Marvell Technology Group Ltd. Aktie

All right. Good afternoon. Welcome back to this session at our BofA Global Technology Conference. And as a real huge surprise and positive surprise, we have the team from Marvell join us. Really honored and delighted to have the CEO, Matt Murphy; and Ashish Saran, the Head of Investor Relations. And I was told that Matt actually just landed back from Taiwan, where I think he had a little bit of a quiet trip, right? It was quieter than usual.

Very calm trip. [indiscernible]

Very calm. Nothing happened at all.

[ Chasing ] wafers.

Exactly. So I'm sure we will talk about market caps and such, but let's get to fundamentals. So Matt, how would you kind of describe Marvell for the next 5 years, right? What -- how do you see Marvell fitting into this ecosystem where you have 1 or 2 really large companies who have both kind of compute and networking resources? And then you have other companies who are good in compute or in networking, right, but not in both. So how does Marvell fit into this ecosystem? What is kind of your strategic plan for the next 5 years?

Sure. Maybe before I talk about the next 5, just for a moment, just a quick recap because it's important to understand how we got here and where we are, right? So actually, I'm coming up on June 20, 2016, I was announced as the Marvell CEO. So you can believe that for those of you guys have been around, I can't believe it's been 10 years.

But the first 5 years of the company was really about rebuilding and transforming into really to try to get to where we were today, which was our aspiration to be a leader in data center and data infrastructure silicon. So a lot of ins and outs in the first 5 years, selling businesses, acquiring businesses.

But net-net, even though there was a lot of ins and outs, and we shrank to grow a couple of different ways, we doubled Marvell actually in the first 5 years. We were at $2.2 billion in the first quarter run rate I was CEO. We ended 2021 post-Inphi at like $4.5 billion, $4.6 billion, something like that, pretty well positioned.

Took a big risk with Inphi, if you guys remember. Actually, fun fact, the banker told me this, the 2 highest multiple transactions of any scale in the chip industry in the last 25 years, number one was Inphi, 13.3x forward multiple. The second was the divestiture of the automotive business of Marvell, which was 12-plus times forward. So I've been involved in a couple of these high stakes ones. The Inphi one worked out really well, but that was sort of 15% growth in the first 5 years getting ready. We took a breather. We digested. We integrated. We fired up the M&A machine again.

In 2025, we did the auto divest. We acquired Celestial and XConn really to address the scale-up opportunity, which we're going to talk about because that's a key sort of call option in front of us. And so if you just take The Street revenue this year of $11 billion plus, we'll have grown the company about 20% a year compounded for the second 5 years I was here.

And actually, even if you go back to the Investor Day we did in 2021, when we closed Inphi, the whole company was at $4.3 billion run rate. So we've actually hit the high end or more than our Investor Day targets, which normally are 3-year targets, right? And then you sort of reset it. 5 years, we've been growing the company. So that's all been great.

I am so fired up on the next 5 years, right? I mean, the last couple of years, we've been growing at a faster rate, right? So first 10 years was like 18% compounded. Obviously, shareholder returns were way above that. Data center went from sub-10% of revenue to 75%. So next 5 years, data center is going to continue to grow in importance for us in terms of the contribution.

Total top line, I mean, if you just look at the last couple of years, data center last year grew 46%. It will be like 50% this year. And then we signaled our current indication for next year would be like 55%. So it's accelerating, but that's all on the back of 10 years of blood, sweat and tears, man. This isn't some overnight success, some new PowerPoint presentations, some new concepts. I mean you had to start on this a long time ago, okay, to be ready.

So why we're well positioned, I think, to keep this kind of a growth rate up and the position we have is that -- and here's where we're unique. As you mentioned, there's a lot of the big scale AI players -- and I said this in my COMPUTEX presentation, there's 3 big compute ones, and these are all $1 trillion or close to and then $5 trillion for NVIDIA, compute-heavy companies with some networking, but they're compute-dominated, right, compute-heavy. And then you have the memory companies, those are obviously like big companies and solid, they're all memory.

If you look at us, we're like, whatever, majority is connectivity, not compute. And that's one of the reasons why NVIDIA invested. By the way, we can come to that investor if you want to talk about it, but we're very complementary to the rest of the ecosystem. We're not battling it out in some compute war. It's a part of our business. But even the people that use us for our custom are typically using us because of our connectivity IP and our I/O IP and the fact that we can stitch it all together with the rest of our solution.

So when you look at the, let's call it, the 70%, whatever your number you want to pick is, the connectivity, which includes a whole bunch of product categories, that has a ton of legs to it because that's -- and I'll pause in just one second, but that's today, the growth has been driven just by the scale-out deployments, which has really been in our traditional DSP business but also even like in scale-out, the switching is really doing well now. But you've got scale-up all in front of us. That market is probably as large an opportunity for us to scale-out in terms of connectivity.

And then you have scale-across, which prior to that was just really DCI, that's moving now AI workloads across. So I don't -- that's where I think a bigger opportunity exists -- and we don't know how big that can be, but certainly, we're well positioned, again, because of the last decade of investment, which at COMPUTEX, I said was basically $22.5 billion of acquisitions, $18 billion of organic investment and $4.5 billion of divestitures. So $36 billion of kind of net spend has gone into this to get to where we've gotten to today.

So Matt, you were just back from COMPUTEX. I was hoping if you could give us an update on what were sort of the key points, right? You made a 102-terabit product announcement as well. But I think bigger picture, I wanted to get your take that everyone seems very positive right now. And you have kind of gone through a number of these cyclical ups and downs. How do you sense kind of the durability of this cycle, right? What are your customer discussions indicating?

And you're saying cyclical from a -- take a long-term view of the semi industry, you're not saying Marvell specific.

Exactly right.

Yes, yes, yes. No, I'm in my 32nd year in this industry kind of on the front line from day 1 as a product manager all the way to general manager and VP and whatever, all the days for different things I did. So I've gotten to see all of the major cycles we all remember. And I also remember the ones that everybody has forgotten, by the way. There's been a lot of them.

Remember the Greece issue? Remember that one? There was like a Greece financial issue and that had a slowdown, 1998 contagion. And anyway, there's all kinds of different ones. But this one is really interesting.

I think one is we're just in a much more thematic cycle here because it's a true global infrastructure build on AI is kind of on the scale of an industrial revolution type of event. And I think -- so one, as you can see in our own numbers, right, just us as a proxy, one company, 46% data center growth, 50% -- 55% as we stand here today is the outlook. That cycle, to me, at least at the moment, looks very healthy.

And the reason is there's -- this is with a lot of natural constraints in the system. There's still not enough logic wafers to go around, and that's going to take the leading companies several years to put that capacity in place. I think the memory shortages are understood and there's a path there, but that's going to take time.

There's just the global power grid. There's permitting. There's -- so there's all kinds of, I think, constraints, Vivek, in the system that I think in the past, we had these kind of unfettered access to supply, which then created these issues like PCs over cycling or smartphones over cycling or -- and then COVID, by the way, was a supply-related shock.

We're in like a demand shock now. But I don't -- we don't see it slowing down. And if anything, we're being accused of being a little conservative for next year because in our model to get to our $16.5 billion target, we do assume CapEx will moderate. We don't think it's going to collapse, but we have kind of a 30% plug in there for next year. We're not the experts. You guys have better data than us. But to the extent CapEx bias is higher than that, we would probably do better. And I think there's a chance that it does. But right now, I think everything is -- bookings have been incredibly strong. Backlog is laid in.

And then I would just say on the design side, by the way, this point about really having the core IP developed and production worthiness of what we do. And I mean just take silicon photonics, for example. We've been in production for 10 years on silicon photonics. We have 15 billion hours of data in the field. We've gone to high-volume manufacturing on 4 generations.

Customers, when they look at us -- and that's just one example, they're looking at us saying, if I really want to get there over the next couple of years with a partner who has all the pieces and it's proven, we're in very rare territory here. There's a lot of companies and concepts out there. There's start-ups. There's PowerPoints. There's proof of concepts. There's a demo at a conference. It's not the same thing, okay, as shipping in extremely high-volume manufacturing with the right capacity, cost structure, yield, et cetera.

So I would say, if anything, not only is the setup great, I'd say in the last 6 months in a very pronounced way, I think our customers are realizing if you don't have all the end-to-end and all the pieces and you're relying on a cobbled together third-party situation of IP and back-end design houses and start-up companies, and I'm just going to cobble it all together, and I think I'm going to make it work, I think we're seeing that recognition that it might or I can have Marvell underwrite my success because it's going to work, and I can trust these guys. So that's kind of all in front of us, right? That would be incremental things that we would go off and do. But I'd say the positioning of the company is extremely strong at the moment.

Yes. And I think what's equally exciting is that we are a kind of revenue diversity as we get there. It's not about just more of the same product line. That absolutely is going to happen. But we, for example, called out 3 new $1 billion businesses, all happening in the next year alone, right? And this is -- by the way, this is before scale-up becomes really big, right?

So I think that's the other nice part. It's not a single socket. It's not a single product line. It's not a single technology, it's not a single customer. I think that's very exciting for us is that the revenue diversity in this company is going to be very significantly growing in the next 2 years.

Got it. Right. I mean the nature of the beast. So one of the arguments you would hear is, well, gosh, you guys are so concentrated. But let's take a step back. First of all, the whole market for semiconductors today is AI. I mean if you just look at like the absolute spend now, what's happening relative to auto, the traditional ones we sell, you have auto, you have PCs, you have industrial, you have smartphones. Just look at the numbers, right, and where it's going.

So AI is the market for semiconductors. Let's just accept that for a minute. So then which again has its own risk, I got it, but that is the market. And then you say, well, within that, how diversified are you? And we're not a one-trick pony. We don't have like one chip, and it sells to one customer, and that's like a bunch of our revenue. And that's exactly what Ashish said.

We're -- and we had some of this issue on the ramp-up on custom. That's why there was a lot of excitement over the last couple of years, hey, your custom business is going to be a big portion of the total, and we had to kind of go through that cycle, but I look out into '27 and '28. And that's why we called out these things. I mean, broadband analog, SiGe components, SiGe, high-performance analog, $1 billion business. Cloud switching, $1 billion business. DCI with very little scale-across $1 billion business. Whole company was $4 billion a few years ago. We're layering in.

And then within those, they have multiple customers to them. So -- and then even with our custom area, again, by the time you get to 2028, you're going to have 15, 18 products in production, some are XPU, some are XPU attached, none of them is going to break the bank here. So I think that's really powerful for investors to see that there's a -- among all the semi companies selling into AI, we actually have, I think, a pretty favorable diversification plan going forward.

And I'm not saying the others are bad because they don't. They might be more concentrated because our customers have gone nuts. That's great. I'm just saying that's where we end up, and that's why we have so many shots on goal. We think that's a healthy thing to have. And we're able to do it in our OpEx envelope and fund all these things. And we think there's strategic value for us going back to like, well, how is Marvell different? We are generally agnostic, so we can work with pretty much everybody, including people that would be considered a rival like take NVIDIA.

Yes, maybe we do custom silicon. They make GPUs. But even in their announcement, they're actually enabling us on custom with a lot of their IP to be more competitive because they realize it's going to be a fungible type of a solution out there. So that's -- if you want to just get to how we're different, connectivity heavy, I/O heavy, a lot more customer -- lot less customer or socket concentration and a diversity of technologies, which in the end, and we can get to scale-up if you want to, you pull it all together, it's very powerful to truly have the end-to-end solution you can provide.

Got it. Maybe a quick word on NVIDIA and the engagement, Matt, that you have, whether it is across CPO or NVLink Fusion or just in your core optics business. And I also wanted to take the opportunity to ask were you as surprised as all of us when you saw Jensen pop up on stage?

Yes. Well, let's cover the partnership first. So yes, that's off to the races, and it's very strategic in nature. Certainly, we have existing -- and it's built on, by the way, years of the companies working together. This wasn't some new concept. I mean Jensen and I fired up the teams to work together. I mean this is -- I think it is even pre-COVID probably, if you go all the way back on just how do we sync up, how do we work together across a variety of their platforms, by the way. So this has been ongoing for some time in a collaborative way. We formalized it, which was great, and then we sort of wrote down some of the things we can do together. But most of it is very strategic in nature, right?

It's about enabling us in custom silicon. It's enabling NVLink Switch and NVLink Fusion to be more broadly adopted. What a home run on the Celestial team joining us and joining forces with our own optics team, right? We had a team that did the 10 years of history of silicon photonics. Now we got Celestial in. That's very attractive because now that whole combined team can work with NVIDIA. We could drive standards that are interoperable. We can make all of our chips work together and all of our silicon photonics work together, huge customer benefit. We're even doing stuff on like 6G and AI RAN. I mean it's quite encompassing. So it's very strategic. It's really based on long-term kind of R&D alignment. So that's all positive.

And then, yes, he was nice enough to support us. I did my first COMPUTEX presentation ever. Hadn't done that before. We got invited. I thought it was a great opportunity. I was honored to be able to have Marvell kind of be on that stage. So I said, yes, quite a while ago. And then we asked him to join us on stage, which he did and also Tien Wu from ASE, who is the largest back-end company in the world.

And yes, I think he knows our story well. There wasn't a lot of prep. I mean, with his team a little bit. But yes, he came on and he did a great job. Actually, he rattled off my whole presentation in about 3 minutes, first of all. That was not -- he was not scripted. That was not a rehearsed thing. He just kind of knows what we're doing. And then, yes, he sort of dropped this very exciting comment about Marvell is the next trillion-dollar company, which was quite a surprise. You can watch the video. That was definitely not scripted.

What was your reaction and then you...

Watched the video. I'm like, I think, I said, we have a little work to do still. But he's a great guy and his team has been just outstanding to work with.

So your reaction was makes sense or...

No way. It's a long-term journey. I mean if you -- I would say, though, if you go back, I mean -- and again, like 10 years ago -- guys, 10 years ago, Marvell's market cap was $5 billion and its enterprise value was $3.5 billion. You could have bought the whole thing for $3.5 billion. It's a real number. Go look it up, $1.5 billion of cash, $5 billion market cap. That was the starting point.

So what could you imagine? I didn't imagine it could go from $3.5 billion of EV to $270 billion today in 10 years. That's like 80x -- so I don't know. What can happen in this market? It's unbelievable actually. It's unbelievable what you can accomplish actually in an industry that's so unique and so vibrant like this.

If you have the right strategy, you align yourselves with the right kind of group of people -- I mean, for us, we stayed very -- if you followed me or invested with me for a long time, we're very consistent. We're very methodical. We grind it out year in, year out, quarter after quarter. We're -- and my whole guiding principle from day 1, and I told the Board this in my first Board meeting on June 11, 2016, I said, my job at this company, #1 job towards everything else is capital allocation. And a few of them said, what's that?

And I said, my job is to steer the R&D efforts of this company and make the best possible bets with the best possible returns and have a vision for where this company is going. And that's pretty soon after I became CEO, I outlined this thing. I said, look, I think the whole market cap of the world is going to move to these data platform companies. And that's where the semi TAM will follow.

And we didn't have anything. We had $200 million of revenue in 2016 in data center. It's actually $190 million, 1-9-0. It's like $9 billion this year or something. But we didn't have anything. I don't have anything to work with. So that was the M&A and the divestitures and all those things.

So my point is, what's possible? I don't know. I don't know what's possible with Marvell. But I do know that we're -- if the market continues and the AI infrastructure build continues, where we're sitting now, we couldn't be in a better spot relative to just how we're performing on the engineering side with our customers.

I mean it used to take us 5, 6, 7 spins to get a chip out. I'm not kidding you. I had some chips to add one product line, took 14 spins to get a chip to production. 90% of our chips now that are 5-nanometer and below -- actually, no, I think it's almost 100% now, but A0 first-time success.

So as long as we keep performing, and I think the products that we have for scale-out, scale-across, scale-up, custom silicon, switching, storage, I mean, all the key pillars we outlined 10 years ago, moving data, processing data, storing data, have those key IPs, bring it all together, that's still kind of the recipe that we're following. It's the same playbook from 10 years ago.

Makes sense. On -- you mentioned the transition from -- before I get to scale-out, I wanted to ask you about any constraints that you're seeing in the supply chain, right? Everyone seems to be tight and Marvell has gotten very big. There are still much larger players also. How are you ensuring, Matt, that you get your fair share of allocation from whether it's foundry or substrate or other parts?

Yes. Well, we definitely retooled the machine here about 5 years ago when we hit the pandemic, and we were kind of woefully prepared for that. I think -- the reason is, I think, Marvell had operated very tactically relative to supply chain and kind of reacted and they have 3, 4 sources for everything and kind of get cute, and it was like how do you save a penny. And I think that was born out of the business model back then. So I'm not -- it's not like a criticism, but that's how it was. And I think for consumer, that made sense.

We started getting into all of these data center projects, and it was like this is not going to work. And so we retooled the supply chain. Chris Koopmans took over as our Chief Operating Officer. We got very, very strategic relative to these engagements that we formed in partnerships, including a very robust long-term forecasting process that we run as reviewed by me and the Board once a year. I push it out to my suppliers and then I show it to them.

So I got to tell you, my supply chain for 2026, even though, again, there's a lot of concerns over the last few years of what's happening with Marvell, guess what, I told them what I needed. I forecasted $10 billion of growth to my suppliers back when I was $3 billion, $4 billion in revenue. I did. So I'm kind of getting what I told them, to be honest.

And if you look at where we were last September, as we kind of started to reorient investors, back then, we were saying $9.4 billion for this year and $11 billion for next year, and then we went to $13 billion for next year, and then we went to $15 billion for next year, and now we're at $16.5 billion. We've been getting the performance out of our suppliers. It's very tight out there, like I mentioned, on not just logic wafers, but across the board. So it's a battle. But it's been a battle since the pandemic. This isn't like some new thing we're not used to.

So we're pushing the supply chain hard, but we form these very deep relationships with our suppliers, and we do more with fewer. That's why Tien Wu. I mean, he's never been on stage at COMPUTEX before. He runs the biggest back-end company in the world. Super strategic, right? And he bet on me and Marvell like personally, like I think you guys are going to crush it. And look, they perform incredibly well for us.

So we'll keep working on it, but the demand is not the problem right now. Supply needs to be -- continue to be worked and planned for. But I think we keep planning our business well. We can keep growing quite aggressively. And we have enough supply today to do everything we've told you guys. So the question is -- yes, so that's all in place. The question is, if CapEx bias is higher or some of these projects really take off, can we do more? And that's what I'm trying to prepare my suppliers for.

Got it. I think on the networking and the optics side, there is kind of broad recognition that Marvell has led important parts of that market, and it's a diversified business to your point.

I think it's compute that I wanted to dig into a little bit that over the last 2 or 3 years, at least from the outside, it looked like the business went through a lot of volatility, right? The expectations got very high, right? Then even though from the inside, it might have been more predictable and kind of more according to a plan.

If you look over the next 2 to 5 years, Matt, do you think that you're now at a place where you have visibility on not just the size of the opportunity but also your share in those opportunities? Because that is what has created, I think, angst in the last 2 years or so.

Yes, yes. And there was definitely some lessons learned, but just to kind of provide the kind of what really happened and set the record a little bit. So we basically had effectively a $0 custom silicon business in data center when we bought Avera, which was out of IBM, GlobalFoundries, IBM. We did a good job winning designs there.

And along the way, we actually kept abreast investors up to speed. At one point, we had said for calendar '24 that we could do. $400 million in revenue, and we did that, and we exceeded it. And then we actually told everybody in April of '24 for calendar '25 that custom would be about $1 billion business. And I never updated that number. I just kept saying it's going to -- it will do better.

And what I learned from that was that, yes, I think people's expectations ran away from what I had told everybody, which is I only told people $1 billion. And we beat it by 50%, by the way. I never said it was going to be $3 billion or $4 billion or $8 billion or whatever people. People put stuff in their model later, and they're like, well, I thought it was going to be this. I said, well, I didn't tell you that.

I mean, guys, if you worked with me for 10 years, I kind of give you a sense of where it's going to go and like we're generally pretty close. You should do better actually. So -- and that's on us, too. Like I'm not blaming people. I'm saying that's what happened.

And so the reset that I really did in September -- and I'm going to get to your question, but I just want to make sure people understand, what should have been a huge success story is you guys took this spin-out of IBM that hadn't -- was stuck on 14-nanometer at Global. And we moved them to TSMC, 5-nanometer, SerDes worked the first time, won a bunch of sockets and got to $1.5 billion in revenue in like a couple of years. So it was like an unbelievable success story. But it was sort of like, oh, you missed and you screwed up. And so that was all shame on everybody.

So in September '24, I just said, look, at September of last year, I said, look, I'm just going to be very prescriptive now. I don't want anybody getting confused. I don't want this to happen again. I don't want my investors to go through this stress anymore. So we just been -- I've been very clear, hey, I think it's going to grow 20%. Why is it going to grow 20%? Can it be more? Shouldn't it be double? No, it's 20%.

And then in the last quarter, I said, well, it will be more than 20%, but just a little bit more, like don't go crazy. And all the way, the reason I'm doing that, too, is I did said at the same time, you say, well, you're so conservative, Matt, you're always so conservative. Well, guess what, April of '24, we did our long-term model for the first time in custom silicon, and we put -- basically, we said it's a $40 billion market in 2028, and we think we can get 20% of it. That's an $8 billion hanging out there. My whole company was $1.1 billion a quarter when I said that. So say I'm conservative or not, we put a model out there.

And then we said, CapEx is probably higher last year, so it's probably going to be more. So call it, $8 billion, $10 billion. So we're tracking to that. We're tracking to that. business is performing better than I said it would last year for this year.

Next year now, we had -- last time I had said custom was going to double year-over-year. It's going to more than double. Again, don't go crazy, go crazy, but more than double. And part of the reason I'm doing this, too, Vivek, is I'm trying to guide everybody to the fact that we still continue to have conviction as we've had for the last 2 years in our model that gets us out into 2028.

And now I think people are starting to realize, oh my gosh, I think these guys might hit these numbers. So we still got to go hit the numbers. We're not there. But it's really about just trying to -- especially on that piece of the business, just being much more kind of prescriptive so that we don't have a situation where expectations get too high. I'm really mindful of that.

And the nature of those, by the way, they're very specific programs, and they're so confidential, like I can't talk about them. You guys want to talk. I can't talk about them. But the nice thing is it's very diverse now. There's like bunch of XPU sockets, a whole bunch of XPU attached sockets. By the time we get to 2028, there's enough shots on goal that even if one does better or one doesn't do better, like we're going to be okay, net-net.

And -- but at the same time, we're extremely well positioned there. We have a great competitive platform at 5-nanometer, 3-nanometer, 2-nanometer. We'll be there when we get to the angstrom era. And every generation, by the way, on XPU and XPU attach, die size, die costs are going up because there's more density. Core counts are going up. I/O is getting more complex. ASPs are going up. Everything is going up and volume is going up. So at some point, I know these sound like big numbers to double or you going to triple it or something, but we're still relatively low share relative to the whole spend, too.

One final thing, one investor say, I can't believe you're -- I don't know, is it $4 billion next year? It's so small. So tiny. Like why do you even care? I'm thinking, sir, this company was $2 billion in revenue. That would be a nice number. So we're good.

Like just going to keep making progress against that and keep investing in the technology, do a good job for our customers, keep our mouth shut and just try to really do a good job delivering for you guys. So you have confidence in that base. But then let's let it ride on the connectivity, right? That's where I think there's a lot because that's the whole market. That's a market we can create, we can control. That's our engineers, like we can make the TAM happen there.

The other one, it's our customer that's really got the control point. And we're their like Oz behind the curtain to make sure that they're successful and help them out and do their business model and let them get all the credit. It's their ship.

Final question, Matt, scale-out to scale-up, how does your scope of opportunity change? And when do you see CPO becoming like a real product with like the depth with which you describe numbers in compute. At what point do you think you'll be in a place to describe the journey of CPO? And do you see customer diversification? I'm just kind of wrapping all these questions.

Sure, yes, I'll do the final one. So scale-up is all in front of us. Like I said at the beginning, the scale-out and the build-out of the AI infrastructure has been led by scale-out, and that's what's in our numbers today. That's what's driving interconnect growing 70-plus this year, right, as an example. It's ripping. Ripping last year. This year, it will continue to do well next year. That's scale-out driven.

So scale-up, a couple of things on that. First, from a switching standpoint, I literally -- on any number I've ever given anybody, there is $0 today ascribed to anything on switching. That's -- I mean -- and by the way, it's not like I'm not investing. I've got UAL projects, ESUN projects, PCIe with XConn and now -- and then an opportunity with NVIDIA, right, with them. So that's going to be a big investment focus for us. That's all in front of us, more to come there. But that's all new incremental TAM that's going to get created, right? That's coming.

And then from an optics standpoint, I think people got it, but we called out actually for next year, $300 million in revenue from scale-up optics, scale-up optics. So that's CPO, that's NPO, that's our light engines. It's not just Celestial. Celestial was a company we acquired in December, which had a lead customer and that's like, say -- that's on track, so let's say half of it. But I mean, it's actually starting next year. And that's like beyond early stages. It's 0 this year. So you're going from 0 to $300 million with -- who knows where it could go. So I think we're at the very early stages on that. So the only number I've given you is $300 million.

Yes. Maybe just one other thing to consider on the switch side in particular, unlike scale-out switching, where we enter the market, but at that point, somebody is already well established and you're -- we've done well, right? We've taken it from 0 to $100 million, $300 million, $600 million this year, $1 billion plus. Scale-up switching is completely greenfield. It's fully available. We could be leading the market from day 1. That's a massive opportunity. That's a very big difference, something to keep in mind. And because we also do XPUs and we also do the interconnect and we do the switch, that's a massive advantage. The scale-up...

And that is the discussion. I'll just end on this. The discussion right now that we're deep into across the board is pick your scale-up switch. We have it. We're not making some holy war bet on this thing, by the way. I'm investing in all of it. I have the R&D capacity to do this now. I don't have to take that bet actually.

So I got them all. I can absolutely prove and show how we can attach and integrate our optics there, whether that's CPO or an NPO-based solution. I can also do the same thing on the XPU or GPU, and I can provide that as a chiplet. And with our expertise, whether we make the XPU ourselves or we don't, we know how to integrate with that. And we know how to make the whole link work. We already do that. We know how to do this.

So that whole discussion -- I mean, always in semiconductors, you talk about the platform and the solution and then a lot of times, people say they have the platform and then you just pick one piece. But this is one, guys, I'm telling you where I think certainly in this first wave to make all this hang together, it's really going to be powerful to have proven manufacturing and reliability and technology for the scale-up optics and have flexibility on the architecture, whether it's CPO or some version of MPO and all the subcomponents around it, have the platform teams to make it all work together, have the switching IP and technology and be able to deliver reticle-size chips with hundreds and hundreds of I/O at the bleeding edge and be able to control that I/O, so you can pack it all in and optimize it and then have the ability to do the attach on the XPU side or build the entire XPU and integrate the chip directly as CPO. So that -- I mean that's not in any revenue forecast today, that thing. And the optics alone are worth almost basically just assume whatever the switch ASP is, that's the content opportunity on the...

Optics side.

So it's potentially very large, and we have not sized it other than saying, hey, the initial scale-up optics only for next year, which is year 1, has gone from $150 million when we bought Celestial to $300 million in like 6 months.

Got it. We have 6 minutes over.

Yes, I know.

Thank you so much.

I'm running on fumes from Taiwan. I'm done. I'm out. All right. Thanks, everybody. Really appreciate it.

Yes. Okay. Great. Thanks. All right. Well, very excited to have Marvell join us. My name is Mark Lipacis. I'm the senior semiconductor analyst at Evercore ISI. And so today, we have Willem Meintjes -- hopefully, I'm getting that name right and who is the CFO; and then Ashish Saran, who is the SVP of Investor Relations. So there's a lot going on in your sector. So I think let's just get right to it.

But maybe just to start out with, you guys just reported last week and maybe if you take a moment and kind of summarize for everybody, like what do you think are the main takeaways?

Yes. Thanks. It's really great to be here, Mark. Yes, so we reported last week. And I think if you go back over the last several quarters, we've really established a different cadence here, and it's really driven by the market. What we've seen is the end markets that we're addressing continue to be extremely strong, and it's really driven by CapEx, right?

And so if you go back to September last year, we, for the first time, sort of indicated some longer-term numbers, right? And for the current year, we sort of took that number to $10 billion. And then outer year to $13 billion, right? And so in a subsequent quarter, we upped that from $10 billion to $11 billion, and then for the outer year from $13 billion to $15 billion. And then just last week, we've now taken it to $11.5 billion and to $16.5 billion for the outer year. And so we've been sort of very constantly been able to drive the supply chain with the upsides that we've been seeing on the demand side.

Yes. I think our interconnect business, in particular, is absolutely on fire. I mean that thing is growing at 70% plus for this year. I think within that, if you think about there's really 2 big underlying drivers. There's a scale-out business that's growing even faster, quite frankly. That's really our PAM DSP, TIA driver business. And then you've got scale across, which is just starting. It's going to get a lot bigger in the outer years, right? So that's -- the short to medium term from a revenue perspective looks very, very strong. And then I think what's even more exciting is all the engagements we are seeing a little bit further out in time, which I'm sure we'll get into today's call. But I would say that's equally exciting as well.

Great. All right. Thanks for the setup. So I think there was -- in Taiwan, there were some news, Jensen Huang said that you guys are the next $1 trillion market cap company. And I'm hoping that you can talk about that a little bit. And before you do, though, I'd like to ask or just reflect on the checks that we have done, which indicated over the last 6 months, we talked to about a couple of dozen sources amongst the hyperscalers. And very consistently across the group, what we heard is that Marvell is being viewed more as a more strategic partner supplier. And that's because you have the broad portfolio of IP. And this was coming through not just on the XPU side, but also on the networking side.

And so can you -- if you want to -- if you care to make any comments about this idea about where your market cap is going to go or the context that, that was put in? But if you could also just remind investors of the relationship you have with NVIDIA and NVLink fusion and to the extent that the checks that we have picked up on, which I believe Matt played back in your call, how that is working into that relationship also?

Yes. Maybe I'll start and Ashish can talk a little bit more about the NVIDIA relationship. But really, if you look at it over a longer arc, I think it's a real validation of the investments we've made and the relevance of connectivity, right? I think beyond that sort of narrow commentary from last night, I think the broader discussion was really how connectivity is driving compute and accelerated compute and the investments that we've made both organically and inorganically have really positioned us to be the leader in that space. And so as you've seen sort of these bottlenecks on compute and then memory, we're really seeing the next major bottleneck being on connectivity, and the technology to drive past that is really the investments that we've been making. And so we can get into more detail, but maybe you can talk a little bit on NVIDIA, Ashish.

Yes, sure. I think on NVIDIA, I think -- I mean there's really 3 big pillars to that agreement we just announced. The first one of them, of course, is a very long-standing relationship we've had on the optics market, primarily historically on the scale-out market, which is on the pluggable side. But now since scale-up optics, which we'll touch upon more later in today's discussion is becoming a much, much -- that's where most of the new differentiation is going to come from, they themselves have their own path they're pursuing. We, of course, are investing in that technology as well. And I think it just makes sense for the kind of the 2 leaders in this space to have collaboration in where does NPO, CPO go in the future, right? So think about optics as basically becoming a much broader relationship, building upon what we've been doing for them, but now really collaborating much more so on the new scale-up optics market. So that's kind of one pillar.

The second pillar, of course, is the one which I think got most interest and it makes sense is today, if you think about hyperscalers, they've really got 2 different infrastructures. They've got a completely merchant infrastructure, NVIDIA primarily, and then they obviously got custom. And they, too, don't talk to each other, right? So think of that in the long run, that's not how you want to build a data center. You want complete fungibility. That's where Marvell comes in, right, with our networking IP, our custom IP, where we have the bridge between both those worlds, right? So that's really a big part of the relationship, where now you can have hyperscalers being completely fungible in terms of how they design their network. They don't have to choose one or the other. They can choose both. So we see that as really opening a much bigger TAM for the 2 companies combined.

And the last one, of course, is innovation on the AI RAN infrastructure. We've been in that market for a very long time with our OCTEON processors, which help on the baseband side. And as you go from the world of 5G to 6G, you need a lot more capability, you want software-defined architectures, and that's where NVIDIA comes in. So we're building in hooks into our baseband processors where you can basically add NVIDIA solutions and very quickly, you can upgrade your RAN without reinvesting in building the base station from scratch. So it's a very important thing for our customers. So again, a very broad set of relationships, but we certainly see all 3 of them being very important for Marvell going forward.

I have to say in hearing you talk about that, it just reminds me when Matt Murphy first came to Marvell, and he said, I identified the top companies to acquire Inphi and Cavium. And if I think about the comments we heard from the -- we heard from our sources, the hyperscalers, this idea, this broad portfolio of products, it really seems that like that vision is playing out, right?

And can you take that to one -- maybe one layer deeper and just talk about like how is this idea, like, the broad portfolio? And I appreciate that the other thing that comes through when we talk to the hyperscalers is everybody has a completely different view on how the architecture plays out. So if you -- maybe you could just take it one layer deeper to help investors kind of sink their teeth into what is this idea of a broad portfolio mean? Why does that make you strategic?

Yes. I think maybe let's start in connectivity, and then we can talk about custom. But if you take connectivity, right, on the longest reach, we have a coherent portfolio, right, which -- this is sort of the DCI original sort of colors portfolio that was part of Inphi, and that Inphi really pioneered and developed. And so what we're seeing today is that technology starting to transform into a scale across opportunity, which if you look at the scale across opportunity over time, that's going to be a lot larger than the DCI opportunity. And we're investing to be the absolute leader there.

And so when you take that coherent technology and then you start sort of moving down inside the data center, there's reaches there where it's sort of too far for PAM technology, and we're able to do something that we're again pioneering on coherent-lite. Obviously, Inphi was famous for the PAM technology that when we acquired them. And we've continued to invest to be the leader on each generation going from 400 to 800 gig to 1.6T going to 3.2T next. And so I think one part there that we didn't necessarily talk about previously, but on the last call, we mentioned that if you look at the broadband analog, TIAs and drivers that's part of that portfolio, it's actually very significant. It's a $1 billion business for us.

And then you can kind of keep going down the stack on connectivity once you get inside the chip on die-to-die SerDes. And so this whole -- the entire stack from connectivity all the way from hundreds of kilometers down to millimeters, it's Marvell IP that's enabling that. And so when we look at the custom opportunity, maybe you can kind of tie that in.

Yes, maybe one other example where all the pieces come together is scale up, right? So if you think about when you want to build a scale-up network, there's 3 key components. There's a compute engine, the XPU. There's the interconnect, right, which is copper today going optical over time, and it's the switch. We have all 3 pieces, right? We build XPUs. We're investing in switches, and we have the leading photonics technology in the market.

On top of that, to your point earlier, that each hyperscaler is looking at multiple different alternatives, this is where our breadth of technology and investment lets us do multiple flavors. So we're not making a bet on any one single thing. In optics, we basically have 3 different modulator technologies for doing essentially scale-up optics. It's MRM, Micro-Ring; MZM, Mach-Zehnder as well as EAM. We're also investing in even more exotic technologies, whether it's microVCSEL, it's microLED. So pretty much all options on the optical side are available. Even on copper, we can do co-packaged copper if that's what customers want initially as they go towards NPO, CPO.

Similarly, on the protocol side, on the switch side, we're not restricted to offering one type of technology. We have a UALink switch coming out now. It's 115T product. We have an ESUN product, which is based on a 100T platform. And of course, with the whole discussion we had earlier with NVIDIA, we have full access through us through NVLink platform.

So I think that's the idea when you really think about having end-to-end, really having all the pieces. Scale-up is almost a perfect example where a customer can work with us on the entire rack-scale infrastructure upfront. We can design all the chips for them, the entire signal path and give them different optimization paths, which is a fairly unique ability. There's not that many other people which you can do it in the world today.

How is the -- you talked about the optical engine, the Celestial AI. Where are you in the development cycle? How should we think about like the milestones? Are our own checks are indicating excitement around this and the potential for this to be huge for you guys. So like how should investors think about Celestial AI and where it is?

Yes. So I think, first of all, I think if you just step back, there's this category of scale-up optics, right? And so Ashish was talking about, there's a bunch -- a whole range of different technologies enabling scale-up optics that we're investing in. So certainly, Celestial is one of them. But at the same time, we're doing CO-OP technology, MRM with TSMC, where we have MZM, right? That's the photonics engine that's been in our DCI product for a long, long time. And so both NPO, CPO, we're playing in all these different technologies.

And so the update that we gave on the last call was that initially, when we said -- when we acquired Celestial, we said $150 million for next year. And we've updated that to incorporate the whole category and actually doubled that number to $300 million, but that includes all the different go-to-market.

And so Celestial is on track. We're also very much excited about that opportunity, but the opportunity is a much larger opportunity. And going forward, we're really updating on the overall scale-up optics opportunity versus just the specific.

Yes, the Celestial product is basically -- it's well past development. At this point, it's really entering a manufacturing cycle. We've lined up sources of supply. This is really going towards -- I mean, basically, it's going to be in production essentially next year. So it's well advanced on its path to get into production. So that revenue for next year still stays where we suggested last time. But then, of course, to your point, that ramps very significantly as you go from calendar 2027 to calendar 2028.

So is this the $500 million bogey?

On a quarterly basis, exiting the year next year, it's still $500 million, and it doubles to $1 billion exiting the year following year. So if you kind of work your way through the math, it's basically saying over this roughly 1.25 years time frame from the start of production. It's about $1 billion in revenue, plus or minus. And that's just the very, very beginning of this scale-up ramp. And then on top of that, we would get revenue, as Willem mentioned, from a broader set of scale-up optics solutions, which we have, including our light engine as well as some of our TIAs and drivers.

And Celestial AI, there's -- there haven't been production revenues, as I understand yet, but there's been 4 generations of the product. Can you just talk about the develop because I think there's a set of investors who are concerned about how can you project that kind of revenue when we haven't seen it yet?

Yes. I mean you should think about that this technology was actually demonstrated several years back. So -- and you should imagine that the lead customer went through a very, very extensive bake-off process, right? With very heavy technical detail, getting physical product in their hand because that is basically their scale-up optics, one of the key solutions they need going forward.

So in this case, it's not like what we tell you, it's basically where you've got a very large hyperscaler, which has gone through that bake-off process, and this is what they picked. I think that's probably the best validation of that particular technology. And at this point, we are -- like I said, we are well beyond -- this is in a productization phase, where we've lined up basically capacity for volume manufacturing essentially starting fairly soon.

Got you. If we could shift over to the XPU, the ASIC business. The news flow, I think, on this business has been the bane of your existence in a way. And I think there's confusion about what you guys do and what other suppliers in this market -- or claimed to be suppliers. Can you help us understand what is that -- I guess, like we think about you guys as an IP business, how should we think about the competitive landscape? There's companies in Taiwan, like what do they do? What do you guys do differently? I think a lot of people understand what Broadcom does. Like how should we think about -- clear up the confusion for once and for all.

Every time. Once and for all, every time.

This time, it's going to happen.

Let me just frame the size of the business for a second and the puts and the takes, and we can kind of get into the more technical piece. But I think the first piece is that we said that, that business is going to grow over 20% this year. So from $1.5 billion to, let's just use round numbers, $2 billion and then, over double the following year. So rough math, $4 billion plus, right?

And we've actually given quite a bit of granularity on the drivers for that growth where we basically said 1/3 is from our current program, 1/3 is from the new XPU program -- XPU attach programs that are ramping, and that's actually 10-plus programs that are going into production or will be in production next year and then 1/3 from the new XPU program.

And so the key point there is that this custom business has actually become quite diversified. There's multiple growth drivers, multiple engagements. And what we've seen is if you look at the XPU attach part, that's CXL and NIC. And when we spoke about that originally, we saw those sort of going to $1 billion each in the outer year. And that's very much on track to sort of get there or even more than that. I think we've seen those attach areas just become more and more relevant. And so...

Yes, maybe in terms of what we do, which is a little different than I think, look, there's companies like us, which is the reason why companies are coming or customers are coming to us is because there's a lot of our interconnect technology, the entire discussion we had for the last 15, 20 minutes or so, which is very relevant when you build large complex XPUs. These XPUs are no longer monolithic single-chip devices, right? These are basically multiple compute die, HBM stacks. You need high-speed SerDes, you need die-to-die interfaces, you need custom HBM interfaces, you need much more optimized custom SRAM for much higher packing density, you need advanced packaging. And those are all the things we do for our merchant business. So the reason we're in the custom business is actually because of the expertise we've established from our merchant business. And those capabilities are very important to some of our XPU customers.

Now when you build a chip, there's an IP portion, which is what I discussed, which is kind of more considered kind of front-end design. But right at the end of the process, there's also a process where you have to go through a back-end process and do like basically layout, which is physical design, right?

So I think some of the companies you're thinking of some of the design service companies have an important role in the ecosystem, but they don't have IP. And the reason they don't have IP is because they don't have a product business, right? So that's the real clear distinction is if we are engaged in a project, it's because our networking IP is what the customer is looking to access and build into the XPU or XPU attach versus when they're partnering up with somebody in the design services side, it's more of a relationship of the design is done somewhere else and essentially, you need someone to do the last part of the process, which is physical design. That's really the clear distinction, I would say, between the 2 business models.

And the traction that we're seeing is really a reflection of that. I think ultimately, when you go through all this [ threads ] and you get out to the other side where we are today, we've seen a massive increase in engagement on the amount of opportunities that we have. And just on the last call, we said we've actually won multiple additional custom design wins, right? And so really, this differentiation in IP is we're seeing that show up in the amount of traction that we're having with the customers.

So may I ask what -- I mean this is -- it seems to me that this is IP that you've had, but it seems like it's like in the last 6 months where our field work is telling us that you're seeing an inflection. And Matt talked about that on the call last night. So what has happened recently?

It's basically -- it's what you're seeing happen in the world of AI, which is that the rate of acceleration is increasing. The complexity is going up. Again, scale-up is a great example of it. Scale-up networking outside of one player didn't really exist at this point in time. But if you look for the next 2 to 3 years, you're looking at much higher densities. You're looking at much more complex inference in an example. You're hearing about agentic AI today. All it's doing is driving a lot more traffic. You need to disaggregate your memory at this point. You need a lot more traffic between XPUs. All of that is, again, back to networking IP.

So yes, there's a reason why you're hearing what you're hearing because all the IP we've developed and we ship in our merchant products is absolutely critical. Kind of goes back to your first question, by the way, like why did we get called out? This is the reason why we got called out is because the role of networking is absolutely critical going forward.

Yes, that connectivity thread that goes all the way through, right?

So the market is coming your way, basically.

That's correct.

Yes. I mean we've had this thesis for the last decade, and that's driven all the investments that we've made. And I think the interesting thing is that what we've been saying has been very consistent.

Yes. And by the way, it's not just the market is coming our way. I think what's also happening is as you go to these longer reaches, higher speeds, higher densities, the market is coming our way on the optics side. And that's an area where we started a long time back.

I think one of the key things when you look at do you not just have all the pieces, when did you start assembling those pieces? It's not something you wake up today and go license IP and go build a product. These are things you have to invest in 5 years or 10 years.

When you engage with these large customers, they're looking at where will you take them not just today and tomorrow, it's what's your road map look like? And do you have data? Do you have field data, which says you've done it and it works? I think that's the other thing to keep in mind.

So I think the shift toward optics in particular, is going to be very beneficial. I mean this is -- we started down this path with Inphi, but we've added more capabilities, right? Very recently, Celestial is one example. We're looking even further out. We acquired a company called Polariton, clearly a pre-revenue company. But they've got some very unique technology, which can take modulators to speeds which are roughly up to 10x higher than current technology. So we're looking out where does this market go in the next 5 to 10 years, and optics is a massive differentiator for us.

Got you. Now Will, you mentioned the XPU attached design wins. And I think Matt mentioned on the call that the forecast continue to come in higher than expected, if I didn't capture that.

No, you got it right.

Spirit correctly -- please do correct me. So we have a -- the other checks that we've done over the last 6 months indicated led us to this called CPU renaissance, where you have more CPUs because of first, inferencing and now agentic AI, you have a change in the CPU to GPU ratio. And I think you guys mentioned this on the call as well.

To what extent are these XPU attached business or sockets that you have? Is it NICs and CXL controllers? Is it all that? And is it correct to say that these all get attached to CPUs that go into AI servers? Like how could we qualify this business?

Yes. I think, first of all, clearly, there's an inflection, right, in terms of like the CPU attached to GPU, right, that rate. It's very dynamic, right? I think the technologies that we have are clearly attaching to some of that. But exactly how that plays out, I think it's not necessarily built into what we've communicated, right?

I think as you look at CXL and you look at the memory wall, that's very relevant as you scale these CPUs. And so certainly, I think that is a significant opportunity, both not just CXL, but on PCIe, on PCIe switch and retimers. If you look at just the amount of data that's getting moved, there's sort of the MOE and then this is agentic and quantifying the impact of agentic, I think that's still very much ahead of us here.

Yes. But the products are very, very similar to your point. I mean if you look at our revenue stack we outlined, call it, in '28, where we said in that $10 billion, at that point, we said $3 billion-ish plus or minus is going to be coming from XPU attach. The 2 biggest pillars of that essentially, each $1 billion plus was CXL and custom NICs essentially.

And if you think about the simple -- if you're going to have more CPUs, just very direct, you're going to need more NICs, just flat out. And if you're using this primarily for inferencing, you're going to have these longer context windows in inferencing, you need a lot more KV cache data. You're going to need more CXL essentially, especially today where DRAM prices are high and capacity is very tight. So I would absolutely expect -- the numbers we've outlined so far don't really include the agentic impact, just to be clear, because that's all happening now. So the revenue we had outlined was things we had already seen based on attaching to XPUs.

Now with agentic, more CPUs, you should expect to see more upside on those. And by the way, it's very similar products. In fact, if you think about it, CPUs in the last few years were designed with CXL already in mind. They just never used them because all the oxygen went to GPUs in this whole AI cycle, but now the tides kind of reversing a little bit, right? So yes, I think there's certainly a lot of excitement around this, and the attach rates can be actually very, very meaningful for us.

And is the XPU attach -- is it all NICs and CXL? Or is it...

There are some other products there.

How can you qualify this?

So think about XPU attach at a very high level as an offload device. You're trying to basically maximize your central compute device for its core function. This is no different than if you go back years back, right, where you started creating NICs to minimize I/O of the server CPU, same idea essentially.

CXL and NICs are the 2 large examples, but there's a couple more we can identify where we have design wins going into production. One certainly would be storage accelerators because while we are discussing memory as CXL, there is storage, right? So you're basically -- whether it's SSDs or HDDs, right, instead of having the core CPU, XPU address them, you'd rather do it for a dedicated device. So that's a storage accelerator.

And the last one is the security offload device, right? This is a unique product for Marvell. This comes from the Cavium acquisition. We're one of the only companies building dedicated security accelerators, and this is a perfect use case, again, optimized for different hyperscalers, right?

And again, the idea is as more and more companies and enterprises -- take Marvell, we don't go buy our own AI infrastructure. We're using AI as a service, but I want to make sure my data when I put it into a cloud network is encrypted. You can certainly do encryption/decryption on a standard XPU or a standard CPU, but why would you? You'd much rather do it on a dedicated device. So that's another example, which is again going to be a fairly large opportunity for us.

I want to kind of shift gears to the DSP business, which has been a great business for you guys. How do you anticipate your share position to evolve as the market kind of shifts from 800 gig to 1.6T to 3.2T over the next several years. Is the competitive landscape changing?

Yes. I think when we acquired Inphi, there was always this expectation that we're going to lose market share. And frankly, Mark, it just really hasn't happened, right? I mean the teams executed really, really well on investing and being first with each technology. And we continue to see that.

Now the market has become so much bigger that clearly, there's a lot of investment and there's a lot of competition. And so at our scale, is there a potential for somebody to take some small amount? Certainly, that's the case. But we're continuing to drive absolutely very, very strong majority market share across each one of these generations. And the team is ensuring that, that continues.

I mean if you look at 1.6T that actually went into production last year, it's become very significant this year. And 800 gig is still growing. But if you look across into next year, we're going to continue to have very, very, very strong share.

I think the basis of competition in these markets is always were you first to market? Did you have the first solution? And that's been the case in every single generation of PAM4, no different at 1.6T.

The second, of course, you can see it in the numbers. I mean we outlined for this year. We're going to grow at 70% plus interconnect. Within that, scale-out is growing faster, right? And then again, we said next year, we will again substantially outgrow cloud CapEx from an interconnect business. And again, the largest part of our interconnect business is our PAM DSPs. And next year, in particular, 1.6T is a huge part of it. So -- and the reality is I don't think we see anything significantly different. I think we've had a leading position. I don't really see that changing going forward.

I think from our standpoint, it seems like you're planning for billions in annual revenues from hundreds of millions last year for 1.6T. How should investors think about the mix of DSPs versus TIAs versus drivers? You have good products in all these buckets. How do we think about those different elements?

Yes, I mean...

How -- where they work together, if that's fair to ask?

Sure. So DSPs would still be the largest portion of our entire interconnect revenue, but it is getting a lot more diversified, right? So if I go back -- and this is all the information we provided a couple of quarters back. If I go back to like last year, our total business in interconnect was about half of data centers, call it, roughly $3 billion. And within that, we said about $0.5 billion was scale across and the remaining essentially was scale-out, which was dominated by DSPs.

Now that business is going to go from $3 billion to 70% plus this year, right, and then outpacing cloud CapEx, you guys can run through your model, you'll end up with a very large multibillion-dollar DSP business. But at that point in time, to answer your question, we also said scale across will also be a $1 billion business. And then TIAs and drivers will also become a $1 billion business. So the business gets very diversified, but obviously, DSPs, given the volume, given our market share position, one would expect will remain kind of the majority.

Specifically, I think it was underappreciated like the size of our TIA and driver business. So we wanted to sort of at a point in time, just break that out for investors so that you can kind of see the scale of that. We'll probably not break that out too regularly. But I don't think most people understood just the scale of that business today.

That was from Inphi. That's a...

That's -- in fact, that is the original Inphi business, right? When they started down, this was before PAM DSPs, right? So the original business from Inphi was actually building these analog components, right? And that's been a critical differentiator. Obviously, with the takeoff in PAM DSPs and that business became kind of the face and it still is. But the reality is it's a lot more diversified. And it's also a key component of technology required for scale-up optics going forward. It's not just scale out.

Got you. How is -- so everything is growing very nicely, right? How do we think about supply constraints here? And maybe starting here, how do you characterize the supply -- laser supply for you guys?

Yes. Maybe just talk about the general and we can talk specifically about laser. I think we spoke about it on the call last week a little bit, but we've really instituted a very rigorous process where we do like a 5-year forecast. And what we do is really proactively share that with our supply chain so that they have visibility to our growth trajectory.

And I think that's been critical in terms of allocation where when you look at like a year like this year, we've been very consistent in terms of what we've seen in terms of demand and how we've communicated that. And so Chris and team have really -- and you can see the progression here over the last couple of quarters. They've done a great job at securing supply, specifically in relation to those forecasts that we were providing.

Yes. I mean supply has been tight. I mean that's worst-kept secret out there. But having said that, I mean, just look at the upside we're driving this year and next year. So I think we've got -- done a pretty good job. I mean take our data center business. We grew at like mid-40s last year. We're going to grow 50%, so accelerate this year. In a tight supply environment, we're going to accelerate even more next year. So I think that gives you a good sense of -- I think we've got a pretty good handle on it. I think we have a very good understanding. We triangulate demand, make sure we serve it well. And I think we've actually done pretty well on the capacity side.

I think in terms of lasers, I mean, I think as far as I can tell, our customers are able to get as much as they need. No one is saying the data center is not going up because they're not on lasers. Would they like more? Of course. But I would just say, I don't think that's a huge constraint in the near term. But given what you're going to expect to see and scale up optics, I think it's important the industry keeps investing in it because the demand for like external laser source is going to absolutely explode. So I think it's a good question, especially in the longer term, where the industry does need to keep investing in it. But look, this is capitalism 101, right? There's an opportunity, people will invest. And I think over time, I think we'll be in pretty good shape.

Does vertical integration make sense on -- for this component for you guys?

I think at this point in time, I think we feel pretty good about the way the industry is set up, right? I think we focus on what we do extremely well, which is build DSPs. Having said that, I think there are places we've built modules. Our DCI business, we actually build the entire module given the complexity of the product. But in general, I think our focus is on silicon. That's where we differentiate. And I think there's other partners out there which are very focused on that particular part of the ecosystem.

Is the -- it looks like we ran out of time. We're going to go a little bit longer, if that's okay with you guys.

Sure.

The -- how should we think about the long-term data center interconnect revenue opportunity? I think Matt previously indicated the TAM is going to expand 5x by 2030. Does that include the more recent commentary as it relates to like the scale across DCI TAM that's got 10x higher bandwidth than the front end? How do we how do we think about that business?

Yes. I think your last comment is really -- I think, the key there where the amount of data on connecting the back end and a scale across network is multiples higher, right? And so that's where the 10x number comes from. And so to be able to address that amount of data, clearly, that market needs to be a lot bigger, right? And so we've, again, pioneered the technology there. And so our expectation is to sort of continue to maintain a really high market share as you see that market develop on scale across.

Yes, I think the TAM numbers are probably going to keep floating up, Mark. I mean, I think, if I compare it to what we said like 2 years back, and that's not just a comment on the fact, yes, clearly, CapEx is higher than we thought it is. So that tide lifts all boats. But I just think that the sheer complexity of the network, right, with agentic, with mixture of experts with all of these different inferencing models is just a lot more than what we all thought even just barely a year or 2 years back, right? And scale across, we're at the very, very beginning stage of it, right?

I mean, today, I would say the majority of the market is still DCI front-end connectivity. I think the first scale across meaningful revenue really probably starts next year, maybe a little bit this year. One of the key things you need for scale across is customers really want to go to 1.6T. Today, most of the market is more like 400, 800 gig. But think of scale across as just taking your existing scale-out network and just stretching it across a much longer reach. But it's the same amount of data, which is what drives our PAM DSP business today, just to give you a sense of how large this can be. And you want the same bandwidth, right? So this needs to be 1.6T, which is why we pulled up our road map, came out with our 1.6T DCI module pretty soon. It's going to be sampling second half of this year with a 2-nanometer coherent DSP first to market. So pretty unique position there.

So I think we have the mayor coming in next, and I'm getting the signal from our organizers that it's time for us to wrap up. So that will have to be the last word.

Great. Thanks, Mark.

Willem, Ashish, thank you so much for joining us today and for all the great insights.

Thanks, Mark. Appreciate it.

Appreciate it. Thank you.

Good afternoon, and welcome to Marvell Technologies First Quarter Fiscal Year 2027 Earnings Conference Call. [Operator Instructions] Please note, this event is being recorded. I would now like to turn the conference over to Mr. Ashis Saran, Senior Vice President of Investor Relations. Thank you. You may begin.

Thank you, and good afternoon, everyone. Welcome to Marvell's First Fiscal Quarter 2027 Earnings Call. Joining me today are Matt Murphy, Marvell's Chairman and CEO; Willem Meintjes, CFO; Chris Koopmans, President and COO; and Sandeep Bharathi, President Data Center Group.

Let me remind everyone that certain comments may today include forward-looking statements, which are subject to significant risks and uncertainties that could cause our actual results to differ materially from management's current expectations. Please review the cautionary statements and risk factors contained in our earnings press release, which we filed with the SEC today and posted on our website as well as our most recent 8-K, 10-K, 10-Q and other documents filed by us from time to time with the SEC.

We do not intend to update our forward-looking statements. During our call today, we will refer to certain non-GAAP financial measures. A reconciliation between our GAAP and non-GAAP financial measures is available in our earnings press release.

Let me now turn the call over to Matt for his comments on the quarter. Matt?

Yes. Thanks, Ashish, and good afternoon, everyone. For the first quarter of fiscal 2027, Marvell delivered record revenue of $2.418 billion, reflecting 9% sequential and 28% year-over-year growth. Revenue exceeded the midpoint of guidance, and as a result, non-GAAP earnings per share of $0.80 exceeded the midpoint of guidance by $0.01. We are seeing strong demand and exceptional bookings across our entire data center portfolio. .

This robust demand is reflected in our guidance for the second quarter of fiscal 2027, where we expect total company revenue to grow 12% sequentially and 35% year-over-year at the midpoint to $2.7 billion. On our earnings call last quarter, we indicated that beginning in Q2, we expected quarterly revenue growth throughout fiscal 2027 to trend in the high single-digit range sequentially on a percentage basis. Q4 revenue exiting the fiscal year at approximately $3 billion.

We are now guiding Q2 revenue to grow double digits sequentially, and we expect Q3 and Q4 revenue to also grow by at least 10% sequentially. As a result, we now expect $3 billion in quarterly revenue in Q3, one full quarter ahead of our prior outlook. We also continue to expect year-over-year revenue growth rates to accelerate each quarter throughout fiscal 2027, reaching approximately 50% by Q4.

As a result, we now expect overall Marvell revenue in fiscal 2027 to grow approximately 40% year-over-year to nearly $11.5 billion. The increase in our revenue outlook continues to be driven by our data center business, which we now expect to grow approximately 50% this fiscal year. Notably, we expect our interconnect business to grow more than 70% year-over-year, well above our prior expectation of 50% growth.

I will provide additional color on our interconnect business later in today's call. For our communications and other end market, we continue to expect revenue growth of approximately 10% in fiscal 2027. Now looking ahead to fiscal 2028, while we are planning for the rate of cloud CapEx growth to moderate into the 30 plus -- the 30% plus range, we expect strong data center revenue growth for Marvell to continue.

We expect our interconnect business to continue to outpace cloud CapEx growth, reflecting strong 1.6T demand from scale-out networking and more meaningful contributions from scale up and scale across networking. We now expect our custom business to more than double year-over-year in fiscal 2028, higher than our prior outlook and expect our Ethernet switching business to continue ramping.

As a result, we expect data center revenue in fiscal 2028 to grow approximately 55% year-over-year, accelerating from fiscal 2027s projected growth rate. For our communications end market, we continue to expect low single-digit percentage revenue growth in fiscal 2028 consistent with our prior view. In aggregate, we now expect overall company revenue to grow approximately 45% in fiscal 2028, off a higher fiscal 2027 base.

As a result, we now expect Marvell's fiscal 2028 revenue to reach approximately $16.5 billion, roughly $1.5 billion higher than the outlook we provided on our earnings call last quarter. This outlook is supported by demand trends we are seeing today and by programs already in execution. Our investments in securing supply are paying off, enabling us to scale the business every quarter.

As we move through the fiscal year, we expect to remain closely aligned with our customers as they continue investing aggressively in AI infrastructure. Now let me turn to the expanded partnership we announced with NVIDIA, which reflects the growing importance of high-speed connectivity, optical interconnect and accelerated infrastructure in scaling AI.

The collaboration connects Marvell's custom silicon and optical networking capabilities directly into the massive NVIDIA ecosystem to help build scalable, highly efficient AI data centers and telecommunications networks. There are 3 core pillars of this exciting announcement. First is our Optics partnership. Marvell has long been a key supplier of DSPs, TIAs and drivers, and we are now extending this relationship to collaborate on silicon photonics technology, which is expected to be a key enabler of scale-up networking.

Second, NVLink Fusion integration. This allows Marvell to build custom chips and networking semiconductors that can seamlessly interface with NVIDIA infrastructure. It increases choice for hyperscalers who will now have complete flexibility to mix and match custom emerging capabilities across their platforms, with Marvell uniquely providing the bridge between these 2 architectures.

We expect this to create new market opportunities for both Marvell and NVIDIA going forward. The third is AI-RAN. Marvell will enhance its existing OCTEON base station processors to work directly with NVIDIA GPUs, integrating AI with wireless infrastructure on a single, software-defined computing platform. This will enable telecommunications operators to run both 5G and 6G radio workloads and high-performance AI applications concurrently on the same hardware.

Since the announcement, both teams are off to the races, and we are working closely together to realize the benefits of this collaboration. We deeply appreciate the partnership and the investment from NVIDIA. Okay. Let me provide more color on our current business, beginning with data center.

In our data center end market, we delivered record first quarter revenue of $1.83 billion, representing 11% sequential growth and 27% year-over-year growth. We achieved sequential and year-over-year growth across multiple product lines, including optical interconnect, custom silicon and switching. Looking ahead to the second quarter, we expect data center revenue growth to accelerate into the mid- to high teens sequentially on a percentage basis and into the mid-40% range year-over-year.

Our networking products, including interconnect and switching, are driving strong revenue growth as networking becomes increasingly critical with each new generation of AI infrastructure. Now in the early stages of generative AI, the primary focus was on addressing compute and memory bottlenecks. As more complex architectures such as reasoning modules and mixture of experts have begun to deploy, the role of networking has become significantly more important and this is the key driver of the increased demand we are seeing today for our scale-out networking products.

Now what is completely in front of us is the massive expansion expected in scale-up networks as these domains become significantly larger, requiring high [ radix ], low latency switches as well as high bandwidth optical interconnects. In addition, these new AI models are also driving innovation and memory architecture, which we expect will benefit our XPU attach business. We expect the emergence of agentic AI to further supercharge demand for our scale out, scale up in XPU attach businesses.

In agentic AI, a single user request may require agents to query AI models many times rather than just once as in traditional one-shot inferencing. These queries may also be routed to different parts of the AI cluster to complete a single task. This substantially increases the volume of data traffic that must be transmitted and switched with very low latency across longer reaches as well as the amount of memory required. Agentic AI is also expected to drive a significant increase in the number of CPUs deployed in AI infrastructure. More CPUs require more NICs, PCIe switches and retimers as well as greater bandwidth and CPU-centric front-end networks.

As a result, we believe agentic AI can provide another significant tailwind for our interconnect switching and XPU attach franchises. It's increasingly clear that optics is the future of data center connectivity, and we continue to invest aggressively in our technology platform to extend our leadership in this rapidly expanding market. Our late addition is the acquisition of Polariton, a developer of high-speed, low-power plasmonic-based silicon photonics devices. Plasmonic offer meaningful advantages over traditional silicon photonics by enabling substantially higher modulator bandwidth which is critical for support faster optical transmission speeds.

Polariton has already demonstrated Plasmonic modulator bandwidth exceeding 1 terahertz up to 10x higher than current silicon photonics and thin-film lithium niobate-based solutions. We are excited to incorporate this breakthrough technology into our DCI and coherent light road maps, extending our technology platform to 3.2t and beyond.

Let me now pivot back to the near term and discuss trends we are seeing across both our established data center businesses and our newer growth initiatives. I'll organize the discussion into 3 categories: interconnect switching and custom. I'll start with interConnect, which represents the largest portion of our data center business. Demand for our interconnect products continues to accelerate. And as a result, we have increased our fiscal 2027 revenue growth expectations for this business to more than 70% year-over-year.

Interconnect is also a major driver of the higher fiscal 2028 company revenue outlook we provided today. We are benefiting from our leadership position across the industry's broadest portfolio high-speed connectivity solutions, spanning scale out, scale across and scale up networking. Within our scale-out PAM franchise, demand continues to strengthen for our 800 gig products while our 200 gig per lane 1.6T solutions are ramping quickly this fiscal year following their production launch in the second half of fiscal 2026.

We expect 1.6T revenue to take another substantial step up in fiscal 2028. We continue to benefit from the first-to-market cadence we have maintained across successive PAM4 generations. We also expect to maintain leadership into the next PAM4 generation with 400 gig per lane technology, which we demonstrated first at the Optical Fiber Conference in April 2025.

In addition to our DSP franchise, we have also built a formidable position in broadband, analog, TIAs and drivers. This business is scaling rapidly, and we expect quarterly revenue of TIAs and drivers to exceed a $1 billion annualized run rate in the next few quarters. To support campus-wide data center architectures requiring longer reach than traditional PAM solutions. We were the first to introduce coherent light products to the market. These solutions are optimized for applications spanning 2 to 20 kilometers within an extremely low power envelope as compared to traditional coherent DSPs.

Over time, as speeds continue to rise, we expect coherent light to penetrate deeper inside data centers, complementing PAM-based solutions for shorter-reach applications. We've already begun shipping the first generation of our coherent light 200 gig per lane 1.6T products. We are now introducing next-generation coherent light products featuring integrated MACsec security as well as higher speed capabilities.

Turning to DCI. This market is undergoing a major architectural transition driven by the emergence of scale across networks, which we believe will significantly expand the opportunity for pluggable DCI modules over the next several years. Marvell pioneered the pluggable DCI market, where the original use case was driven by hyperscalers replacing public WAN connections for intersite connectivity using pluggable modules, with traffic between data centers originating primarily from traditional front-end networks.

This has become a highly successful business for Marvell and today, we shipped DCI solutions to all 5 major U.S. hyperscalers. What is now changing is the push to build significantly larger AI clusters, which increasingly must span multiple data centers due to power and space constraints. In these architectures, the back-end AI network must also extend between the data centers, creating the scale across use case where massive amounts of data move continuously between XPUs during AI workload processing.

Aggregate bandwidth requirements for scale across networks are projected to be more than 10x higher than those of current front and DCI networks. As a result, industry forecasts project a pluggable DCI TAM to grow significantly by rapidly increasing speeds and rising future complexity, including integrated MACsec security. while traditional DCI networks today primarily deploy 400-gig solutions and are now transitioning to 800 gig, scale across architectures are expected to rapidly adopt 1.6T connectivity.

Marvell is exceptionally well positioned to lead this transition with the industry's first secure 1.6T ZR and ZR+ DCI modules powered by our new 2-nanometer coherent DSP announced earlier this year. These modules are expected to begin sampling this year. This position Marvell to extend our technology leadership into the emerging scale across market, supported by our proven expertise in high-volume manufacturing of these highly specialized and complex modules.

Our leadership position here is translating into strong revenue momentum for our DCI module business, giving us line of sight to a $1 billion annualized revenue during fiscal 2028. This would represent approximately double the revenue we achieved in fiscal 2026 when the business generated roughly $500 million in revenue. As scale of cross deployments become a larger portion of the market, we expect growth in our DCI business to accelerate further. Let me now transition to scale-up optics.

Scale-up Interconnect represents one of the newest and most strategically important opportunities emerging in AI infrastructure. Marvell is uniquely positioned to enable both NPO and CPO implementations with a broad silicon photonics platform spanning all 3 mainstream modulator technologies, including MZM, EAM and MRM, fully supported by our market-leading broadband analog TIAs and drivers. We are also investing in emerging approaches such as micro LED and [ micro Voxel-based ] solutions. Marvell has already shipped more than 1 million DCI modules powered by our silicon photonics over the past decade.

Across 4 generations of silicon photonics deployments, we have accumulated more than 15 billion hours of field data with demonstrated world-class reliability. We have leveraged this experience in developing our silicon photonics-based light engines, and we are deeply engaged with multiple Tier 1 customers with our third generation 6.4T light engine for NPO and CPO implementations. Our acquisition of Celestial AI added photonic fabric technology, including EAM modulators in the industry's leading low-power analog SerDes. The solution has already been selected by a Tier 1 hyperscaler for its next generation of XPU scale-up networks.

The full strength of Marvell's engineering and operations organization is focused on bringing CELESTIAL's first-generation chiplet into high-volume manufacturing. We are also making significant progress with MRM-based scale-up interconnect solutions. We completed our MRM device demonstrations last year and continue to continue -- I'm sorry, continue to collaborate closely with TSMC on its Coop platform.

We believe scale-up interconnect represents a massive new TAM that will likely be served by multiple photonic technologies and architectures, and we are investing aggressively to establish leadership across all of them. We are seeing market adoption accelerate from multiple CPO and NPO engagements. And as a result, we expect our scale-up optics business to ramp significantly next fiscal year, with revenue forecasted to more than double our prior outlook of approximately $150 million, which was based at that time solely on Celestial AI.

Turning to data center switching. We continue to benefit from sustained demand for our 12.8T and a strong ramp up of our next-generation 51.2 T switches for scale-out networking. We are seeing strong engagement for both existing and new customers for our 51.2T platform as well as our 100 platform which we believe delivers industry-leading power efficiency and low latency, attributes that are increasingly critical for AI infrastructure.

And our engineering teams are already executing a road map towards 200T Ethernet switching and beyond. Given this momentum, we expect scale-out switch revenue in fiscal 2027 to exceed $600 million, doubling from fiscal 2026, and we currently see the business tracking to more than $1 billion in annualized revenue in fiscal 2028. The Scale up switching is an emerging market where we have significantly increased our investment, both organically and through the acquisition of XConn which substantially expanded our team and capabilities. While some customers are currently deploying PCIe switches for their current generation of scaleup networking, the ratings and bandwidth limitations of PCIe are expected to drive a rapid transition towards purpose-built, large ratings, high-bandwidth UALink, ESUN and NVLink solutions.

Marvell is uniquely positioned to support all of these scale-up protocols through our internally developed UAL and ESUN switches as well as our expanded partnership with NVIDIA around NVLink Fusion. We currently have multiple engagements with Tier 1 customers for our scale-up switch portfolio. Given the size of the scale of TAM, each of these engagements represent a multibillion-dollar lifetime revenue opportunity.

We believe we are exceptionally well positioned in this market, leveraging decades of extensive experience developing large reticle size switch silicon as well as our in-house best-in-class high-performance series technology. Now let me touch on a few additional growth opportunities in data center.

In the AEC market, we are seeing strong interest in our golden cable program, and we have already secured design wins with 3 Tier 1 U.S. hyperscalers along with several other customers. We are also seeing strong traction for our retimer products. Both AECs and retimers are now ramping, and we expect combined revenue to more than double year-over-year in fiscal 2027 and continue growing rapidly in fiscal 2028.

The acquisition of XConn also advanced PCIE and CXL switch solutions to our portfolio, and we are seeing strong interest in our PCIe Gen 6 and CXL 3.1 solutions. Marvell is well positioned in both of these markets to provide customers with complete end-to-end solutions and reference designs consisting of our PCIe switches paired with retimers as well as our CXL switches paired with our memory expanders.

Okay. Turning now to our custom business. Custom revenue remains on track to grow more than 20% year-over-year in fiscal 2027, led by our flagship XPU program, which we expect to drive multiple years of growth across multiple generations. Several XPU attach programs are also ramping in fiscal 2027, including our CXL and NIC products. Looking ahead to fiscal 2028, we now expect custom revenue to more than double year-over-year, which is higher than our prior outlook. The growth is expected to be driven by 3 primary factors, including: first, continued growth from our existing custom programs, including our flagship XPU. Second, over 10 XPU attach programs reaching high volume -- higher production volumes with demand continuing to exceed prior forecasts, particularly in NIC and CXL memory attach use cases driven by increasing inference KV caching requirements.

And third, the ramp of our new Tier 1 XPU program into volume production. This program continues to progress very well through development, and we already have firm requirements in place for all of next fiscal year. Since last quarter, we have won several new designs as customers continue to expand their adoption of custom silicon. We expect these new sockets to begin contributing incremental revenue following their typical development cycle of approximately 2 years.

The level of custom engagement with key customers remains unprecedented, and we continue to be deeply involved in a broad set of significant additional opportunities. We remain confident in achieving our target model for our custom business to deliver on over $10 billion in revenue in fiscal 2029. Turning to our communications and other end markets. We delivered first quarter revenue of $585 million, up 3% sequentially and 29% year-over-year.

For the second quarter, we expect revenue to decline in the mid-single-digit range sequentially on a percentage basis, while growing in the high single-digit range year-over-year on a percentage basis. The communications and other end market has now largely recovered from inventory corrections at our customers. And going forward, we expect revenue in this end market to broadly reflect the underlying trends in our enterprise networking carrier and consumer businesses.

In summary, our business continues to accelerate, and we have increased our revenue outlook multiple times over the past several quarters. Today, we are again raising our outlook, increasing our fiscal 2027 revenue forecast by more than $0.5 billion, and our fiscal 2028 outlook by approximately $1.5 billion versus the projections we provided last quarter.

Our data center revenue grew 46% year-over-year in fiscal 2026, and we are now projecting growth to accelerate to approximately 50% in fiscal 2027 and accelerate again to 55% in fiscal 2028. Our customers continue to signal robust demand, not only for this year but for the next several years. Our results and outlook reinforce our confidence that Marvell is in a strong multiyear growth cycle with substantial runway ahead.

Now today, we sit here in a unique position to simultaneously: one, drive incredibly strong top line growth; two, increase R&D investments strategically in the highest growth AI opportunities while continuing to drive operating leverage. Three, make necessary capacity investments to fuel the next wave of growth; and four, continued strong capital returns to shareholders. The Marvell team is firing on all cylinders with strong momentum expected to continue across the business.

We have built a well-diversified company anchored by multiple large existing franchises and complemented by several emerging growth engines. I look forward to updating you on our progress as we continue this exciting journey as a key enabler of next-generation AI infrastructure.

With that, I'll turn the call over to Willem for more details on our recent results and outlook.

Thank you, Matt, and good afternoon, everyone. Let me start with our financial results for the first quarter of fiscal 2027. Revenue was $2.418 billion, growing 28% year-over-year and 9% sequentially. Data Center was our largest end market, contributing 76% of total revenue. GAAP gross margin was 52.1%, non-GAAP gross margin was 58.9%.

Moving on to operating expenses. GAAP operating expenses were $921 million, including stock-based compensation, amortization of acquired intangible assets, restructuring costs and acquisition-related costs. Non-GAAP operating expenses came in at $577 million. Our GAAP operating margin was 14%, while our non-GAAP operating margin was 35%. For the first quarter, GAAP earnings per diluted share was $0.04 lower than our guidance, reflecting the impact of purchase accounting for the Celestial AI and XConn acquisitions and the related earn-out obligation.

We expect this to normalize in the second quarter as reflected in our strong GAAP net income guidance. We have now delivered 6 consecutive quarters of positive GAAP net income and expect to continue to drive strong GAAP profitability going forward. Non-GAAP earnings per diluted share was [ $0.80 ] above the midpoint of guidance, reflecting year-over-year growth of 29%.

Now turning to our cash flow and balance sheet. In the first quarter, cash flow from operations was a record $639 million. Our inventory at the end of the fourth quarter was $1.4 billion, almost flat from the prior quarter. During the quarter, we repurchased $200 million of our stock through our ongoing capital return program and returned $54 million to shareholders through cash dividends in the quarter.

As of the end of the first quarter, our total debt was $4.96 billion, with a gross debt-to-EBITDA ratio of 1.44x and a net debt-to-EBITDA ratio of 0.32x.

Turning to our guidance for the second quarter of fiscal 2027. We are forecasting revenue to be in the range of $2.7 billion, plus or minus 5%. We expect our GAAP gross margin to be between 52.1% and 53.1%. We expect our non-GAAP gross margin to be between 58.25% and 59.25%. Looking forward, we anticipate that the overall level of revenue and product mix will remain key determinants of our gross margin in any given quarter.

We project our GAAP operating expenses to be approximately $960 million. We anticipate our non-GAAP operating expenses to be approximately $600 million in the second quarter. We expect our GAAP other income and expense, including interest on our debt to be an expense of approximately $68 million. We expect our non-GAAP other income and expense, including interest on our debt to be an expense of approximately $35 million. We expect a non-GAAP tax rate of 11%. We expect our basic weighted average shares outstanding to be 899 million and our diluted weighted average shares outstanding to be 915 million. The increase from the prior quarter reflects the full impact of shares issued for the Celestial AI and XConn acquisitions as well as the shares issued as part of the NVIDIA investment.

We anticipate GAAP earnings per diluted share in the range of $0.32 to $0.42. We expect non-GAAP earnings per diluted share in the range of $0.88 to $0.98. As we look ahead, we intend to continue to invest in growing our business while driving operating leverage. For fiscal 2027, we expect non-GAAP earnings expense of approximately $2.45 billion, which includes the acquisition of Celestial AI and XConn, both of which closed in the first quarter of this fiscal year.

For fiscal 2028, we expect non-GAAP operating expense to grow year-over-year approximately in the mid- to high teens on a percentage basis. This is significantly below the 45% revenue growth outlook Matt provided in his remarks for that year. As a result, we expect to achieve the upper end of our target operating margin model of 38% to 40% as we progress through fiscal 2028.

Based on the [indiscernible] we have secured and our confidence in sustained customer demand, we are aggressively locking in additional capacity to ensure our growth. We are following the same successful playbook we established during the last major supply crunch, which includes sharing our long-term demand outlook with key suppliers and making strategic prepayments to ensure capacity. This approach has served us well, enabling Marvell to scale revenue significantly during a period when the broader industry has remained supply constrained.

We are forecasting approximately $1 billion in prepayments during this fiscal year with the first payments beginning in the second quarter. These prepayments will be applied against future material purchases. We expect to fund these prepayments through our strong balance sheet and robust operating cash flow generation. In parallel, we plan to continue to repurchase shares to manage dilution. I'm very pleased with our execution driving strong revenue growth and operating leverage as well as robust cash flow generation and ongoing stock buybacks. We are looking forward to continuing to deliver strong earnings growth to our stockholders. With that, we are ready to start our Q&A session. Operator, please open the line and announce Q&A instructions. Thank you.

[Operator Instructions] Our first question comes from the line of Vivek Arya with Bank of America Securities.

Matt, towards the end of your presentation, you mentioned something along the lines of your custom XPU is on target to hit $10 billion in fiscal '29. I just wanted to make sure that we heard that correctly because that would mean that fiscal '28 if your XPU is a little over $4 billion and then it gets to $10 billion. So that's an increase of at least $5 billion to $6 billion year-on-year in sales.

So I just wanted to make sure that I heard that. And then Matt, at what point will you feel more comfortable talking about that large customer and the progress in this new program? Do you expect to be exclusive in this? Because they have a really large CapEx profile. So when do you think investors should start to give Marvell more credit for that new XPU program that should start next year?

Yes. Thanks, Vivek. Yes. So yes, you heard it right. And just for context on the $10 billion plus for fiscal '29, the context for that is, back in April 2024 we set long-term targets out through calendar '28, which would be our fiscal '29.

At that time, we had identified a data center custom overall market, overall custom silicon market, which would generate about $8 billion revenue for us was sort of what you -- if you assume 20% of our market sizing at that time for fiscal '29. And at the -- at our custom silicon event last summer in June of 2025, we then said basically the TAM is bigger. So the implied again, if we achieved our share targets in fiscal '29, would indicate something over $10 billion.

It was like a $55 billion TAM. Take 20% on that. That's $11 billion. So call it in that range. And yes, we're still tracking to it. That's a key part of our assumption. Two years ago, it looked like a very steep hill to climb. We're clearly making progress there. And yes, between our existing programs, our new ramp and just a plethora of XPU attached programs, all of which have sized up significantly since we won them.

We definitely see line of sight to hit those targets. And again, this is just updating investors along the way about how we're progressing. On your second question, program remains on track. I think we're hitting all of our milestones. We -- and we see for next year, I think as I stated in my prepared remarks, kind of across the board on all of the custom programs we have, we're seeing indications of greater need for demand. We we had judged that went down pretty significantly.

And I think as we progress through the year to answer your questions, I think we'll all gain confidence in magnitude that ramp, but it's on track and it's a key part of our plan for next year, but it's not the only piece. It's probably still about 1/3 of the total growth we're expecting in our custom business next year. But yes, it's on a very strong trajectory for that from fiscal '27 to '28 to '29 and beyond on the custom side overall, which includes XPU and XPU attach.

Our next question comes from the line of Harlan Sur with JPMorgan.

Matt, with the aggressive evolution of the inference in compute workloads, this is really opened up a lot of opportunity for these XPU offload engines, like an LPU processor, an STX processor, right? That's what motivated NVIDIA, for example, to acquire that start-up Groq, right, that has this unique SRAM-based memory architecture that enabled them to design a very efficient sort of low latency influencing engine called LPU.

I seem to recall that the Marvell team actually designed the start-up's first generation LPU inferencing chip. And we all know that Marvell has always focused on memory and in particular, SRAM-based IP, like highest density, lowest power. In fact, I think you guys bought the first -- industry's first 2-nanometer SRAM IP to the market last year. Is the Marvell team leveraging this differentiation? Are you seeing more interest in SRAM-based XPU offload ASICs and do you already have XPU attached design wins for SRAM-based offload architectures? .

Yes. Thanks for the question, Harlan. Yes. So first, you definitely pay attention. We talked about this actually quite extensively in June of 2025 at our custom silicon event. We had a whole dedicated presentation actually on our best-in-class SRAM design capability and IP.

And that's, by the way, a legacy in a history investment area that goes all the way back to the Avera acquisition and prior to that, the team being GLOBALFOUNDRIES and IBM. So there's a long legacy there. That technology has evolved from networking products into AI products and you can see that trend in the market as inferencing is happening. So that continues to be a key part of our IP portfolio sure as we go win designs. And it's one of the reasons why we are winning our designs in XPU attach. But it's part of the bigger strategy really, which is I think having the full suite of solutions, right, from advanced packaging, best-in-class high-speed I/O and just the ability to really dive in and get the develop these solutions with our customers in very aggressive time to market time frames, leveraging our manufacturing expertise and our capacity, et cetera. So it's 1 piece of the puzzle, but it is an important piece for sure. And we definitely see that as a trend out there, Harlan, and we'll keep you updated.

Our next question comes from the line of Timothy Arcuri with UBS.

I wanted to ask about the breadth of the customer base. I know we've talked about the existing XPU you have. There's this new customer on the XPU and then you do have some XPU Attach with a third customer as well.

But there's some speculation that you actually might be moving into the compute TAM, and that's by far the biggest custom compute wallet out there. So -- so is that included in the forecast? Or would that be incremental to that? I'm just -- I guess I'm just trying to ask about the breadth of the customer base because I know you did -- you mentioned about the engagement broadening. So wondering if you could talk about that.

Yes. Well, I think a couple of things. The first is, and we've said this for some time, we have custom engagements across the board at all the U.S. hyperscalers. And we've had that for some time, some XPU, some XPU and XPU attach and some just XPU Attach. That's -- and that's part of the AI custom silicon event we did last summer is indicating kind of that broad pipeline of opportunities and sockets. So everything we've laid out, which is the greater than 20% growth this year, more than doubling next year and then still having line of sight to our long-term targets. That's all based on the designs we've already won and locked and even going back really the last summer.

Now if -- some of these programs, again, the timing on them, the newer ones we won since last summer, if you can get them done in like a 2-year time frame, then you might get some contribution from that in fiscal '29, but that's not needed. That -- Think of that as like an insurance policy.

Great. If some of those [ hit ] then if you're worried about some of the existing programs not quite getting there. So when we look at the whole picture, net-net, we feel very comfortable with the trajectory of our custom business. and it's not requiring anything incremental. And I would just say broadly across the board, our technology platform is very competitive, and we're out there competing every day for the most important sockets in the world.

And again, those would contribute later. But at our 2-nanometer platform and then beyond, our road map is very compelling, and it's only gotten -- it's only strengthened since our custom silicon event last summer. So this business is inflecting, and it's definitely on the right track. And again, I highlighted a number of reasons why that is with Harlan's question, but I'll just tell you that what we've really seen, and it's been pretty pronounced, I would say, in the last 6 months or so is that -- the performance of our high-speed I/O, our SerDes performance, our die to die and our ability to integrate that very densely in XPU, switching applications, attach applications.

is the amount of activity we're seeing on demand for that, and I think people realizing that really, it's us and just maybe a couple of other people that can do this at the level that's needed for this level of performance integration, it's driving a whole new set of opportunities, quite frankly. So I -- again, design activities through the roof at Marvell across the board, but including on custom silicon.

Thank you. Our next question comes from the line of Aaron Rakers with Wells Fargo.

This is Michael [indiscernible] on behalf of Aaron. I think you mentioned in your script that you had, I think, new XPU or custom, I guess, [indiscernible]. Could you just provide any additional color on, I guess, for those XPUs or expat or maybe the type accelerator or just type of chip, -- that would be really helpful.

Yes. No additional details at this time. I think it's just another data point we're trying to give people that based on the 50-plus type of opportunities we outlined last June. We continue to close on those. That opportunity pipeline continues to grow in terms of sockets and dollars, by the way. I think every sort of program we looked at a year ago is larger when we look a year later. But no additional details at this time. But at the right time, we'll do a more comprehensive look back and update you on our progress. But right now, at least from the current business we have, the revenue line is definitely moving in the right direction and strong validation for the capabilities that we have.

Our next question comes from the line of Blayne Curtis with Jefferies.

Matt, I wanted to ask about specifically the CECL opportunity alongside accelerators. How real is that opportunity? Is there a way to think about the content per accelerator for that?

Well, yes, it's a very real opportunity. I think we've -- we've -- one is we engaged in this business actually a few years back, and this was when the applications were really driven by sort of traditional server dynamics for x86 compute and that's obviously vectored over into AI.

I think we had a plan in this area, which we've talked about getting that sort of custom line item of XPO attach alone over $1 billion in revenue in the next couple of years. That continues to expand both because we continue to expand the customer base, which is very compelling. But also, I think the concerns around the memory cycle we're in are driving additional adoption of CXL-based design.

So it's sort of no secret at this point that I think the memory architectures are [indiscernible] and critical on how people think about deploying their next-generation infrastructure. And I think us coming in with these very proven solutions now for CXL are really playing in our favor. So I think it's just the trend line continues and just to continue to size up both because of the CapEx, higher penetration due to some of the memory issues and then just more and more of these solutions moving into inferencing. So those are some of the trends that we're seeing, Blayne.

Our next question comes from the line of Chris Caso with Wolfe Research.

Question is about some more color on how you're addressing capacity constraints right now. You spoke about some of the prepayments. And so I mean, first part of the question is, how you're managing to get that additional capacity. And then as a follow-on to that, with the guidance. The increase in guidance, is that a function of you managing to get more capacity out of your customer -- out of your supply base -- or is it really more of a factor of becoming more comfortable with the forecast your customers already given you know that you didn't guide us to everything your customers had put in their forecast in the first place.

Yes. Thanks. I'll hand this one over to Chris, who's been knee deep in this, but I just to give a shout out to our team. Our supply chain operations has been doing an outstanding job, and our suppliers have been doing an outstanding job continue to react to the changes upward in demand. We've been very pleased with that, and we are in great shape overall to deliver on what we just talked about.

But I'll have Chris give some more color on how we're going about that.

Sure. Thanks, Matt. So yes, I think as Matt mentioned, I've been dealing with the operations side since 2020, 2021, and I don't think we've been in an unconstrained environment since then, everything that touches AI has been constrained basically since the beginning of this.

And ultimately, the way we've been able to manage this is by building very tight relationships with a small number of key suppliers, giving them a 5-year forecast of what we need and hitting what we need each time along the way and doing what we said we were going to do. And that goes a long way towards getting which you need when you take everything that you need as you go along.

And we work very closely with all of our key customers and with our key suppliers to give them that forecast. And yes, there's really only a handful of companies that are driving this AI infrastructure TAM build-out and in order to do that, part of that is the prepayments that William mentioned that we're making with our suppliers to back our forecast with confidence and cash. And ultimately, that's what helps us deliver the revenue capabilities we have.

Our next question comes from the line of Ross Seymore with Deutsche Bank.

I want to go to the interconnect side of things. Matt, you talked about the growth rate going to, I think, over 70% this year. If I recall right, it was at the beginning of this year, 30% and then 50% and now 70%. So it's clearly accelerating. I guess the question is, why would you think other than just conservatism that, that would slow to kind of closer to, but still above the hyperscaler CapEx rate next year? Is that just conservatism? Or because everything that you rattled off before about the DCI side of things, AEC, 1.6T, the scale up, et cetera, et cetera, all sounds like those are still very, very strong tailwinds. So I just wanted to get a little more color as to what you're thinking in fiscal '28 for that business.

Yes. I think the fact is this is where we are right now, Ross, and very encouraged to see as we've progressed, as you mentioned, from where we were starting probably around last September to now with the growth rate continuing to inch up as CapEx has gone up and also the attach rate of our interconnect products have gone up. The new initiatives, those businesses are ramping. It's definitely I mean we've had a lot of success across the board here, but this has been the star of the show here. Look, I think when you look out to next year, I think this is where we are today. But if you start building a bottoms-up model, you can see that there's definitely a possibility for a lot of upward bias because our traditional business in DSPs, right, obviously, we talked about a big step-up next year. in 1.6T. That's higher content.

You have DCI ramping. You have the new initiatives, things like retimers and ACs, but also you have scale up optics, which we're effectively calling at this point to be about $300 million, which is true NPO and CPO-based solutions this is like the beginning of a major growth cycle for us. So I think there's a lot of optionality is what I would say at the moment, Ross. But I think today, net-net, you look overall, [ 16.5 ] where we're comfortable overall, but I think there's upward bias for sure. The trends continue .

Our next question comes from the line of Tore Svanberg with Stifel.

Congrats on the record quarter. Matt, I was hoping to zoom in a little bit more on the interconnect business and specifically scale up. Obviously, a lot of that is in front of you. But there's also a lot of different dynamics there, right, whether it's copper, optical, NPO, CPO, copper. What has been surprising to you the last few months as far as what 2 or 3 products are you seeing the bigger upside from? Because obviously, it's a very dynamic market. So given your new growth forecast here, there must be a few that are worth highlighting. .

Yes. Well, I think the good -- I'll start at the top. The good news is, as you said, you just rattled off a number of different technologies, and there's an equal number more you didn't mention. We're in all of them. And I think that's one of the unique advantages that we bring to the table is we have the absolute broadest range of connectivity and scale up and scale out solutions for interconnect in the industry.

And I think, as I was indicating to Ross, I think everything's got an upward bias to it. The area where I think the most intense activity has sort of emerged -- and I think it's just -- it's been just a home run, getting the Celestial team in here and combining them with our own very, very strong capable optics team. that combined team, we're able to go into customers now and outline full end-to-end solutions from XPU to switch with any type of interface and optical or copper connection between that the customer can envision and really optimize around that.

And so that's why we mentioned we're seeing for next year, not just the Celestial products ramping on scale up optics, but also Marvell products as well. And I think that's an area that could [indiscernible], but that upwards, sorry. But that's in the context of us presenting a very comprehensive total system solution.

Because in the end, especially for this first generation, you're going to want to have these solutions bookended, Tore, and you're going to want to have the XPU integrate the same photonic element as the switch side. And we are able to walk in and show our capability and switches, our ability to integrate into XPUs or build the XPU and then all the optics in between. So it's a very, very powerful combination. And I think the end-to-end is getting a lot of attention. And the fact is we have the proven technology. It's not a PowerPoint. It's not a concept. I mean, we have 15 billion hours as an example, of device data on 4 generations of silicon photonics already.

We have 224 gig SerDes in production. We have die-to-die in production. We built switches for multiple generations that are reticle sized with high yield. These are -- this is like the most complex stuff you're going to do in the semiconductor industry. And we're able to simultaneously do it all and bring it to our customers. So if I were to point out a trend, which is kind of what you're getting at, it's that trend right now is how do we help our customers enable their scale-up network for the future.

And this is like Gen 1 ground zero, I mean, there's a lot of room to go here. That's what I'd say the most sort of intense discussions we're having in. And we're very much investing to win here. That's why you see us, in some cases, betting on multiple standards or multiple technologies. We don't want to miss out, and we will pivot at the right time. We'll follow our customers, we'll follow the market. But today, we have a little bit of a lag sheet, and it's very refreshing to our customers because we don't just go in there with our agenda based on the 1 piece that we have and say, well, this is what we have, this is what we can offer you. We go in with here's the technology that you need, and it's real engineering. It's real proven silicon. It's real proof of concept that we can show. So I think more to come on this one.

Our next question comes from the line of Srini Pajjuri with RBC Capital Markets.

Matt, I want to zoom in on the switching side. I think you talked about scale out switching doubling this year. and potentially reaching $1 billion annualized run rate next year. I know it's a large market and you have relatively small share in that market. But as we go from scale out to scale up, you mentioned 3 different, I guess, [ ESUN ] NVLINK and UAVLink. So it seems to me that, that could be even larger market. And obviously, it's all greenfield market for you. So I'm just curious as to how you're thinking about that opportunity in '28 and beyond.

Yes. Thanks, Srini. Yes, you're right. On the scale-out side, it's a large and established market. We're an emerging company there. I would just say though, I think it's a huge milestone to look out to next year and have line of sight to $1 billion of revenue here when, if I go back 6 years ago, when we were able to bring the Innovium team in that was -- it was effectively a pre-revenue company or very little with, at that time, kind of line of sight to $100 million, $150 million of revenue. So that's gone really well if you look over the last 5 years, right, and just kind of where we're heading. And a lot of those assets and capabilities that we got from that are now directly being leveraged into scale-up networking particularly on the -- obviously, on the Ethernet side on ESUN. So that, in our mind, is a bigger opportunity from the standpoint that the market share isn't established yet -- and that's what I was referring to earlier in Tore's question, there's so many important architectural discussions going on, but it's not just the switch, and it really requires kind of a very broad set of capabilities and track record for customers to bet on you here.

And I think the stakes are higher because we do continue to see the adoption of CPO and NPO technologies being much more robust and sticky inside the CLF networks, which is just a lot more TAM for Marvell as well. So you've got switch ASPs that are roughly the same, but you've got a whole new emerging market segment where nobody's might establish the leadership position there. And we think based on assets and our strategy, we should do really well there. That will be -- all the numbers I've rattled off to everybody here, which is a lot, I get it, but the concept is trying to give people visibility on the different pieces of Marvell that you might not think of.

But on the scale up and on the switching side, that is not really in any numbers I'm talking about right now. I mean that's not very little or nothing in the [ 16.5 ] And then the year after, we'll probably get some contribution, but that's never showed up in an Analyst Day we've done or any kind of discussion. So that's all in front of us, and that's all upside. And I think it can be very meaningful over time. So we'll see.

Operator, we'll do 1 more question, and then I'll give some closing remarks and we'll end the call.

Our last question comes from the line of Simon Leopold with Raymond James.

I'm wondering if we look at the fiscal '28 outlook for the custom to double and the 3 drivers that you outlined, can you give us a little bit more color as to which of these is the biggest or how to think about the contributions between existing customer expansion, the XPU attach and then the new Tier 1? Just trying to get a sense of relative size of each

Yes. No, no problem. Thanks for the question. So last quarter, when we talked about it, it was about 1/3, 1/3, 1/3 in those different buckets, the existing programs, XPU attach and then in our new program. And it's about the same. But what's happened is what we said it's going to more than double. So all of those have sized up. Every -- I think across the board, I would say, all of our different custom programs, which, again, I mentioned, was quite a few when you add the XPU attach in there.

Every one of those has sized up and wants to be bigger for next year. So it's a double plus which is great, and I'd say it's roughly the same ratio. We'll know more, obviously, as we get through the year here and kind of lock the production plan for next year in our allocation. But at the moment, everything wants to be more, and we will do that, we'll more than double the business from this year to next year and about the same increments in terms of growth. So thanks for the question.

Operator, do you want to give some closing comments? Do you want to close the call first then I'll do it. Do you want to do it?

No. I'll just close it out once you're finished.

Excellent. Okay. Thanks for your help today. All right. So thanks, everybody, for joining. I appreciate it and appreciate the interest in the company. Marvell is in the middle of really an incredible growth period. We're seeing record demand. We're seeing record bookings the last few quarters. Our data center business is on fire, and we're projecting accelerating revenue growth for this year and next year already from a strong base. I mean, basically, we were 46% data center growth last year, this year's 50% and next year is looking like 55%. So it's only getting better.

The team is doing an excellent job winning new designs, so we can keep the growth engine humming for the foreseeable future. And this is a company that was put together and purpose-built and going back almost 10 years actually in the making to get to this point, we've had a lot of kind of high-profile M&A and integrated that well. But incredibly strong organic investment by Marvell engineers as well to build really a best-in-class leading portfolio across the board. And this came up in some of the questions, but I'll just touch on it. I mean what really is resonating with our customers is we have all the pieces.

We have all the pieces. We have the pieces that can help our customers architect their fully AI -- their fully optimized AI infrastructure, and that can be built on Marvell end-to-end technology. I mean having custom under one roof, high-speed optical interconnects switching, leading at SerDes and IO and then the things we talked about earlier in terms of our capacity, our scale, and our ability to manufacture in yield and high volume and customers' ability to trust us to do that.

So very exciting time for the company. It's been a little roller coaster over the last year. The revenue has been up into the right. But there's been ups and downs along the way. And we have a very, very dedicated and loyal and committed employee base in this company. And I want to thank all the Marvell employees that are listening and all the engineers in this company and everybody in every function who's working your butts off every day to get to where we've gotten to, your focus and commitment is highly appreciated.

So look, we're well on the way to be one of the big winners in this AI cycle. We know what we're good at, and we're going to keep doing that, keep our head down, keep executing and keep driving it forward. So we look forward to seeing all of you. I know there's a whole bunch of different investor events, bus stores, huge amount of people coming through here at Marvell. I'll be in a lot of those meetings, but you'll also see -- you'll get a chance to see Chris and Willem, Sandeep, Ashish and me. So with that, I want to conclude the call. Thanks, operator, and thanks, everybody, for your interest in Marvell. Take care.

Thank you. Ladies and gentlemen, thank you for your participation. This does conclude today's teleconference. Please disconnect your lines, and have a wonderful day.

Good afternoon, and welcome to Marvell Technology Inc. Fourth Quarter and Fiscal Year 2026 Earnings Conference Call. [Operator Instructions] Please note that this event is being recorded.

I will now turn the conference over to Mr. Ashish Saran, Senior Vice President of Investor Relations. Thank you. You may begin.

Good afternoon, everyone. Welcome to Marvell's Fourth Quarter and Fiscal Year 2026 Earnings Call. Joining me today are Matt Murphy, Marvell's Chairman and CEO; Willem Meintje, CFO; Chris Koopmans, President and COO; and Sandeep Bharathi, President, Data Center Group.

Let me remind everyone that certain comments made today include forward-looking statements, which are subject to significant risks and uncertainties that could cause our actual results to differ materially from management's current expectations. Please review the cautionary statements and risk factors contained in our earnings press release, which we filed with the SEC today, and posted on our website, as well as our most recent 8-K, 10-K, 10-Q and other documents filed by us from time to time with the SEC. We do not intend to update our forward-looking statements.

During our call today, we will refer to certain non-GAAP financial measures. A reconciliation between our GAAP and non-GAAP financial measures is available on our earnings press release.

Let me now turn the call over to Matt for his comments on the quarter. Matt?

Thanks, Ashish, and good afternoon, everyone. Let me begin by extending a warm welcome to the Celestial AI and XConn team. We recently closed both acquisitions and the teams are working closely together with joint product road map discussions in full swing with customers. These highly strategic additions further strengthen our technology platform and significantly enhance Marvell's position in the rapidly emerging AI scale-up networking market. I'll provide additional detail on these acquisitions later in today's call.

Now turning to our results and business outlook. For the fourth quarter of fiscal 2026, Marvell delivered record revenue of $2.219 billion, reflecting 7% sequential growth. Revenue exceeded the midpoint of guidance, driven by strong demand in our data center end market. As a result, non-GAAP earnings per share of $0.80 exceeded the midpoint of guidance by $0.01.

Turning to our full year results. Fiscal [indiscernible] was an exceptional year for Marvell. Revenue grew 42% year-over-year to approximately $8.2 billion as reported, and approximately 45% year-over-year, excluding the divested automotive Ethernet business. Our data center revenue surpassed $6 billion, growing 46% year-over-year. This performance was driven by robust demand for our interconnect, switching and storage products, along with a strong ramp in our custom business, which doubled in fiscal 2026.

As we begin fiscal 2027, we are seeing very strong demand across our entire data center portfolio with [indiscernible] accelerating at a record pace. This robust demand is reflected in our guidance for the first quarter of fiscal 2027, the total company revenue forecasted now to grow 8% sequentially at the midpoint to $2.4 billion. Looking ahead, we expect to grow revenue every quarter this fiscal year at a similarly strong sequential rate, which would result in Q4 revenue exceeding $3 billion exiting this year.

[ Store cast ] also implies that our year-over-year revenue growth rate will accelerate each quarter throughout fiscal 2027. As a result, we now expect overall Marvell revenue in fiscal 2027 to grow more than 30% year-over-year, approaching $11 billion. Notably, this outlook is meaningfully higher than what we communicated in our prior updates.

Some of you may recall, in September 2025, during an investor call hosted by JPMorgan, we provided a fiscal 2027 revenue outlook of approximately $9.5 billion, which at that time was received positively as it was significantly higher than the market expectations. In our December 2025 earnings call, as CapEx growth forecasts continue to increase, we updated our fiscal 2027 revenue forecast to approximately $10 billion. Today's outlook approaching $11 billion raises our forecast by almost another $1 billion. Importantly, this outlook is driven by Marvell's organic businesses as the recently closed acquisitions are not expected to contribute meaningfully until fiscal 2028.

The increase in our overall revenue outlook is all being driven by our data center business. Since December 2025, cloud CapEx expectations have continued to increase, and we have seen our bookings continue to accelerate. As a result, we now see our fiscal 2027 data center revenue growing by 40% year-over-year. We expect all our key product lines in data center to be stronger than our prior outlook. Notably, we expect our interconnect business to more than 50% year-over-year, well above our prior expectation of 30% growth. For our communications and other end market, we expect 10% revenue growth in fiscal 2027.

Looking ahead to fiscal 2028, while we assume the rate of CapEx growth moderates from the current fiscal year, we expect continued robust data center revenue growth for Marvell. We expect our interconnect business to significantly outpace cloud CapEx growth, our custom business to at least double year-over-year, and our Ethernet switching business to continue to ramp meaningfully. In addition, we expect Celestial AI and XConn to contribute approximately $250 million in aggregate revenue in fiscal 2028. As a result, we expect data center revenue and fiscal 2028 to grow close to 50% year-over-year.

Achievement of our forecast would result in 3 straight years of data center revenue growth [indiscernible] at well over 40%. For our communications end market, we continue to expect low single-digit percentage revenue growth in fiscal 2028, consistent with our prior view. So in aggregate, we expect Marvell's overall revenue in fiscal 2028 to grow close to 40% year-over-year, reaching approximately $15 billion, roughly $2 billion higher than the outlook we [indiscernible] in our December earnings call, and driving our non-GAAP EPS to well over $5. This outlook is based on demand we are seeing now and designs that are already in execution.

As we progress through the fiscal year, we plan on remaining closely aligned with our customers as we expect them to continue to invest in AI infrastructure. With that, I'll provide more context on our numerous growth drivers across our end markets, beginning with data center.

In our data center end market, we delivered record fourth quarter revenue of $1.65 billion, representing 9% sequential growth and 21% year-over-year growth. Revenue exceeded our guidance, driven by increased demand across our interconnect portfolio. We achieved sequential growth across all key product lines, including optical interconnects, custom silicon, switching and storage.

Looking into the first quarter, we expect our data center revenue to grow approximately 10% sequentially, including a seasonal sequential -- including a seasonal sequential decline in on-premise data center revenue. Let me now highlight the broader trends across both our established data center businesses and our newer growth initiatives, including recent acquisitions. I'll organize the discussion into 3 categories. Interconnect, switching and custom.

I'll begin with Interconnect, where we offer the industry's broadest and comprehensive high-speed connectivity portfolio, addressing scale out, scale across, and scale up networking. In our scale-out [ PAM ] franchise, demand remains robust for our 800-gig products. We are also seeing very strong bookings from multiple Tier 1 customers for our 1.6T solutions which entered production in the second half of fiscal 2026. Reflective demand in our first-to-market technology leadership, we expect our 1.6T revenue ramp -- to ramp very rapidly in fiscal 2027 and with substantial additional growth projected in fiscal 2028. As a result, we expect to continue to maintain leadership in the PAM market at 1.6T just like we have at every PAM generation.

Marvell is the first company to productize 200-gigabit per [ lane ] technology, enabling the 1.6T transition now underway. While this generation is expected to continue to grow through the end of the decade, Marvell has already demonstrated 400-gig per [ lane ] technology. We expect that this will position us to enable the industry's subsequent transition to 3.2T, once 1.6T reaches full maturity.

To support campus-wide data centers requiring longer reach than traditional PAM solutions, Marvell has introduced [ coherent ] light, optimized for 2 to 20-kilometer applications within a highly power-efficient outlook. We have already begun shipping first-generation 1.6T coherent light products and are now introducing a second generation with integrated [indiscernible] security.

Turning to scale across interconnects, a technology we pioneered with our 100-gig DCI modules, we continue to lead the market with coherent 400-gig and newer 800-gig solutions. We are winning new customers and expect to supply DCI modules to all 5 major U.S. hyperscalers this year. We see significant long-term growth in this market, as the global data center footprint expands and bandwidth requirements between data centers continues to increase.

Industry forecasts project that DCI pluggable TAM to grow by more than 5x by calendar 2030, with speeds doubling each generation and feature complexity increasing, including the integration of [indiscernible]. To that end, earlier today, we announced our latest innovations and scale across interconnects, including the industry's first Secure 1.6T ZR and ZR+ DCI modules powered by our new 2-nanometer coherent DSP. We also introduced a new 2-nanometer 800-gig DSP, which enables second-generation lower-power 800-gig DCI modules. DCI modules powered by these 2-nanometer [ MACsec-enabled ] DSPs are expected to begin sampling later this year. This positions Marvell to maintain technology leadership, supported by our proven expertise in large-scale manufacturing of these highly specialized and complex modules.

Now let's move to scale-up interconnects, which is an entirely new and rapidly emerging market. We are very excited about what we believe to be a [indiscernible] opportunity unlocked by Celestial AI's photonic fabric, or PF technology, as well as growing customer traction for our AEC and retimer solutions. As discussed last quarter, Celestial AI's PF technology is expected to enable large-scale commercial deployment of CPO for scale-up connectivity starting next year. Our chiplets will be [indiscernible] into both custom [indiscernible] and the scale-up which is connecting them together on both sides of the length.

With the acquisition now complete, Marvell's engineering and operations teams fully engaged in bringing Celestial's first generation chiplet into high-volume manufacturing. We remain on track for our forecast for our CPO revenue from Celestial to reach a $500 million annualized run rate in the fourth quarter of fiscal 2028, doubling to a $1 billion annualized run rate by the fourth quarter of fiscal 2029. We have seen strong interest from a broad range of customers in Celestial's photonic fabric technology following the deal announcement. We look forward to updating on our progress in the scale-up interconnect market, which we believe could exceed $10 billion by 2030.

In the [ AEC ] market, we have secured design wins with 3 Tier 1 U.S. hyperscalers and several additional customers, including model builders and hardware OEMs. We are also seeing strong traction for our retimers. As a result, we expect combined AEC and retimer revenue to more than double year-over-year in fiscal 2027. We continue to abate through our [ Golden Cable ] initiative, a strategic program that delivers a complete solution, including industry-leading software and validated reference designs, enabling ecosystem partners to rapidly design and deploy AEC products at scale. Hyperscale customers benefit from access to multiple high-volume cable OEMs offering fully compatible ADCs, both on the same high-performance Marvell DSP and reference design.

[indiscernible] data center switching, we delivered strong growth in fiscal 2026 with revenue exceeding $300 million, driven entirely by scale-out applications. Given sustained demand for our current 12.8T products and a strong ramp of next-generation 51.2T products, we now expect data center switch revenue in fiscal 2027 to surpass $600 million, up from the $500 million we had indicated last quarter. We are seeing strong engagement from both existing and new customers for our 51.2T platform, and our upcoming 100T platform, which we begin to -- should we expect to begin sampling in the first half of this fiscal year. Our 100T switch delivers industry-leading power efficiency and lower latency, attributes that are especially critical for AI applications.

In scale-up switching, the combination of Marvell and XConn creates a significantly larger team to address rapidly emerging UAL and Ethernet-based opportunities. UA Link builds on decades of PCI ecosystem development and incorporates high-speed interface innovations from Ethernet to meet the bandwidth, latency and reach requirements of next-generation accelerated infrastructure. XConn expands Marvell's switch team's deep PCIe switching expertise, enabling a comprehensive -- enabling comprehensive support to customers building next-generation AI platforms. We are fast tracking our scale-up switch road map by leveraging our extensive experience in developing large reticle size scale-out switch chips, and best-in-class in-house high-performance series.

We remain on track to sample our [ UA-LINK 115T ] solutions in the second half of this fiscal year with volume production expected in fiscal 2028. In parallel, we continue to advance the Internet-based road map with key customers. We're able to further enhance our scale-up road map by enabling integration of our CPO technology from Celestial directly with our switches, delivering a purpose-built, fully optimized end-to-end optical scale-up platform.

XConn also adds advanced PCIe and CXL switch solutions, another completely incremental TAM for Marvell. The PCIe Gen 6 and CXL 3.1 solution is based on a monolithic switch architecture supporting up to 256 lanes, delivering the industry's highest ratings and lowest latency. PCIe switching remains foundational in standard compute architectures connecting CPUs to peripherals and increasingly [ the AI ] infrastructure to connect CPUs to [ XPUs ]. In parallel with next-generation protocols like [ UA Link ], PCIe is also adopted for XPU to XPU connectivity, particularly in AI inference systems and small- to medium-sized clusters.

CXL is rapidly becoming essential for memory disaggregation in modern data centers. We have been investing in CXL for several years and XConn switching portfolio, combined with Marvell CXL [indiscernible] create the industry's most comprehensive CXL platform. XConn was already engaged with more than 20 customers prior to the acquisition. As part of Marvell, XConn now benefits from our global sales and marketing reach and strong presence in the data center. As a result, we expect to drive strong growth in both the PCIe and [indiscernible] switch markets over the next several years.

Turning now to our custom business. This remains one of the most compelling growth drivers for Marvell. In just a few years, we have scaled from zero revenue to $1.5 billion in fiscal 2026. As you may recall, the first meaningful ramp again in the second half of fiscal 2025. Fiscal 2026 marked the first full year of production for those programs. And as a result, we doubled our customer revenue year-over-year. We expect custom revenue to grow more than 20% year-over-year in fiscal 2027, higher than our prior view.

We continue to see growth from our [ Lead XPU ] program this year, including a transition to its next generation. As I noted last quarter, we have purchased orders covering the entirety of this fiscal year's forecast for this next-generation program and are now ramping production. In addition, we are expecting the growth to continue in fiscal 2028 from this program. We are also deeply engaged on the follow-on generation of this XPU.

In addition, several XPU attach programs are ramping in fiscal 2027, including our initial CXL and NIC products. CXL demand is accelerating, partly driven by tight memory supply. Our custom CXL expanders enable customers to reuse prior generation DRAM with new XPUs, GPUs and CPUs, while also supporting near-memory compute operations. A recent white paper from a leading hyperscaler on next-generation LLM inference architectures highlighted, near-memory processing is a key opportunity to improve model performance. They cited Marvell [indiscernible] a processor as an example of a CXL-enabled solution that improves programmability and simplify system integration.

This all provides a great setup for fiscal 2028. We continue to expect custom revenue to at least double year-over-year from 3 primary drivers. First, continued growth from our existing custom programs. Second, Multiple XPU attach programs reaching high volume, particularly in custom neck and CXL applications. As I mentioned last quarter, we have line of sight to revenue exceeding $2 billion by fiscal 2029 from just these two use cases, and we expect to make significant progress towards that outlook through fiscal 2028. Third, our new Tier 1 XPU program ramping into high-volume production. This program has continued to progress well -- very well through development, and we have firm volume requirements for all of next year and are planning for high-volume manufacturing.

Beyond programs already won, we are encouraged by strong new design engagements with both existing and new customers. Custom compute is proliferating across the hyperscale ecosystem with inference optimized hardware becoming increasingly important. We are seeing an unprecedented level of activity across multiple new engagements as hyperscalers increased their cadence of custom chip development. We are engaged in deep technical discussions on innovative new architectures, and are seeing a massive opportunity on 2-nanometer and below process technologies.

Okay. Turning to our communications and other end markets. We delivered fourth quarter revenue of $567 million, up 2% sequentially and 26% year-over-year. For the first quarter, we expect low single-digit sequential growth on a percentage basis and approximately 30% year-over-year.

In summary, we concluded fiscal 2026 on a strong note with revenue growing 42% year-over-year and non-GAAP EPS increasing 81%, roughly twice the rate of revenue growth, demonstrating the strong operating leverage in our business model. Fiscal 2026, we were very active on the M&A front, divesting our automotive Ethernet business for a double-digit revenue multiple, and rapidly redeploying the proceeds into two highly strategic acquisitions. These positions [indiscernible] Marvell at the forefront of the large and incremental AI scale-up networking market.

At the same time, we continue to execute our capital return program returning $2.245 billion to stockholders through share repurchases and dividends. So far in fiscal 2027, we are seeing strong bookings across our entire data center portfolio with customers signaling robust demand not only for this year but for the next several years. We believe we are still in the early stages of a strong multiyear growth cycle for Marvell.

Our first quarter fiscal 2027 guidance represents 27% year-over-year growth at the midpoint, reaccelerating from 22% in the prior quarter. We expect year-over-year growth to accelerate each quarter throughout fiscal 2027, with revenue exiting the fiscal year at over $3 billion in the fourth quarter. We have reached our fiscal 2027 forecast meaningfully. And in fact, the revenue growth rate we are projecting today for fiscal 2027 is roughly double the outlook we provided just a few months ago in September.

This is an exciting moment for Marvell. I want to take a moment to thank our global team for staying focused despite the external noise, and delivering consistent execution, which has enabled record results and positioned us to capitalize on what we expect will be a massive AI opportunity ahead. I look forward to updating you on our progress in the coming quarters.

With that, I'll turn the call over to Willem for more detail on our recent results and outlook.

Thank you, Matt, and good afternoon, everyone. Let me start by summarizing our full fiscal year 2026 results, which were very robust across the board.

In fiscal 2026, Marvell delivered $8.195 billion in revenue, growing 42% year-over-year. This growth was primarily driven by AI demand in our data center end market, as well as the continuing recovery in our communications and other end markets. For the full year, on a GAAP basis, our gross margin was 51%. Operating margin was 16.1%, and earnings per diluted share was $3.07. On a non-GAAP basis, our gross margin was 59.5%. Operating margin was 35.3%, expanding by 640 basis points year-over-year, and earnings per diluted share was $2.84, growing 81% year-over-year. We also significantly increased capital returns to our stockholders, returning $2.245 billion through share purchases and dividends in fiscal 2026, an increase of approximately $1.3 billion from the prior year.

Moving on to our financial results for the fourth quarter of fiscal 2026. Revenue in the fourth quarter was $2.219 billion, growing 22% year-over-year and 7% sequentially. Our data center end market was 74% of total revenue, with our communications and other end markets contributing the remaining 26%. GAAP gross margin was 51.7%. Non-GAAP gross margin was 59%.

Moving on to operating expenses. GAAP operating expenses were $744 million, including stock-based compensation, amortization of acquired intangible assets, restructuring costs, and acquisition-related costs. Non-GAAP operating expenses came in at $517 million, in line with guidance. Our GAAP operating margin was 18.2%, while our non-GAAP operating margin was 35.7%. For the fourth quarter, GAAP earnings per diluted share was $0.46. Non-GAAP earnings per diluted share was $0.80, above the midpoint of guidance, reflecting year-over-year growth of 33%.

Now turning to our cash flow and balance sheet. The fourth quarter cash flow from operations was $374 million. Our inventory at the end of the fourth quarter was $1.39 billion, growing $374 million from the prior quarter. Our working capital has increased to support the significant revenue growth we are driving. During the quarter, we repurchased $200 million of our [ stocks ] through our ongoing capital return program, and returned $51 million to shareholders through cash dividends in the quarter. We expect to continue to return capital through repurchases and dividends.

As of the end of the fourth quarter, our total debt was $4.47 billion, with a gross debt-to-EBITDA ratio of 1.38x, and a net debt-to-EBITDA ratio of 0.57x. Our debt ratios have continued to improve as we have driven an increase in our EBITDA.

Turning to our guidance for the first quarter of fiscal 2027. We're forecasting revenue to be in the range of $2.4 billion, plus or minus 5%. We expect our GAAP gross margin to be between 51.4% and 52.4%. We expect our non-GAAP gross margin to be between 38.25% and 59.25%. Looking forward, we anticipate that the overall level of revenue and product mix will remain key determinants of our gross margin in every -- in any given quarter. We project our GAAP operating expenses to be approximately $872 million. We anticipate our non-GAAP operating expenses to be approximately $575 million in the first quarter. This is stepping up from the prior quarter due to the typical seasonality in payroll taxes, and employee salary merit increases, as well as the addition of Celestial AI and XConn. The two acquisitions in aggregate are expected to add approximately $75 million to our fiscal 2027 annual non-GAAP operating expenses.

We expect our GAAP other income and expense, including interest on our debt, to be an expense of approximately $51 million. We expect our non-GAAP other income and expense, including interest on our debt to be an expense of approximately $48 million. We expect a non-GAAP tax rate of 11%. We expect our basic weighted average shares outstanding to be [ $876 million ], and our diluted weighted average shares outstanding to be [ $883 million ]. We anticipate GAAP earnings per diluted share in the range of $0.26 to $0.36. We expect non-GAAP earnings per diluted share in the range of $0.74 to $0.84.

As we look ahead to the rest of fiscal 2027, we will continue to invest in growing our business while driving operating leverage. On a sequential basis, we expect non-GAAP OpEx to remain flat in the second quarter and then grow in the low to mid-single digits on a percentage basis in each of the third and fourth quarters, well below the rate of revenue growth Matt provided in his remarks. We are seeing strong growth from our existing franchises and scale out and scale across AI as well as custom, and we are investing to drive new revenue streams from the rapidly emerging AI scale up market. We have entered a robust multiyear growth period and are looking forward to delivering strong earnings growth to our stockholders.

With that, we are ready to start our Q&A session. Operator, please open the line and announce Q&A instructions. Thank you.

[Operator Instructions] Your first question comes from Ross Seymore with Deutsche Bank.

Matt, thanks for all the updates on the out year -- well, fiscal year, both this and next. Beyond the magnitude of the revenue growth, can you just talk about the profile of it? Is the customer base broadening? People are always worried especially in your custom business about the concentration of it. So I just wanted to get a little bit more color on the shape of the demand from a customer perspective?

Yes. Thanks, Ross. Well, first of all, we're deeply engaged across the entire ecosystem, extremely strong position with the top 4 U.S. hyperscalers and then the next level. And each of them, we have a different concentration and revenue mix. But just to be super clear, if you look at this year and you look at us driving the company to $11 billion, and then you unpack things like custom, it's not that big a percentage of the total. So that's not what's driving our concentration.

I mean by design because of the top 4 U.S. hyperscalers is spending the bulk of the CapEx, that's where the dollars are going to go. But we're quite diversified across each of them. And some of them we sell a different mix, obviously, of product to. But in the case of all 4, within our portfolio, which I just went through the laundry list of all the different types of products that we provide, we're highly diversified within each of these customers.

So -- so yes, custom is something that gets a lot of attention. But if you just look at the numbers I gave you and the context as I said, it's a piece of the equation, but not all of it. And then over time, even on the custom business, as you look out through fiscal '28 and fiscal '29, Remember, we've got 20-plus design wins, or products now, sockets that are either in production or going into production, it's going to layer in across all those companies as well. So the diversification is only going to get better over time.

But we're very unique in sort of the breadth think of the products that we offer and the product lines we have to really serve end-to-end the needs of all of our key hyperscalers. And the last two M&As we just did really round that out nicely in terms of adding PCIe, getting -- beefing up the UAL, and then also adding key silicon photonics capabilities.

Your next question comes from Harlan Sur with JPMorgan.

Congratulations on the strong results [indiscernible]. Matt, on your custom [indiscernible] attached subsegment, open AI recently inked a partnership with your lead XPU, customer to consume, I think, something like 2 gigawatts worth of your lead customers, next gen and next-gen XPU. So it feels like the overall demand for AI compute continues to accelerate. Right on top of that, like you said, you're ramping 15 to 20 XPU attached custom programs this year and next year.

Within our better outlook for custom this year, and with you already starting to ramp your lead customers next-gen XPU program, do you still anticipate a stronger second half step-up of this XPU program? Or is it more of a linear ramp through the year now? And I think you previously thought that you would exit this year with custom driving about a $2 billion sort of annualized growth rate. What does that exit run rate look like today?

Yes. Thanks, Harlan. I think the first part of your question is absolutely seeing strong validation in the market for the AI compute spend, and the fact that a significant portion of that continues to go to companies that are building their own XPUs. So that's a positive trend. We certainly see it. And you're right. Even where we don't necessarily have the XPU, we have XPU attached. So all the [indiscernible] is going [indiscernible] in customers where we're not participating. So we're -- we participate across every one of those large companies and more on [ XPU attached ]. So that's a very positive trend for us that's driving our positive outlook for sure through this year, which we said custom was going to grow faster than we thought, but more meaningfully into fiscal '28 and '29.

And then from a linearity perspective, under the hood, we kind of give you a view of what the sequentials would look like throughout the year. But yes, custom, we have said was going to be a stronger second half due to a program transition. That's still the case. And that -- the type of exit rate you're talking about is certainly still intact and probably has an upward bias to it. If you look at the exit rate we're talking about for the whole company now, we're looking at north of $3 billion. So within that custom continues to have some real upside to it.

But that's going to improve meaningfully and the revenue growth is going to continue into fiscal '28 which is basically those programs from the second half now having a full year. So that's going to provide some nice growth. Content increase, then layering in the XPU attach, and then layering in our new program with a new Tier 1 hyperscaler, which is in its early stages, but just even the rough plug we have for them, is significantly lower than actually the wafers that we're planning on starting the material and the production plan we have with our manufacturing supply chain. So I think it's a very reasonable setup for next year with a lot of upward bias depending on if these trends continue.

Your next question comes from Aaron Rakers with Wells Fargo.

I guess my first question is on the optics, the electro-optics business. I know Matt, you've talked about in the past that your ability to kind of outgrow the pace of what we're seeing in CapEx spend. So I guess my question is, we've seen some massive upward provisions in CapEx. I think most people look at that and say, hey, we're looking at like 60% plus growth this year. Do you think you can grow at that level? And how do you think about the durability of that growth as we move into fiscal '27 -- or fiscal '28?

Yes. Aaron, your observation is absolutely correct. And that's why even as we look at the upward momentum we see in the business for this year, a big part of that change is in that electro-optics portfolio. We had been calling it kind of closer to CapEx as we were modeling what we thought we could do this year back in the September call and then even in the -- in my December call. But now it's clearly growing more like -- more like accelerator growth and more like this sort of accelerated CapEx growth. So yes, it's growing like 50% plus this year now. And that momentum is going to continue, okay, into fiscal '28. A couple of things are happening there.

The first is that as new XPU, GPU, et cetera, generations are released. There is -- we are seeing some increased concentration on the attach rate of optics. So that's a positive. You get more 1.6T, which has just -- because of its performance, commands higher ASP. So that's going to roll in. And then we have -- yes, we just have some pretty new exciting programs happening in that area. So that business has been growing at like [indiscernible] a year-ish. You can give it plus or minus, I get the exact data. But it's been at that rate for some time since we acquired Inphi and the data center stuff really took off. We see that continuing not only through fiscal '28, but that momentum should continue beyond that. Maybe it's not the exact same magnitude, but it's significant. We have a real head of steam on the electro optics business at Marvell.

Your next question comes from Blayne Curtis with Jefferies.

Matt, I don't want to ask on the custom business. So I think you feel very confident about the trajectory. I'm just kind of curious, one, can you just help us with [ '26 ]? Because I mean, you have the big broad swath, but I mean, is that custom business growing 30% this year? I just want to figure out the base that you're going to double. And can you talk about that second major XPU customer? I mean, kind of give this type of guidance, like what kind of confidence do you have in the timing of that program?

Sure. Yes. And I think you're talking about -- just to be clear, calendar '26, fiscal '27 set on custom, kind of what numbers are we talking? Is that the first question? The second one is the...

Yes, sorry, your fiscal year. But yes, fiscal '27 is at around 30%. And then your confidence level on that second major [indiscernible] customer and timing as we try to layer that in to get to that double?

Yes, great. And by the way, I don't feel bad. I've been in this job for 10 years, and I still have to translate every day between my fiscal year on my calendar year. So don't feel bad.

For fiscal '27 we had been indicating after the double from last year, it would grow 20% this year. So we're just saying that's north of that. So I'm -- I can't give you the exact number now, but it's biased upwards, but it's just -- so just take what I read before that 20%, you can make an estimate but higher, but not significant enough where I would like give you a new number, but just say it's [indiscernible] higher. So in the ballpark, but higher. So then next year, obviously, gets a little bigger than we thought.

And then the reason we're confident is we have line of sight in terms of -- well, first of all, we have history, right? We've built these large scale custom programs before. We've done these ramps before. We have a good sense of when the product is going to go through its key milestones through [ NPI ]. We have had very detailed discussions and alignment around manufacturing plan, and we've aligned up a corridor for fiscal '28 for production on this that would be a lot higher than what I'm indicating to you.

I think we're budgeting at the moment for -- is there a delay? Does it take longer, et cetera. And plus, I think at the moment, it seems like a lot of folks aren't really believing it's maybe going to do anything. But I think our plug is very, very reasonable for next year in terms of what's there. And I think it would bias quite a bit higher if we could just achieve what we're planning on reserving in terms of capacity. So more to come there. But I think we try to call the ball as best we can. And in general, we've done a pretty good job over the years [indiscernible] to size and judge things in advance. And then usually, we're pretty good and then they [ buy us ] upwards. So we'll see where it lands. But I think it's not a big stretch for this custom business to double next year.

Your next question comes from Ben Reitzes with Melius Research.

Matt, nice to see the beat and raise. I wanted to ask the question about what got better in a different way? I mean, if you could just unpack since December, the $2 billion, especially the -- how fiscal [ '28 ] got $2 billion better since December? What -- if we can unpack that and what exactly got better? And then potentially, I'm going to be a little greedy, what can carry into the next year as well, calendar '28 of those signs that you saw since then?

Yes. Thanks, Ben, and great to hear from you a long time. So I think -- one is you just kind of look at it as progression. I mean, it's the first point I'd make is we tried to give a view for investors to be helpful because there's a lot of concern and angst back at the end of last year. So in September, we talked about 9.4-ish for this year. And then that's now -- in December, we said that looks more like [ 10% ], and now I'm saying it's more like [ 11% ]. So some of that is just the progression in terms of time and getting better visibility and more concrete. And then that just ripples into the -- I'll use calendar for a second, calendar '27.

But on top of that, I mean, one, we've now got very firm requirements and understand the profile, in particular the interconnect business. And that is, I think we had called it very conservatively, to be frank. And I think even a few analysts last quarter kind of [ dinged us ] saying, well, you're plugging your interconnect business at CapEx, but it really looks more like it should be tied to GPU, XPU. And that's really the case. So I think we're seeing that now in terms of the forecast. So that's come up quite a bit, which then again, the upward revisions we're seeing for this year then ripple into next year.

And then I'd say this is all underwritten Ben, by extremely strong bookings and backlog layering in and then the detailed conversations with our customers around supply planning. It's just given us a much more concrete view. And by the way, the other reason I think it's important and why we felt it was important to continue to update on this metric is that we set targets back in April of '24 for calendar '28. We did that around some assumptions around data center market share of 20%. And those numbers looked enormous at the time we talked about it. I think you guys were there.

We were doing low billions a quarter in revenue at that time. $1 billion -- I think we had guided [indiscernible] when we put out this number that was like $15 billion in data center revenue in 4 years. And I think everyone thought we were nuts. At our June AI investor event, we said the TAM went up, so that data center revenue bogey kind of moved up to like if you just did the math, moved up to more like $18 billion and change. But you kind of look at it and you see where we're landing in calendar '26. And now we're sitting here in '27. I mean, it's -- we're very much on track actually to those targets that we had set in.

And so in a way, yes, it's some upward revisions and that's part of it is just because we have more data, but it actually is also validating, I think, the plan we set actually 4 years ago about what we thought we could go off and do, which were very lofty ambitions, and we're not there yet, and we have to go execute [indiscernible] me and the whole team. But we're very encouraged by what we're seeing, and the proof is in the pudding that we're getting in terms of the backlog forecast and alignment with our entire [indiscernible] chain to be ready to go make that happen both this year, next year and in calendar '28.

Your next question comes from Tom O'Malley with Barclays.

I think in the preamble, Matt, you talked about AEC and retimers more than doubling in the fiscal year. Could you maybe give us some perspective on the base there? And then been really helpful in the next few years kind of giving the contributing factors of what is a pretty impressive growth profile. Maybe talk a little bit about how much that can contribute in this broader overview?

Yes. Tom. Yes, this is still an emerging area for us. So we're -- it's doubling -- over doubling this year, but it's probably in the $200 million range is what I would say. I mean, we actually -- I think based on some of the things we're looking at, maybe that goes higher, but that just gives you a sense of the magnitude. But it's going to keep going from there. I mean this is -- we've seen this in a lot of our emerging product areas when we get into them. Once they start doubling, they kind of keep doubling, and you know this market quite well. There's quite a bit of room, I think, for a bunch of people to participate. So yes, we're very encouraged by what we see based on the traction we have on our products, especially on product leadership.

We leverage a lot of our DSP and PAM technology in this area. We inflected when both on the retimer side and [indiscernible] move from NRZ to PAM, and that was -- that was our kind of conscious decision to do that. So we're earlier in the cycle because we're coming in, in later generations than some of the existing sockets, but we intend to really invest here in a significant way and participate.

Over the long term, we see that as complemented. There's a place in the market for this, and we're going to participate. But obviously, we made the bet when you go back to even the Inphi acquisition 5 years ago on optics and pluggable optics, in particular, and then now with Celestial also, on CPO on the scale upside. So there's a period of time we're going to participate. I think it's going to be great, and the business is going to do well and it leverages what we have. And I think it's going to be just part of our goal to be the end-to-end provider for our customers of all of these types of solutions. From electrical to optical to silicon photonics various reaches various distances, various form factors.

And that's what our customers are looking for, okay? They want to have an interconnect partner that could be the one-stop shop and do it all and have high amounts of leverage on the IP, so they can trust it, because we do it ourselves and also on the firmware and the software, and the system implementations, they also want to make sure that they have reusability. So it's been a virtuous cycle here, just the scale-up part relative to [indiscernible] smaller but growing rapidly.

Your next question comes from Vivek Arya with Bank of America Securities.

Matt, I just wanted to first clarify what your XPU attach was last year and what contribution you expect in '27 and '28? And then, kind of, my more strategic question is, when we look at the pattern of your first large XPU program, right, you had a very strong start, followed by competition from another supplier. How would you handicap kind of your exclusivity at the large new XPU customer you plan to start at next year?

Yes. Thanks, Vivek. So maybe I'll answer the second one first. So yes, we're -- I think you're asking specifically about our newer program that would ramp next year, and we feel very good about our position. These are very deep engagements we have with our customers. We're two hands on the steering wheel on this. This is multi-generational in nature. Given the rate of innovation and the pace that the technology is moving at, it's really in everybody's best interest to plan, not just one generation out but even farther. And so we've really been able to do that, I think, across the with our customers. And so we feel really good about our position there. And the sustainability of that.

It still needs to ramp obviously. But certainly, the CapEx envelope is out there to really consume a lot of product, and we're very encouraged by what we see from a road map perspective. And we're investing heavily as a company to be there across the board on all of the key attributes that these big XPU customers care about. So I think more to come on that, as well as future opportunities on XPU for the company. But we feel very good about our position in the next few years in terms of line of sight to hitting the revenue targets that we talked about over the last couple of years and then growing beyond that.

And then, yes, I'm sorry, then the -- on the XPU attach, we [indiscernible] give the exact numbers, but just maybe big round numbers. And maybe we'll first start with the line of sight just on the NIC and CXL I gave you, which was kind of $2 billion out in '28, and then you layer more on that. So -- and by the way, just -- we had sized for everybody on the call, the XPU attached TAM in the future at about $15 billion in calendar '28. We didn't break it out exactly, but we had a total market share goal of about 20% in that time frame. So I'll just call that $3 billion, we're driving in that area. So let's take a step back now.

XPU attached probably in the couple of hundred million ballpark [indiscernible], like this last year, doubling this year, maybe over doubling again the year after. So I think by next year, this thing is probably a $1 billion type business. We'll see how it all shakes out. It's all going to happen under the hood of our custom business with that. But just to give you a sense, it's on a massive trajectory upward, and it's in that category of kind of double plus each year.

Your next question comes from Tore Svanberg with Stifel.

Congrats on the record quarter. Matt, I was hoping you could give us a bit of an update on the mix of the [ opto ] electronic business. So you talked about [ 1.6 ] already shipping. But my understanding is that 800-gig is definitely going to be the bigger volumes this year. So any sense for what the mix is going to look like for fiscal '27 between [ 1.6 ], and I guess [indiscernible] some [ 400 ]?

Yes. Well, I think you got it right. First of all -- and we've been saying this for a while that [ 800 ] was going to be sort of stronger for longer, and I think that was our mantra even last year. And that's still the case for sure. But as I mentioned in the prepared remarks, we had significant shipments actually of 1.6T at the end of last year, and it's going to ramp again pretty hard this year. But [ 800 ] will still be the majority. I think it's going to take probably through -- I mean, even next year, 800 is still going to be strong.

So I can't give you the exact breakout at the moment, Tore, but part of the -- I think, the uplift as well in terms of just our outlook for interconnect for the year was also based on, kind of, all of our customers revising up in terms of what they were going to need, but maybe a little bit more pronounced in 1.6T and it's really ramping strong with those initial customers we had and more will layer on throughout the year and next year. So yes, maybe more on that later, Tore, but probably not in a position to give you the exact number.

And also, I'd say the reason why too, is it's been moving around a lot. I mean, this has been very dynamic in terms of the bookings environment and the demand environment. So I think the mix will have a better view of what that looks like as we progress throughout the year.

Your next question comes from Joe Moore with Morgan Stanley.

With all the growth that you're looking at here, I wonder if you see anything on the supply chain, that could be challenging for you. My sense is you've come a long way in terms of supply chain management since a couple of years ago, but just any updates there would be great?

Yes. Joe, great to hear from you. I'm going to have Chris answer that, our COO. He's been knee deep and had that job for about 5 years, and he's knee-deep in the supply. So Chris, go ahead.

Yes. Thanks, Joe. Look, we've been in a tight supply environment for anything that touches AI, advanced [indiscernible] fabrication, advanced packaging, large body [indiscernible] since the launch of ChatGPT. And against that backdrop, to your point, we were still able to grow the company north of 40% in total revenue last year. So we clearly have very, very good relationships with our suppliers.

But I would argue that really what helps us we've been forecasting this growth for quite some time. And by giving them multiple years of visibility of what we're going to need and ramping into these numbers is really helping us. And so I'm very confident we've secured the supply that we need for all the growth that Matt outlined this year, next year and beyond.

Your next question comes from Jim Schneider with Goldman Sachs.

It's great to hear the increased visibility you have business in the next year. If I think about the guidance for $15 billion of revenue next year, and $5 of earnings, roughly speaking, that's about 15% to where I see the peak consensus being for next year's revenue, but only about half of that on the earnings side. So can you maybe unpack a bit of what are the moving pieces below the top line? Whether that's gross margin mix, or increased investments to sort of get to that? Or is the [indiscernible] number just relative to [indiscernible]?

Jim, yes, yes, that's like just a floor like it's 5-plus. I mean you can run your own pro forma income statement. But just to give you a sense of how to think about it.

So on the top line, we gave you a framework. And then you can also take, basically where we're going to exit this year and you could use whatever number you want to model finally in your model, but we're saying put in 3, or a bit more. And then if you actually just kind of roll through some of the guidance we've given you already for this year on OpEx and the moving pieces on gross margin, we actually start to get to our target operating model, margin model exiting the year. And that probably continues through the next year is a safe assumption. So the number if you put in 15, and you put in that, it probably -- it floats above $5. So that was not a prescriptive number, or a firm number. It was just a [indiscernible].

People are going to have their own estimates, and you guys will sort of come up with your own view. But yes, no, I'm not making any comment about any kind of margin changes, or dilution, or losing leverage at all. We're going to get leverage -- we're in the mid-30s [ op ] margins right now, if you kind of look at where we were last quarter and what we're guiding and that should float up throughout the year. And then not calling it exactly for next year, but it probably would be consistent with our -- certainly our exit rate of this year. And so that's a simple way to think about it. So it's -- that would pop out a number above $5.

And your next question is from Christopher Rolland with Susquehanna International Group.

Matt, thanks for answering the question. So mine is around kind of big picture, like the CPO scale-up world. Perhaps if you could describe what it looks like, what it looks like from [indiscernible]? But also in your prepared remarks, you talked about integrating Celestial, it sounds like into the [indiscernible] platform I was wondering, are there potentially like [ UA Link ] switch trays that you might be able to integrate this into as well? And just the timing around such products would be cool.

Yes. Great. Thanks. So yes, on the initial plan on Celestial -- and where we -- and just by the way, on the big picture side, our view pretty consistently for some time now has been that the deployment of CPO and scale out would be relatively limited relative to the -- especially relative to the amount of pluggable transceivers that we're going to get deployed. And you can go back many, many [ OFCs ] ago, and that's been our view. And that's been the case to today, for sure. And then I think on the go forward, relatively wise, it's still the case, although you may see some of the industry. That's not our current plan today, although we could absolutely do that and do that integration with the Celestial technology, and our [ Innovium ] CareLink products. And we've done POCs and we've done some work there, but we'll be ready to react to the market there, Chris, when it's needed.

On the scale up, and you mentioned UAL, that's a perfect use case where that is where we see that CPO technology inflecting in a pretty big way and Celestial brought us a pretty significant design win and engagement in that area. And that's what we're trying to drive for next year. So when we ramp it next year, at the end of next year, the -- that would be serving the scale-up application and it would be both an integration of the photonic fabric chiplet into the XPU, as well as on the switch side. That's the first one.

There will be a whole bunch of shipments on scale up switching that will be copper-based, and that's going to exist for some time, too. But we're seeing very, very strong interest across the board for kind of beyond the next few years of where the CPO for scale-up really starts to inflect. And this has been -- and that's sort of been our recognition over the last year or 2, is that's where that's going to happen and that's why we did the M&A and we brought the team on.

So to sum it all up, we'll be shipping next year CPO for scale up at one large customer, and then we're working on more for beyond that. And then the rest of those deployments would be still copper-based.

I think we'll do one more question and then I'll -- I think we'll wrap it.

Our last question then will come from Mark Lipacis with Evercore ISI.

Congrats on the great quarter. Matt, I'm wondering, when you look at these AI systems that your customers are building, it sounds like the way you're talking about it that there's a bigger bottleneck on the connectivity side and the processor side. And so -- but that would be like the part one of the question. And if you agree with that, what's the argument for, why not shift your process or resources to focus more on connectivity?

It seems like your lead is a lot more obvious on the connectivity side, it's a higher margin business. That's where the bottleneck is. And seems like there's a higher chance to add more value, to get paid for that value. And by contrast, the processor side sounds like it's quite noisy on the competitive front. And I think you guys probably get [indiscernible] multiple because of that noise. And so what's the kind of -- what's the rationale for not doing something like that?

Yes. Thanks, Mark. It's a valid question. So first of all, on the interconnect side -- well, on your first part of your question, I'm not sure what's more of the bottleneck or not? I know for sure, the interconnect is a bottleneck, but you could also argue industry-wide, there's a lot more to do on the processing side. But just to be clear, we are absolutely investing to win on interconnect. Like we're not sort of trading anything off there. I mean we're going all in to make sure that we're the leader here. And that's why you can even see when we did our M&A moves last year, we put all that towards that market.

I would say, though, at the same time, we're in the custom business. The -- and you got to break it into two pieces. On the XPU attach side, obviously, that is more margin-rich. And leverages a lot of the Marvell IP, and technology we have, and those typically are our chips that we do. And we've made quite a nice business out of that.

And then on the XPU side, we want to be a big time supplier to our customers. We do get strategic advantage, okay, in being in that market though, Mark, which was actually even a reason thinking all the way back to Avera when we acquired it out of IBM -- or sorry, out of GLOBALFOUNDRIES back in 2019. And one of the reasons that I wanted to do that acquisition and get Marvell into custom -- I mean I never envisioned it would be this significant business for us. Okay. Let's be clear, back in 2019, we weren't thinking that we were going to buy an asset for $650 million, and it was going to open up a $50 billion TAM, but it did.

And one of the strategic rationales that I had for that deal was that it would put Marvell in a product area where we had to be at the bleeding edge. We had to be at the bleeding edge on nodes, packaging, IP development, and it was a tip of the spear type of product line that I felt would be really good for us to really have a driving force to keep Marvell best-in-class on technology. Because at that time, we were making the move from fast follower to trying to be a technology leader. So now we're in that business. I agree with you. It's got a lot of noise around it, and it's got a lot of controversies over the last year and all the different things that have gone on, and maybe it's affected a multiple. But the fact of the matter is, we're in that business.

Our customers are counting on us. We've grown that business from zero to $1.5 billion. It's going to grow again this year. It's going to double the year after. And it's going to be a significant revenue growth driver for Marvell. So I'm not compromising anything on the rest of the portfolio to be in that business. And remember as well, that business also gets significant funding and contribution from our customers who pay us NRE and put their commitment in to make sure that those programs are successful. So we do get a written in terms of the support to go do them.

And so I'm going to keep -- at this point, I'm in the AI market. I have the full portfolio. I'm going to follow what my customers want me to do. And I'm going to ignore the noise. I mean, if you actually look at last year and all the different things that came out, and all the different noise that was out there, it was all wrong. I mean you actually analysts retracting notes. You had articles that weren't even accurate at all. I mean you had -- honestly, it was all noise. Look at our results that we're guiding, look at our outlook for this year. Look at our outlook for next year. Do you see me blinking? You don't. So yes, we're in the business. We're going to be in the business. Our customers want me to be in this business, and we're going to drive a major significant revenue company at Marvell. I'm very fired up on this topic. Thank you, Mark.

All right. Ashish wrap it?

Go for it.

Yes, I got a couple of closing statements. That wasn't it, by the way, everybody. All right. So first, thank you, everybody, for joining the call. I appreciate the interest in the company. It's always fun. Look, our business is on a very strong trajectory, okay? I mentioned on our prepared results. We had record design wins over the past year. Team did a great job. We're seeing record demand. We're on track to grow our data center revenue at or above 40% for the third straight year. And by the way, if I go back 4 years, 5 years, 10 years, this business has been growing at like 35%, 40%, 45% for a very, very long time, and it's going to continue to do that.

In fact, for next year, we're looking at that growth accelerating closer to 50% next year, and we're driving the company to try to get this company to $15 billion of revenue next year. It's -- I've been doing this job for 10 years. The team has been incredibly dedicated and we have this massive opportunity in front of us. So as I said to Ben, who asked one of the great questions, we set some very ambitious targets for the company for calendar '28, fiscal '29. Almost 2 years ago, it looked crazy. We're on track. We're on track to achieve the goals that we set.

This is the start of it. We're going to continue to update you guys on the progress of the company. And I want to end by just thanking all the Marvell employees for your focus, your commitment, and your commitment to our customers to drive the execution they're looking for, and our goal is to make Marvell one of the big winners in this once-in-a-lifetime episodic AI infrastructure build-out. So thanks, everybody, for your interest in the company. I'll see a bunch of you guys on the East Coast in New York next week.

Thank you. Ladies and gentlemen, thank you for your participation. This does conclude today's teleconference. Please disconnect your lines, and have a wonderful day.

All right. Welcome back to the Barclays Global Tech Conference. We have Chris Koopmans and Ashish Saran from Marvell. Thank you very much for joining.

Thanks, Tom.

Why don't we start off with the big question that I'm sure you're getting asked in a lot of meetings, $3 trillion of announced spend, the ability to deploy the spend in question, are we in a bubble, et cetera, et cetera. You guys are increasingly one of the largest providers of AI silicon in the world. How do you feel about this trend? Are we in the early stages of a long investment cycle? Are we going too fast? Any comments there would super helpful.

Sure. I'll start then maybe Ashish can add. So first of all -- of course, we're in the early stages of a long investment cycle. I mean this -- how you can tell that is the number one, we're constantly being asked to go faster constantly. And we're not able to deliver enough computing power for what the workloads are trying to do today. And so the whole industry is scrambling to build higher speed I/O, faster compute, faster acceleration to solve these big problems that aren't there yet, and you're seeing architectural explosion across the board as well.

So ultimately, I think if you took and said, hey, from 2022 from the launch of ChatGPT, that's probably the beginning of a 10- to 15-year investment cycle. Now those don't normally go straight line. Normally, there's ups and downs along the way, and I would certainly expect that here. But that's kind of -- I'm not really thinking about that. I'm more thinking about sort of like how do we actually get to 5 years from now and deliver everything that is needed. And ultimately, right now, I would say, if anything, our customers are asking us to accelerate and do more, right, in terms of what they're trying to do in 2026 and in 2027. So it looks really good for now.

Great. Well, you guys attack AI through a variety of different ways from the XPU all the way through interconnect. I thought it was interesting, your customers are asking you to do more. You did the Celestial deal at earnings. Maybe talk about what that adds to your product portfolio and why you felt like this was the right time to add that piece to your puzzle.

Absolutely. So if you think about the data center architecture today, and we talk a lot about the connectivity inside the data center architecture. Marvell did the, Innovium and Inphi acquisitions in 2021, which were targeting switching and the optical interconnect technology in the scale-out portion of the network. That's the rack-to-rack communication across the data center as well as the data center interconnect, which is data center to data center. Those are the longer distances. 80% to 90% of the traffic in the data center is in the scale-up portion of the data center, which is the XPU to XPU traffic inside the rack. And so today, the entire optical market, the entire switching market is only 10% of the traffic. The rest of it is really all passive copper today. There's only one major architecture that has adopted a switch, and none of them are using optics.

So our view is in the next 5-plus years, you're going to see many of those architectures move towards a switching architecture, and Marvell is investing in the leading scale-up switch road map, including both UALink and [ ESA. ] We'll be sampling our 115T -- UALink switch second half of next year. And the optical interconnect technology that needs to happen in the scale-up portion of the network, where you're talking about millimeters and inches and meters versus tens of meters and kilometers is a totally different type of technology.

And that's what Celestial AI brings to the table. And just to give you an idea, the very first chiplet that we're going to be delivering is a 16-terabit chiplet, that's 10x the state-of-the-art 1.60 in the scale-out. It's a totally different form factor, a totally different density, totally different bandwidth and totally different technology because it has to be co-packaged directly with these kilowatt XPUs and switches. And so really, between Marvell's UALink road map and ESUN road map and the Celestial AI, Photonic Fabric technology, we're positioned to be the leader in this new TAM that's growing basically from 0 today to probably be bigger than the scale-out TAM 5 years from now.

Yes. I mean just to give you a sense, right, today, we have a $3 billion interconnect business, which is all really scale out and that's 15% of traffic. So you can think about the opportunity we have with this acquisition where you have a much, much bigger TAM opening up over time. So Inphi was awesome. It was fantastic. It's done extremely well. But the market size available in scale-out is significantly larger.

And you put those two teams together, we have the leading optical interconnect portfolio in the industry. And I think that team can together solve any challenges inside the data center going forward.

So you noticed with the deal, very comprised of a lot of earn-outs. And then you also think about CPO from a scale-out perspective. It's taken a lot of years to get to fruition, and now we're talking about it more and more, but it does seem more real now than it has over the last couple of years. Could you talk about when that crossover from a technology perspective really happens? I mean you can obviously look at when you talked about revenue contribution from the company as like your first metric. But is that really a statement on market adoption? Or is it a statement around Celestial being able to compete first in that industry?

So I think generally, it's a statement of two different markets. So most of the CPO technology in the industry, including Marvell's own [indiscernible] investment in our light engine technology was focused on scale out. It's a sort of much longer reach, interoperable market. That's the one that is slow to be adopted. And if you think about that market still, I mean, pluggable, we're in -- we launched our 800 gig in 2022 pluggable technology. It's still growing into 2026. So we already are ramping 1.6T. So you should assume that's probably going to be growing through at least 2029.

We've already demonstrated 3.2T technology. So for the next 5, 6, 7 years, pluggable is still working. So that's why CPO and scale-out is still not being adopted. This is a totally different market. This is scale up. And ultimately, this is not going from pluggable optical to CPO optical. That's a world where people are going to stay pluggable as long as possible. This is a -- electrical has run out of steam and pluggable won't work because of the density and the bandwidth required. So you need a new piece of technology.

So it's really kind of an apples and oranges. I think ultimately, the CPO and scale out, it will still come, but it's still years away. I think in the scale-up world, although everybody is working as hard as possible to make electrical go as long as they can, when it goes optical has to go CPO from the beginning. And so really, the time frames that we put, which is significant revenue. I think we said $500 million in run rate exiting calendar 2027 and doubling to $1 billion run rate by the exit of calendar '28, which you'll notice is probably half of what they would need for the earn-out. So it should be much bigger than that. That's the time frame. And that's really just based on first product, first customer. I mean this -- at the end of the day, the reach of this technology can be massive inside of data center...

Yes. And Tom, in terms of timing of that revenue, yes, I would say that the reason why it's as quick as it is, is I think it's both things. It's one is the maturity of their technology, which is part of the reason why we felt they were the best choice. There's obviously a number of folks trying to go after the scale-up market from a photonic perspective. And the second one, of course, is their ability, especially along with us, where they've already got a lead customer, which is very, very public and that customer's ability to go drive this into production very quickly. We believe this will be the first large commercial deployment of photonic technology within the scale-up network. I think it's, again, it really based on the maturity of the platform they're bringing to the table.

In the intermediary period of time before we move to CPO or optics-based scale-up, you do see other technologies that are inserting themselves into the space, whether it be active copper cables, whether it be AECs. AECs have been used between racks, but also there's a view that you could be used intra-rack as well. So I saw -- it was either yesterday or this past week, time was a flat circle, the Golden Cable initiative. What are you guys doing there? Is that just your effort to work with multiple cable manufacturers, ease of use to the supply chain? Why were you so formal about that announcement there?

Yes. If you think about it, something like a cable is no different than an optical module. And our hyperscale customers want to multisource that. They don't ever want to be -- because these things, you need -- like these are copper cables. You need to be having access to copper long term. There's factory issues that could come up, right? They need multi-sourced solutions in this area, and they came to us and they asked us to provide a multisource solution. So what we've done and what we've been working on now for a year plus is enabling the entire copper cable ecosystem to develop and deliver active electrical cables that meet the standards and quality requirements of those customers.

And so our Golden Cable initiative was basically, think of it as like a reference design that's a full working design to sort of bring it to those companies and show them how to make it work. And I think Foxconn Interconnect Technology so they got it working in 2 months. Our goal is really to be able to come to the entire hyperscale ecosystem and provide multiple partners that they can purchase a working active electrical cable with our DSP. And by the way, the reason we can do that is because our DSP is so strong and so capable in this PAM4 technology that we're able to work across all these cables.

We don't have to control and define the whole thing end-to-end. And ultimately, that's what these customers want, right? Just like in the module ecosystem, we may have a customer that is almost 100% using Marvell's DSP, and we won the design with them, but they still have 3 or 4 optical module companies meeting all of their different needs with all of their different laser types.

If you look at earnings, and I want to just focus in on the PAM4 modulation and your optical DSP. The biggest change, I think, where people's numbers were and where they went was the step higher in optics, where you talked about cloud CapEx growing faster. And as a function of that, your optical business, which has tracked largely in line with CapEx, not with large hyperscaler spend has gone higher as well, excuse me -- you were very careful about splitting these up. So maybe I'll use this forum as well. What does your optical business track with? Why are you seeing that big step up to that 35-plus percent kind of range? And why does it track broadly with that number that you described in earnings?

Yes. I mean that optical interconnect technology that came in through the Inphi acquisition has been on fire, right? It's obviously very high market share. And the attach rate within the data center actually continues to grow. I think we've grown at 50% CAGR for like 4 or 5 years since we acquired that company. And so the way that we look at it going forward is demand is off the charts. I mean if -- even if you just look recently, total Marvell, I'm talking now, but optics is a big portion of it being half of our data center our Q2 was our record bookings quarter ever. Q3 was our record bookings quarter ever. We're only 5.5 weeks into Q4. We've already booked more than all of Q2.

I mean the business is doing incredibly well. And yes, we think it should be tracking above cloud CapEx because it's attached to the AI, and that's tracking above total cloud CapEx. Now I will say that the vast majority of cloud CapEx today is going into something related to AI and how do I [ stand up ] AI. So it's not like it's double on cloud CapEx, but it's higher than cloud CapEx.

Yes. And the reason we end up growing faster typically is because it's the same formula we outlined, which is we are first to market. We're first to sample. We're first to get the customer up and running, and we are first to come out with the next optimized solution. And if you can keep doing that, especially on this accelerating cadence, I think that's what gives you market leadership. So I'm not surprised that we would continue to be growing faster than kind of the underlying market.

And then as you shift to new speeds, 1.6T coming more in volume in '26, still 800G, the vast majority of connections next year. But as you move to 1.6T, what does that share profile look like? And how can you assure that you're in the same kind of position that you were in prior generations?

Yes. I mean it's sort of piggybacks on what Ashish just said, which is like you've got to be first to market with technology that works and has the right power envelopes to deliver these solutions. Marvell's 1.6T, we were first to market with our 5-nanometer solution to demonstrate and make it work with every accelerator. We were first to market with our 3-nanometer, which has the right power for this particular growth vector that's coming.

And ultimately, we have deep long-standing partnerships with all these companies. This is not something where you can just come with the DSP and win share. The qual cycles tended to have happened a year ago, right? Even if you come to market later with a product that works, getting it qualified across multiple module ecosystems, multiple laser types and the exact form factor required by all the hyperscalers takes a long time. And ultimately, what they are focused on first and foremost today is time to market.

Yes. And we said on our call, by the way, that we're seeing literally exceptionally strong demand for 1.6T. While to your point, 800 gig is growing, and it's going to be very, very strong next year. I would suspect at this point in time, we are probably the primary driver of 1.6T in the industry.

Switching gears to the ASIC business. Optics drove the big uptick in 2026 numbers. And off a very large base, it's still growing quite nicely, whatever you think cloud CapEx is into 2027. But I think the most surprising thing on the ASIC side is you talked about receiving purchase orders through the entirety of next year. Can you talk about -- is that with just the one large customer? Or were there smaller deals as well? And then secondly, into 2027, you talked about that business doubling, which was a big step-up from kind of where people had thought before. That actually has you growing at a faster rate, generally speaking, than what cloud CapEx would grow in that year. We'll see. But maybe talk about what gives you the assurance about 2026? And then what's driving that doubling into 2027?

Sure. So first of all, our 2026 numbers, calendar 2026 numbers, our custom business is still a relatively small number of sockets, right? We won -- we started this business just a few years ago. We announced a certain set of sockets last year. We announced we have more this year, but the revenue is still being driven by a small handful of sockets. And so ultimately, the reason why we're showing the growth next year is because we know what the customers want to do. And they told us -- and the reason we talked about the purchase orders for the one, and we have first of all, the lead time on these products are very long.

So you're already getting orders for products we're going to deliver in 8, 9 months from now. So within a few months, probably have orders for the entire business for next year. The comment on the one program, which is the next-generation XPU for our lead XPU customer is really just to try to assuage any concerns that we were going to participate in the next generation. We have the purchase orders for the whole year to support that growth.

The following year, there's really a couple of big things. Number one, you take the base business that we have this year that's going into next year and grow again. Number two, you add in a next XPU customer, which is really more meaningful and the more first meaningful revenue in 2027. And then number three, what we've said is our XPU attached business is starting off a really small base. We said it doubled from last year to this year, and it will probably continue to double into the future off of a small base. And so by the time you get out to 2027 calendar, it starts becoming very significant. And that's growing obviously much faster than cloud CapEx because it's sort of a newer portion of the custom market. So those are really the 3 aspects that give you context...

Yes. I mean just to put some numbers around it, right, just so everyone on the same page, you're basically implying that a business which will exit next year. For full year next year, we implied that number is probably just shy of $2 billion, call it, big large numbers. But the exit run rate, to Chris' point, is probably going to be north of $2 billion annualized when we exit Q4 next year. Just assume, again, just using like 20% cloud CapEx, that base alone grows to like close to $2.5 billion.

And then the next $1 billion plus or minus left to get to our target is really from like two big chunks, right? One of them is multiple XPUs attach socket. And then the balance is from, quite frankly, a very conservative view at this early point of the next big XPU, right? So it's not like we're expecting this $1.8 billion additional revenue from just one single place. It's actually, if you think about it, it's from really three big chunks of revenue, which is, in my view, still a fairly conservative outlook.

And you talked about $2 billion by fiscal year '29 from that XPU attach. Could you help us zoom in on this bucket? At the AI Day and the Analyst Day, you kind of looked at different sockets that you would say contributed to that. What are some of the bigger ones that we should be looking at that would be helpful for us to track?

Yes. So first of all, what we said at our custom AI event was that we thought it could get to $3 billion by calendar '28, which is 20% of what we thought would be the $15 billion TAM. What we said in this last earnings call is that just with two specific applications, we'll get to $2 billion, right? The others are with the other applications. And those two applications -- so when we announced this or when we talked about this in June, we sort of talked about all the different sockets that are in there.

At our recent earnings, we sort of highlighted that we now have two specific applications that we have multiple design wins for multiple generations and at multiple hyperscalers. So they're emerging as their own independent applications. The first one being Smart NICs and NICs, network interface controllers and the second one being CXL memory expanders controllers. And both of these are attached to both standard infrastructure and increasingly into AI infrastructure. And what we said is just those two alone will combine to be $2 billion by calendar '28. And really, it still is just starting. So it's going to grow significantly beyond that based on those two applications.

And if you think about how we got here, these are both areas that Marvell has been investing in IP for a long time, right? Smart NICs are both ARM complexes as well as high-speed network interface control and IP around SerDes and things like that. In CXL, we announced a standard product line for CXL memory expanders and near-memory accelerators a few years ago. And ultimately, just like a lot of the other things in the hyperscale market, it helps them see it and test it and say, okay, I want something custom, right? But it's sort of similar idea but I want something customized for the standard infrastructure and then I want something custom increasingly for the accelerated infrastructure. So those are two of the big use cases. We also have storage accelerators, and we also have security products that are kind of making up the rest of it.

I mean two things to add, I think, which is what we've seen and why these markets are getting more exciting. First, on the CXL side, CXL originally was designed for really basically disaggregating memory from a CPU. And that very much is a use case, and that was our initial design wins. But especially as we now look into a world where you have LLMs, which are essentially going to be storing all the what's called a key value cache, KV caching, where you don't have to regenerate the text stream every time you go and ask the next question, that requires a whole bunch of DRAM sitting very, very close to the XPU.

And that's the point which Chris was trying to make. It's not just attaching to CPUs, but once you start attaching to XPUs, as you know, in traditional AI infrastructure, the ratio of an XPU to a CPU is anywhere from 4:1 to probably 8:1. So the use case is significantly larger than what we'd anticipated. Similarly, on the Smart NIC side, our original assumption is on original design wins were really around attaching only to the custom portion of the rack level infrastructure. But as you know, all of these large hyperscalers have their own networking teams, right? So the attach rate is not just to the custom portion, but it's across their entire network. And some of these guys are driving 1 million and well over 1 million plus AI servers per year. So you can see why we have really started to move this up and it's become a much bigger use case.

I want to ask on the ASIC side. This has been something you've dealt with for over a year. First, there was concerns around your largest customer ramping into '26. And more recently, there's been concerns about, I believe, what you called Customer C at your Analyst Day a long time ago ramping in 2027. Matt was on TV last night kind of offering his opinions as to what's going on with the situation, but maybe clear it up for us. Is this a function of things not being set yet? Is this a function of there just being a lot of noise and we shouldn't be paying attention? Maybe set the record straight on where we are because we've now introduced Customer C in 2027.

Yes. I think Matt said it last night, which is there's a lot of noise. And I mean there's a lot of noise you have to pay attention to the signal -- the signal -- like we've said, we're on track with these programs. I mean those customers -- actually, I saw an article this morning, there was an article by Barron that was sort of reporting on this whole loss of a particular product and actually updated their story this morning saying that they got a response from the customer saying that, that's not true. That's just flatly not true.

The other customer, I mean, the senior leader at that customer spoke at our event a few months ago, right? That program is on track. It's doing well. We just said that and gave you a forecast for our customer revenue ramp into 2027 based on that fact. So nothing changed in the last 7 days that has changed that other than that it seems like these days, people can report whatever they want without having to do any actual research. I mean, by the way, I'm sure anybody that wants to can report that I am also talking to all of our competitors' customers and trying to win business around. Every single day, you can imagine, every single day, I'm trying to win business away from all of my competitors. If that's newsworthy, please report it.

Very helpful. I wanted to talk about NVLink Fusion. If you look at T4, there was announced compatibility with NVLink Fusion. Over time, there's a view what is going on with scale-up, back-end networks. How do you feel like this changes the dynamic? Was that something that you would have expected? And then also, there is a view that with this type of deployment, you're going to need a chiplet or an ASIC that's sitting next to all the silicon that's going into NVIDIA rack. Is that -- that seems like something you could do? Is that something that you guys would be interested in doing?

Sure. So let me take it sort of a step way back. This scale-up domain is brand new, right? I mean, obviously, NVIDIA has the NVLink and announced NVLink Fusion and Marvell has announced that we're part of that capability when any customers want to use that and want us to build custom products that would fit into that NVLink Fusion environment. UALink is new, ESUN is new. All of these things are fairly new. And I think most customers right now are working through how they're going to plan their overall architecture. And ultimately, I agree with this idea of chiplets giving them optionality, right?

When you're trying to build everything into a monolithic die, you have to make a decision right now. Now who would make a decision right now to say, I'm going to go with bet my whole farm on one of these protocols when the industry hasn't yet finalized, and there's no switch on the other end that's available yet to even make sure that it was going to work and what happens if it doesn't come out on time. So I think that, that level of optionality is absolutely there. And Marvell is also going to participate.

Now we think that UALink is the open standard protocol that has been purposely designed for this application. It's effectively taking what is generally today a PCI-based scale-up network that's used today in most of these custom XPU programs and then attaching a much higher speed Ethernet style interface to it. That's kind of what UALink is. So it's kind of purpose-built for a scale-up network in an [indiscernible] world. Having said that, there's a huge Ethernet ecosystem out there of -- and so having that capability makes a lot of sense as well. So as a leading Ethernet switch vendor, we'll build that. We're building our UALink switch. We'll sample our 115T next year alone. And so we're investing in that. And we're involved in NVLink Fusion.

So really, this is about optionality for our customers. And from a Marvell perspective, I think we're -- this move towards chiplets actually helps a lot because that's where we can help, right? And we don't -- if you think about the main XPU die building compute cores, that's not using as much of our IP. But if you think about all these high-speed SerDes and all these interfaces and things like that, that's our bread and butter. That's what we do all day long.

Would that categorize itself in like an adjacent AI opportunity, not a piece -- so that would -- you would put that in the bucket of not just an accelerator, but it would be maybe you talked about...

Well, I mean, if you remember, we announced that NVLink Fusion is also part of our XPU basically portfolio essentially. So we -- that was -- if you remember it was the beginning of the year actually, right? So I think this is not something new for us. I mean this is something which, as you can imagine, we work with all our hyperscale customers well in advance of what you see actually coming out. So this is already part of our portfolio. It's basically that flexibility we are able to provide.

You've seen accelerator guys, general purpose accelerator guys go after more system-based architectures. You've seen IP providers move potentially into more accelerator design. You guys own the capability of doing most of the design for accelerators. You own the interconnect. Would you guys ever consider going after a system-based architecture or something of that nature in throwing your hat in the ring?

Well, if you take a look at everything that we've built, including the recent announcement that we have, we're absolutely comprehending rack scale system-level architectures. Now we believe our place in the market is to be the best partner for companies that are trying to build that, whether that be an OEM that's looking to build a merchant-based platform, rack scale architecture. We have a huge breadth of IP. We can help you do that. If you're a hyperscaler looking to build your own custom rack scale system architecture, we have an incredible amount of IP. We're here to help you do that. So that's really where we believe that we fit is that we have all the capabilities from scale out, scale up, optics, electrics, switching, XPU technology and now the Photonic Fabric technology, everything that you need to help you build the best hyperscale rack size -- rack scale AI data center.

Got it. Pivoting off data center, you compressed your other businesses into a single line item saying 10% growth next year and then in '27 kind of GDP-ish growth. Is there anything that we should be looking at that changes the dynamic? Obviously, I would imagine something like 6G would offer a little more life there. But is there a reason why you're kind of saying that kind of grows at the market plus...

Just as a baseline. I mean you should imagine that the person that's running that business has a target much higher than that. We want to gain share in that...

Who reports to Chris, by the way.

And we expect to be able to gain share. But look, from a baseline perspective, that's not going to be a super exciting market that's going to grow 50% at some point, right? Yes, some of these markets like 6G, that will be a little bit lumpier, but those are markets that don't move in quarters, right? I mean those moved markets that they tend to take years to develop. And -- but ultimately, we expect that, that market will grow at least with GDP and then Marvell should go at least with the market. And in fact, I think we should be able to gain share in that market with the breadth of our capability because the way we position it is, like all this investment we're doing for the data center, and it trickles down to the rest of those markets, where they can't necessarily afford the investment in like the highest speed SerDes in the world, they certainly will take it a few years later, right, and be able to invest it and put it across enterprise networking, telecommunications networks, 5G networks, industrial networking, all those other kind of communications networks.

Yes. I mean within that, I think the enterprise portion probably does grow. I mean we use GDP again, was a very, very simple. We're talking 1.5 years, 2 years out from today. But the reality is enterprise spending typically is faster than GDP. Historically, it's IT spending roughly 2x GDP is more typical. So I think that would be my base expectation. Carrier spending does tend to track carrier CapEx historically is more on the GDP side. So again, look at that as a very, very simple framework for now, and I think we've generally done better than that.

With all these baseline assumptions, I'll tip my cap to you guys. You gave a ton of color on the earnings call around the next 2 years. Napkin math, it kind of equates to something over $5 of earnings in fiscal year '28, calendar year '27. If you look at those numbers, it's always risky to go out just a year, and you guys have expressed that. Going out 2 years, we really appreciate. But when you go out further, there's introducing more risk over time.

When you look at the broader market, you've heard a lot on memory issues and availability. You've heard about potentially digestion of CapEx. You've heard about potentially lasers. There's a variety of different constraints that are limiting some deployments. When you think about your calendar year '27 outlook, what is the one area you stay up at night worrying about kind of from your heritage in the business of what limits you in getting those deployments out?

Well, I think there's probably a couple of things. First of all, you should imagine that we comprehended all of these worries when we gave the numbers. And we could have even a bigger number. I think we gave the numbers because with the comprehension of potential supply constraints, potential -- all these issues, right? That's why I mean we mentioned that some of these parts of these are fairly conservative that we think we can actually do a lot better than that. That includes both the new XP that's ramping. That includes everything else that we talked about.

So if you step back and you say what can go wrong, right? I mean, there's obviously the -- if we just hit the brakes on spending and said slow all the spending down for a little while. I said earlier, my view is that if you take a long enough view, that's not an issue. But it could certainly be an issue for a few quarters. And if that's few quarters happen calendar '27, then that would be a problem. But ultimately, all that other stuff, if you think about the design cycles, they're on track. We won all the designs. We've won the business. They're all in execution.

We've been kind of coming to this for years. If you think about this optical momentum that we've built, right, and the strength that we've built in that business, switching, we bought that product line. It was single-digit millions of dollars in revenue. We grew to $150 million and then $300 million. Now we're saying $500 million next year. I mean these things take a long time to kind of get to this point. And now we have a number of these growth drivers that have now are sizable. Our XPU attach, we said it was really small and it doubled into this year. It will probably keep doubling going forward. So that starts to get towards multiple billions of dollars. So once you sort of have these different sort of growth engines, the flywheel spinning, we have a lot of confidence in hitting those numbers.

Yes. I mean most of the -- which is why we use CapEx as kind of as an index, right, so that you all can basically follow it along with us. And in most of our businesses, we are talking about growing in line or perhaps a little bit above CapEx. I think the only place where we're actually going a little bit above that number really is in custom. And even in custom, it's not for the entire business. It's really for a couple of product cycles, which obviously we're very, very far along on, right? So if you really think about it, really, all I'm really saying is, look -- and I'm assuming a fairly modest 20% CapEx growth assumption for calendar '27. So it's not even close to what this -- I mean, this year is way higher.

Next year, I think the whole market thinks is somewhere north 30%. So I've actually taken that down to 20%. And in most businesses, I'm saying it's at or roughly at that rate. It's only in the custom part where because we obviously have very good visibility to some very strong product cycles is where we are saying, hey, hey, that business doubles, that grows way above CapEx, right? So I would -- that's kind of my point of -- it's really not -- once you kind of go through the chunks, it isn't that big growth, even though it sounds like it, it's really tied to CapEx primarily.

Well, very helpful. Things sound great. Thank you, Chris. Thank you, Ashish, for being here. Appreciate it.

Thank you.

Good afternoon and welcome to the Marvell Technology Inc.'s Third Quarter of Fiscal Year 2026 Earnings Conference Call. [Operator Instructions] Please note, this event is being recorded. I will now turn the conference over to Mr. Ashish Saran, Senior Vice President of Investor Relations. Thank you. You may begin.

Thank you, and good afternoon, everyone. Welcome to Marvell's Third Quarter Fiscal Year 2026 Earnings Call. Joining me today are Matt Murphy, Marvell's Chairman and CEO; Willem Meintjes, CFO; Chris Koopsmans, President and COO; and Sandeep Bharathi, President, Data Center Group.

Let me remind everyone that certain comments made today include forward-looking statements, which are subject to significant risks and uncertainties that could cause our actual results to differ materially from management's current expectations. Please review the cautionary statements and risk factors contained in our earnings press release, which we filed with the SEC today and posted on our website as well as our most recent 8-K, 10-K, 10-Q and other documents filed by us from time to time with the SEC. We do not intend to update our forward-looking statements.

During our call today, we will refer to certain non-GAAP financial measures. A reconciliation between our GAAP and non-GAAP financial measures is available in our earnings press release. As we discussed in our second quarter earnings call, going forward, we are consolidating our non-data center end markets into a single new communications and other end market. The composition of our data center end market remains unchanged. Our earnings press release for the third quarter reports leveled by end market in both the [ buyer ] format as well as the new go-forward format.

Please note that today's call will be longer than typical as we will be discussing the acquisition announced today in addition to a number of extensive updates on our business. You may also find additional details on this transaction in the press release and Form 8-K be filed with the SEC today and a presentation posted on our website on the Investor Relations page.

Let me now turn the call over to Matt for his comments on the quarter. Matt?

Yes. Thanks, Ashish, and good afternoon, everyone. Settle in, okay? We have a lot of good stuff to talk about today. For the third quarter of fiscal 2026, Marvell delivered record revenue of $2.075 billion, reflecting a 3% sequential increase and strong 37% year-over-year growth. Revenue was above the midpoint of guidance, driven by stronger than forecasted demand in our data center end market. As a result, non-GAAP earnings per share of $0.76 exceeded the midpoint of guidance by $0.02. Excluding revenue from the divested automotive Ethernet business, the implied revenue growth for Marvell's go-forward business was approximately 6% sequentially and 41% year-over-year. Momentum in our data center business remains strong with revenue growing 38% year-over-year, fueled by robust AI demand.

We also saw a strong recovery in our communications and other end market, where revenue grew 34% year-over-year as reported and nearly 50% year-over-year, excluding the automotive Ethernet business. We expect growth to continue in the fourth quarter with total company revenue forecast at $2.2 billion at the midpoint. We expect this momentum to continue throughout next fiscal year and beyond. I will provide more context on our numerous growth drivers later in the call. Before discussing our end market, I'm excited to share details on the strategic acquisition we announced today of Celestial AI, which brings an entirely new disruptive technology, a photonic fabric platform purpose-built for next-generation scale-up interconnect. This acquisition is the latest in a series of decisive moves to further strengthen our data center portfolio. Since 2019, we have continued to increase our focus on data center, divesting our WiFi business and acquiring Avera, Aquantia, Inphi and Innovium. These transactions have driven significant revenue growth and scale and have each proven to be an absolute home run. This year, following the divestiture of our automotive Ethernet business, we are continuing to double down on data center with the acquisition of Celestial AI. This positions us to further capitalize on the massive opportunity in accelerated infrastructure.

The acquisition is expected to close in the first quarter of next year, subject to customary closing conditions, including regulatory reviews in the United States and will remain a separate independent company through the regulatory process. AI is reshaping data center architecture as an unprecedented speed. Next-generation accelerated systems are no longer confined to single rack, they are evolving into multi-rack scale-up fabrics that connect hundreds of XPUs in a high-bandwidth ultra-low latency and to any fashion. These advanced fabrics demand purposeless interconnects, engineered to deliver the performance and efficiency required at scale, creating a new TAM for companies like Marvell. Industry analysts are forecasting the merchant portion of the scale-up switch market to approach $6 billion in revenue in 2030. On the interconnect side, we are seeing the dollar content for optics of the same magnitude as a scale-up switch as the optical interconnect attaches to both the XPU and the switch, the opportunity actually doubles, meaning over $10 billion. These are both very large and exciting incremental opportunities for Marvell.

As first evaluated Celestial AI, it reminded us of our early look at Inphi and the products we saw in their PAM technology to transform the scale-out interconnect market. We see even greater potential for Celestial AI's photonic fabric to transform the scale up interconnect market. Interconnect technology is as critical as switching and scale up networks to enable hundreds of XPUs to be tightly coupled together. This is driving a massive increase in the number of links in the network and overall system bandwidth, therefore, creating the need for a fabric which can span across racks. Copper-based interconnects used in today's scale-up systems are approaching their fundamental limits in reach and bandwidth, creating a compelling need for optical solutions. Celestial AI's photonic fabric technology platform was purpose-built for this inflection. It enables large AI clusters that scale both within and across racks using a high bandwidth, low latency, low power and cost-effective optical fabric. This breakthrough enables a true optical solution with greater than 2x the power efficiency of copper interconnects, but with far longer each and significantly higher bandwidth.

In addition to exceptionally low power consumption, Celestial AI solution provides nanosecond class latency and excellent thermal stability, which enables deeper levels of optical interconnectivity into XPUs and switch systems. The thermal stability of Celestial AI's photonic fabric technology is a significant competitive differentiator. It enables reliable operation in the extreme thermal environments created by large multi-kilowatt XPUs. This allows the photonics technology to be co-packaged vertically with the high-power XPUs and switches in a 3D package, enabling the photonic connection to be made directly into the XPU rather than from the edge of the die. This stands in sharp contrast to many other CPO implementations for the photonics engine sits adjacent to the XPU and must connect at the die edge. Celestial AI's approach results in a more compact and integrated solution, freeing up highly valuable die edge beachfront which can be repurposed to significantly increase the amount of HBM within the XPU package. Eliminating beachfront I/O constraints also significantly increases the amount of package bandwidth possible for XPUs and switch systems.

Celestial AI's first-generation product is a photonic fabric chiplet or PF chiplet, which integrates all the required electrical and optical components, including drivers, TIAs, equalizers, SerDes, microcontrollers, modulators, photodiodes and wave guys, all into a compact form actor. This is the industry's first scale-up optical solution delivering an unprecedented 16 terabits per second of bandwidth in a single chiplet, 10x the capacity of today's state-of-the-art 1.6T ports used in scale-out applications. Its compact form factor allows multiple PF chiplets to be co-packaged with XPUs and the scale-up switches on the other side of the link to further increase total bandwidth. Celestial AI is deeply engaged with multiple hyperscalers and ecosystem partners who recognize the disruptive potential of this technology. Notably, Celestial AI has already secured a major design win with one of the world's largest hyperscalers who plans to use Celestial AI PF chiplets in its next-generation scale-up architecture. These PF chiplets will be co-packaged into both the hyperscalers custom XPUs and the scale of switches providing connectivity. This is expected to be the industry's first large-scale commercial deployment of optical interconnects for scale-up connectivity.

Beyond connecting XPs and scale-up networks, the photonic fabric technology platform can enable a wide range of transformational applications over time. First is a pooled memory appliance that uses Celestial AI's photonic fabric to optically connect multiple XPUs to large shared external segregated memory bank. A second use case for Celestial AI's photonic fabric to replace traditional electrical die-to-die connections in multi-die packages. This is just the beginning of a broad set of new applications which can be enabled from this technology. After close, we expect meaningful revenue contributions from Celestial AI to begin in the second half of fiscal 2028. Our base case forecast show Celestial AI's revenue reaching a $500 million annualized run rate in the fourth quarter of fiscal 2028, doubling to a $1 billion run rate by the fourth quarter of fiscal 2029. Following the close of the transaction, we look forward to welcoming the Celestial AI team to Marvell. Celestial AI brings one of the industry's strongest photonic interconnect engineering groups with deep expertise in optics, advanced packaging and high-speed interconnect architecture and systems. In addition, the CEO, founders and key executives from Celestial AI will assume leadership roles at Marvell, continuing our successful integration blueprint from prior acquisitions. These leaders have been at the forefront of Avation and scale up switching and photonic interconnects and their technical depth and strategic insight will play an important role in shaping Marvell's next phase of growth.

Okay. Now let me transition back to Marvell's current business and outlook. As you may recall, on September 24, I hosted a virtual call with investors where I outlined a framework for Marvell's revenue growth for fiscal 2027. At that time, we indexed our data center growth potential to cloud CapEx, which was expected to grow 18% next year. Since then, cloud CapEx growth expectations have increased to over 30%. Additionally, we have seen strong demand increases for our products for next year. As a result, our outlook for next fiscal year is even stronger than the expectations we discussed in September. We expect our Interconnect business, which is roughly half our overall data center revenue to continue growing faster than cloud CapEx next year even with the higher outlook. We expect our custom business, roughly 1/4 of our overall data center revenue to grow by at least 20% next year, also from higher than prior expectations. As a reminder, in the near term, this business remains tied to a few specific sockets. We expect custom growth next fiscal year to be higher in the second half and do not expect any air pockets in custom revenue.

Next year's custom revenue forecast comprehends a transition to a next-generation XPU at a large customer. And I would note that we already have purchase orders for the entirety of next fiscal year's current forecast for this next-generation program. Our revenue forecast for this program remained consistent with our prior expectations. As we look beyond fiscal '27, we have several high-volume customer designs in development with meaningful revenue expected from these programs in fiscal 2028, consistent with our prior communications. For the remaining quarter of our data center business, which includes storage, switching and other products, we now expect revenue to grow by at least 15% next year, up from our prior expectation of 10% growth, driven in particular from increased demand for our switching products. Adding all of this up, we now expect Marvell's data center revenue to grow year-over-year by more than 25% next fiscal year. Please note that this forecast does not include any revenue from the pending acquisition of Celestial AI. And for our communications and other end market, we continue to expect 10% revenue growth next year. Putting it all together, we are looking forward to a strong fiscal 2027.

Let me provide more details for each of our end markets. In our data center end market, we delivered record third quarter revenue of $1.52 billion, representing 2% sequential growth and 38% year-over-year growth. Revenue exceeded our guidance for flat sequential performance driven by increased demand across our networking portfolio. Our industry-leading PAM DSPs, DIAs and drivers continue to see strong demand, with revenue from our optical interconnect businesses growing by double digits sequentially on a percentage basis. Our data center storage and switch businesses also posted double-digit sequential revenue growth on a percentage basis. As expected, this strength was partially offset by a sequential decline in our custom revenue due to lumpiness in demand.

Looking ahead to the fourth quarter, we expect revenue growth from our data center end market to accelerate growing sequentially in the high single digits on a percentage basis and approximately 20% year-over-year. This growth is expected to be driven by a rebound in custom and continued growth in interconnect, switching and storage. Let me now highlight broader trends we're seeing with our established data center businesses and our newer growth initiatives. I'll start with our interconnect business, where we offer the industry's broadest and most comprehensive high-speed connectivity portfolio. As our PAM DSP products enter into their fifth year of 800-gig production, demand for our solutions continues to accelerate underscoring strength of the platform we have built through more than a decade of sustained investment in core technology. Our playbook is simple. First to market, first to ramp with timely follow-up optimized solutions to maintain leadership. We did this at 400 gig, 800 gig and now 1.6T, where we established early leadership with our first 5-nanometer solution, which sampled in February 2024, we then accelerated the launch of our Optimized 1.6T solution and sampled our 3-nanometer product just 1 year later in February 2025. As a result, we are enabling volume production of pluggable 1.6T transceivers across the industry.

We began shipping our 1.6T products in the second half of this fiscal year and are seeing exceptionally strong demand heading into next year. This consistent execution enables us to secure qualifications at major customers well ahead of competitors, reinforcing market leadership. While 1.6T has a long life cycle ahead, we have already demonstrated at the Optical Fiber Conference this past April, 400 gig per lane technology to drive the next industry transition to 3.2T. The demonstration was on 3-nanometer technology, but we expect production deployments, which are expected in calendar 2028 to require 2-nanometer solutions to optimize module power. In addition to our PAM portfolio, we are also enabling longer reach connectivity with our coherent light solutions to support campus-wide data centers in the era of million GPU AI clusters. We introduced our 1.6T coherent life solution last year, expect to start shipping next year, and we are on track to deliver our 3.2T solution the year after. Now complementing our DSPs, our high-performance analog TIAs and drivers remain foundational to our electro optics leadership. Our TIAs have a significant performance lead at 1.6T, and we are seeing strong broad-based demand for our products, which are enabling the entire ecosystem. We have also secured several LPO sockets across multiple hyperscalers and are leading this emerging category as well, although deployments remain relatively small today in the context of a very large transceiver market.

Turning to 2 of our newer interconnect growth drivers, AECs and retimers. Both markets are undergoing a shift to high-speed PAM-based solutions, an inflection point that is perfectly aligned to Marvell's strengths. For the past year, we've been collaborating closely with the cable ecosystem to enable 100 and 200 gig per lane AECs and we are now on the cusp of substantial product ramps. We have secured design wins with significant share positions at 2 Tier 1 U.S. hyperscalers, along with multiple wins at emerging hyperscalers. We are seeing strong demand for our AEC DSPs, and we expect our share to continue to grow as PAM-based 100 and 200-gig technology becomes dominant. Our PCIe Gen 6 retimers are also gaining broad traction. We are currently engaged with more than 30 customers and partners, including hyperscalers, cable partners and system OEMs and ODMs. We already designed in more than 10 sockets, and we expect to enter production in the second half of next year with full revenue contribution in fiscal 2028. We expect our AEC and retimer revenue in aggregate to more than double from this year to next year.

Turning to our data center switching business, which continues to gain momentum. We expect revenue to exceed $300 million this fiscal year. We expect strong sustained demand for 12.8T products, reflecting our key customers' plans to rely on 12.8T as a workhorse in their scale-out network for several more years. In parallel, we've begun shipping our next-generation 51.2T products with a strong ramp expected next year. As a result, we now expect our day center switch revenue to surpass $500 million in fiscal year faster than what I had indicated last quarter. We will also introduce our 100T products next year as we continue to execute our long-term road map. We are also accelerating our scale-up switch efforts. These next-generation solutions are complex as 100T scale-out switches with high ratings supporting up to 576 ports. We are fast tracking our scale-up switch development by leveraging our in-house high-speed, low-power SerDes and experience in developing extremely large reticle size chips. We are deeply engaged with key customers and partners and are on track to sample our UALink 115T and 57T solutions in the second half of fiscal 2027 with volume production expected in fiscal 2028. In parallel with our UALink development, we are also collaborating closely with key customers under E.SUN solutions, completing a scale-up road map to address both standards.

Turning to our custom business. We expect accelerated growth over the next several years. fueled by our growing portfolio of design wins. At our custom event in June, we disclosed a total of 18 XPU and XPO attach socket design wins. Several of these are already in volume production with the remainder on track to ramp over the next couple of years. Since that event, our team has secured additional custom sockets, which represent more than 10% of the $75 billion lifetime revenue opportunity funnel we outlined in June. These new wins include multiple XPU attached sockets and XPU, an emerging hyperscaler and most recently, a design win for an electrical I/O chiplet inside an XPU. This is a new trend we see emerging where customers and partners are partnering with Marvell to gain access to our high-performance networking technology to be integrated along with their core compute engines within multi-die packages which are becoming more prevalent in a reticle size constrained world. This provides Marvell another avenue for custom growth and sockets, which were otherwise not available to us as full XPUs.

Now let me provide additional perspective on the rapidly developing XPU attach market. These attached devices offload specific functions such as network I/O, memory expansion and security, member expansion and security from XPUs, GPUs and CPUs, freeing them up to exclusively focused scarce compute resources on the primary AI workload. We now have more than 15 XPU attach wins, and today, let me highlight 2 major use cases emerging across multiple hyperscalers as they architect their next-generation custom accelerated infrastructure. The first use case is for custom foundational and smartNIC, and Marvell has already secured multiple design wins across several hyperscalers. Our customers plan to attach these mics, not only to their custom accelerators, but increasingly to their broader AI server fleets, which at large hyperscalers can exceed 1 million units or more annually. The second use case we are seeing emerge in the X2 attached market is for CXL-based products that enable memory expansion and acceleration to overcall the memory wall challenge. We made early strategic investments in CXL several years ago, and we have now secured 5 unique sockets across 2 Tier 1 hyperscalers, U.S. hyperscalers and are deeply engaged with the third.

The first custom CXL design win started shipping already in the first quarter of this year and is entering volume production now. A second socket focused on near-memory compute is expected to enter production a year from now. The remaining CXL design wins are slated for production in calendar 2027. Our solutions technical advantages include support for both DDR4 and DDR5, larger memory capacity and compression along with deep partnerships with the leading memory and CPU providers. While our initial wins centered on offloading from CPUs, more recent wins attached directly to XPUs, which are deployed in far greater numbers were several upcoming high-volume CXL production ramps, Marvell is leading the transition to next-generation memory architectures. We expect the XPU attached market to continue to evolve at a rapid pace, and we are very encouraged to see the attach rate of our solutions exceeding our initial expectations. Based on designs we have already won just for the NIC and CXL use cases, we have line of sight to revenue exceeding $2 billion by fiscal 2029 and a significantly higher forecast in the following years. This is why we are so excited about our data center business, interconnect switch, XPU, XPU attach, storage, scale up, scale out, we are everywhere in the AI rack. And we are just getting started with what we expect to be a massive TAM ahead of us.

All right. Let me now move to our communications end market. where we delivered $557 million in third quarter revenue, which grew 8% sequentially and 34% year-over-year. Excluding revenue from the divested automotive Ethernet business, the implied revenue growth for Marvell's communications end market for the third quarter would be closer to 20% sequentially and 50% year-over-year. These strong results were driven by normalizing customer inventory levels and strong adoption of our refreshed product portfolio at both our enterprise networking and carrier infrastructure customers. Looking ahead to the fourth quarter, we expect revenue from our communications end market to grow sequentially in the low single digits on a percentage basis with year-over-year growth of approximately 25% as reported and closer to 40% year-over-year, excluding our former automotive Ethernet business. We expect strong sequential growth from Carrier and ongoing growth from enterprise to be partially offset by a steep seasonal declines in our consumer business. We expect the enterprise networking portion of our communications end market to reach an annualized revenue run rate of approximately $1 billion in the fourth quarter, which would reflect the complete normalization of customer inventory levels in that business. Going forward from this $1 billion annualized revenue run rate, we expect this business to grow in line with enterprise IT spending. While our carrier business has also been recovering and our fourth quarter guidance implies the business to almost double from the year ago quarter, we see continued recovery until this business also settles into its long-term growth trajectory, which would be in line with carrier CapEx.

So in summary, during the third quarter of fiscal 2026, we continue to expand operating margins, grow earnings per share and set new revenue records. We executed our $1 billion accelerated stock repurchase program in addition to repurchasing $300 million of stock through our ongoing buyback program funded by our growing operating cash flow. Looking ahead, we expect momentum to continue in the fourth quarter with total company revenue forecast at $2.2 billion at the midpoint, representing 6% sequential and 21% year-over-year growth. Excluding revenue from our former automotive Ethernet business implied year-over-year revenue growth for Marvell's go-forward business would be approximately 24% at the midpoint of our forecast for the fourth quarter.

As I noted in my opening remarks, we are seeing robust demand signals and strong bookings across our entire portfolio, positioning us for a strong fiscal 2027 and even faster growth in fiscal 2028. Customers are planning to add substantial AI capacity over the next several years and are partnering closely with us on long-term technology road maps and coordinated capacity planning. In addition to benefiting from rapid market expansion, we have several of our own unique growth drivers. Taken the other, we expect strong market tailwinds and new product cycles to drive significant growth inflections ahead of us. As a result, we see a path for our data center revenue growth in fiscal 2028 to accelerate meaningfully above the 25% growth we expect in fiscal 2027. So look, we covered a lot of ground today.

And so before I close, let me just quickly highlight a few key takeaways. First, we have activated Marvell's M&A playbook and expect to close the transformational acquisition of Celestial AI in the first quarter of next fiscal year, enabling us to fully capitalize on the massive scale-up opportunity. Second, our interconnect business is firing all cylinders. Our electrooptic interconnect platforms continue to lead the market with world-class road maps across the board and accelerating demand. Finally, when you put it all together, we are positioned for several years' exceptional performance, building on this fiscal year's projected revenue growth of more than 40%. And I look forward to updating you on our progress over the coming quarters.

And with all of that, I'll turn the call over to Willem for more detail on our recent results and outlook.

Thank you, Matt, and good afternoon, everyone. Let me start with a summary of our financial results for the third quarter of fiscal 2026. Revenue in the third quarter was $2.075 billion, growing 37% year-over-year and 3% sequentially. Data center is our larger end market, contributing 73% of total revenue. Our communications and other end market contributed the remaining 27% of revenue. GAAP gross margin was 51.6%. Non-GAAP gross margin was 59.7%, an increase of 30 basis points sequentially.

Moving on to operating expenses. GAAP operating expenses were $712 million, including stock-based compensation, amortization of acquired intangible assets, restructuring costs and acquisition-related costs. Non-GAAP operating expenses came in at $485 million, in line with our guidance. Our GAAP operating margin was 17.2%, while our non-GAAP operating margin was 36.3%. I'm pleased that we drove a 150 basis point sequential increase in non-GAAP operating margin. For the third quarter, GAAP earnings per diluted share was $2.20, including the gain from the divestiture of the Automotive Ethernet business. Non-GAAP earnings per diluted share was $0.76, reflecting year-over-year growth of 77%, which is more than double the pace of revenue growth demonstrating the significant operating leverage in our model. Non-GAAP earnings per diluted share increased 13% sequentially.

Now turning to our cash flow and balance sheet. Cash flow from operations in the third quarter was a record $582 million, growing approximately $121 million from the prior quarter. Our inventory at the end of the third quarter was $1.01 billion, a decrease of $37 million from the prior quarter. During the quarter, we executed our $1 billion accelerated repurchase program. In addition, we repurchased $300 million of our stock through our ongoing capital return program and returned $51 million to shareholders through cash dividends in the quarter. As of the end of the third quarter, our total debt was $4.5 billion, with a gross debt-to-EBITDA ratio of 1.47x and a net debt-to-EBITDA ratio of 0.58x. Our debt ratios have continued to improve as we have driven an increase in our EBITDA. As of the end of the third fiscal quarter, our cash and cash equivalents were $2.7 billion, an increase of $1.5 billion from last quarter reflecting the addition of proceeds from the divestiture of our Automotive Ethernet business and ongoing cash generation from operations, offset by a capital return of $1.35 billion between stock repurchases and dividends.

Turning to our guidance for the fourth quarter of fiscal 2026. We're forecasting revenue to be in the range of $2.2 billion, plus or minus 5%. We expect our GAAP gross margin to be between 51.1% and 52.1%. We expect our non-GAAP gross margin to be between 58.5% and 59.5%. Looking forward, we anticipate that the overall level of revenue and product mix will remain key determinants of our gross margin in any given quarter. For the fourth quarter, we project our GAAP operating expenses to be approximately $741 million. We anticipate our non-GAAP operating expenses to be approximately $515 million, growing from the prior quarter as we continue to invest in the business, and anticipate higher employee bonus payouts, reflecting a strong expected finish to the fiscal year.

For the fourth quarter, we expect GAAP and non-GAAP other income and expense, including interest on our debt to be approximately $30 million. We expect our non-GAAP tax rate of 10% for the fourth quarter. We expect our basic weighted average shares outstanding to be $850 million, and our diluted weighted average shares outstanding to be $857 million. We anticipate GAAP earnings per diluted share in the range of $0.31 to $0.41. We expect non-GAAP earnings per diluted share in the range of $0.74 to $0.84. Looking ahead to fiscal 2027, Matt already provided an update on our strong revenue growth expectations. We intend to continue to invest in growing our business while driving operating leverage and expect our non-GAAP operating expenses to increase at roughly half the rate of the revenue growth next fiscal year. Keep in mind that we typically see a mid-single-digit sequential increase in OpEx on a percentage basis in the first quarter. This forecast for next year's OpEx does not include any additions from the acquisition we announced today. I will provide that forecast separately. Regarding taxes, we expect our non-GAAP tax rate to move to approximately 12% next fiscal year.

Turning to the acquisition we announced today. Post closing, we expect the addition of Celestial AI to add approximately $50 million in annual operating expenses. We expect Celestial AI to start generating meaningful revenue in the second half of fiscal 2028, at which point it is expected to become accretive to our non-GAAP earnings. We plan to fund the acquisition through a combination of stock and cash on hand and do not intend to take on additional debt. We have a strong balance sheet and are generating robust operating cash flow. As a result, in parallel with paying for the acquisition, we plan on continuing to return capital to stockholders through dividends and buybacks. In summary, we're executing on our strategy to drive strong revenue growth while delivering operating leverage. In addition to organic investments, we are actively deploying our strong balance sheet to acquire a transformational asset that we expect to further strengthen our capabilities and increase our addressable market.

With that, we are ready to start our Q&A session. Operator, please open the line and announce Q&A instructions. Thank you.

[Operator Instructions] Our first question is from Ross Seymore with Deutsche Bank.

Matt, I appreciate all the details you gave about how to think about next year. If I run through those numbers, basically, it sounds like you're implying somewhere around $10 billion in revenue for next year. So I guess, first of all, is that in the right ballpark? And then back in June, you gave a longer-term target for fiscal '29 for your business, especially on the AI side of things. How does what you're looking for next year gets you aligned to those long-term targets?

Yes. Thanks, Ross. I think that's actually a great way to tee up the Q&A here. So yes, a couple of things. I think you're absolutely in the ballpark when you add up the numbers I gave you on $10 billion for next year. And I think that's a great target, actually, by the way, that motivational for us as a team to go drive. And just as a reminder, this is just based on the Marvell organic plan, no M&A contribution. A couple of things about next year, an then I'll actually give you some commentary on how you think about that slope we're on for next year. Another thing I would note, I didn't say in my prepared remarks, but we do expect sequential revenue growth next year, every quarter. So year-over-year, we see a nice growth throughout the year. But I would say the second half of the year stronger than the first and with, I think, a really, really compelling exit rate to fiscal '27.

Now it's a little far out to go all the way to '29, let's say, but let me just give you a path sort of on where we're headed and how we think about fiscal '28 and how we build on that strong second half we see next year. So a couple of things maybe to talk about those same businesses. So first on custom. And just as a reminder, I mean, basically, we quadrupled that business from calendar '23 to '24, we doubled it from founder '24 to '25, we're saying it's going to be up about 20% this next year. But then when I look at the year after with all this goodness from XPU attach plus a new meaningful XPU socket ramping and the other programs continuing, we see the custom business action in fiscal '28 doubling off of '27. So we see a reacceleration big time in that year. And that puts us on a nice trajectory towards our growth targets fiscal '29 in custom coming off a strong second half next year and then in fiscal '28.

On interconnect, look, it's early, but we do expect that business should continue to outgrow CapEx. It's done that for a long time. Look, you just said, hey, 20% CapEx growth in fiscal '28 and nobody knows what the number is, but just pick it in there for now, we'll definitely grow above that. I would say our customer forecast support a much higher number than that. But just as a base case, just assume CapEx is 20% that year. And so -- so optics will grow faster. And then for storage switch and the other part of data center just assume 10% growth in fiscal '28 over '27. I think that's a reasonable set of assumptions to use based on what we see. And so when you actually add all that together, you come up with a number of bottoms up, which looks more like 40% growth in data center in fiscal '28. So that's what I was talking about in my lead-in, which was strong data set 45% growth basically this year in data center, targeting 25% next year and then 40% the year after. And if you actually look at it over like the cycle where we started back in calendar '23, that's like a 50% kind of compounded growth rate, we'll be growing our data center business. So the '28 numbers I'm laying out for you are not totally crazy.

And then just for reference, if you just sort of plug in comms and other growing at GDP for fiscal '28 just to keep it simple. Again, it's too far out you basically get Marvell growing like another 30% in fiscal '28, which would be above where we looking for '27. So kind of a long answer, but maybe what's been a lot of investors' mind is how do you get from where we are to where we're going, but certainly, we're very optimistic about our outlook over the next couple of years. And then things like Celestial AI and some of these other big growth drivers, those are going to kick in starting in fiscal '28, but then really in '29 and '30 and beyond. And I think through the end of the decade, it's looking very bright for Marvell. Thanks for the question.

Our next question is from Harlan Sur with JPMorgan.

And congrats on the Celestial acquisition. Today, Matt, your lead AI customer announced their next-generation 3-nanometer AI XPU product. And I think you just said you have secured purchase orders for this program for the entirety of next year. But your lead customer also preannounced their next-generation 2-nanometer XPU product today as well, which we believe we're also involved with, especially now with the Celestial team. I remember at the June custom AI event that the team talked about concurrent design programs. In other words, at the same time, you're towards the tail end of your customer's 3-nanometer design. You're already starting to work with customers on their next-generation 2-nanometer designs. During our fireside chat in September, you talked about the team being heads down focused on 2-nanometer designs. You even talked about next-generation A16 and A14 technologies. Can you just give us an update on your sub 3-nanometer design win pipeline does include both XPU, XPU attach programs? And what's the time line for these programs to ramp into production?

Yes. Thanks, Harlan, for the question. And yes, just in the spirit of customer confidentiality and details, I can't go into too much. But what I would say, which is incorporated into our numbers is that our product transition from where we are today with our lead XPU customer to the next one is baked into all the numbers I gave you and I got the backlog and I got the orders and we got great visibility there. On the 2-nanometer, very exciting. I mean there's a number of programs that we're working on in this area, and that's going to be a workhorse process technology for us. But same as I said, and I would just say that the design funnel keeps increasing there. And the power benefits you see are compelling. And I think that will continue. I mean that's really where AI has sort of kicked in to keep the Moore's Law train running is really the just the power savings are worth real OpEx dollars when you can save power dissipation from one generation to the other. So nothing really new to report there other than just heads-down execution and do see strong product ramps coming over this time period I gave you, especially in the fiscal '28 where you'll start seeing some of the 2-nanometer products ramp. And my team internally just to give them a shout out is executing extremely well. The whole team you saw that got up there at the AI investor event in June by engineering leadership, executing extremely well across the board on the core IP, the nodes, the packaging, you name it. So we're really firing on all cylinders internally. Thanks, Harlan.

Our next question is from Tore Svanberg with Stifel.

Yes. And congratulations on the acquisition. I had a question on Celestial AI, Matt. So when you gave those $500 million and $1 billion target, would that be for the PF Link products only? Or would that also include some of the potential businesses with memory?

Yes. Yes. A couple of things to note, maybe at the highest level. The first is, yes, the revenue targets and also the earnout that we're all going to drive for is all based on Celestial AI in totality. Now the reality is from a lead perspective, the PF chiplet is sort of what's going to go first. But everything is on the table, and there's just tremendous activity that, that team has driven punching way above the weight in the industry in terms of the engagements they have. So those are all in numbers, but clearly going to be driven more from the PF chiplet side in terms of the revenue build in end of fiscal '28 and then the end of fiscal '29.

Our next question is from Chris Caso with Wolfe Research.

Also a question about Celestial. And with that revenue ramp that you're expecting at the end fiscal '28, beginning of fiscal '29, can you talk about the breadth of that? And obviously, I'm sure you're not willing to name the customers right now. But is it a fairly narrow customer base, what's the -- and going over time, how diversified is that revenue stream?

Yes. No, great question, Chris. I think that the engagement is certainly broad, but remember, we're -- there's a bold effort across the industry to really bring this product into volume stable production, and it's going to take real big companies and a few of them to do that. And so yes, we have strong engagement across the board, but there will be -- and we're fortunate with this that we have -- as I said on the call, we have one Tier 1 hyperscaler that's we're engaged with on this that I think is a great partner and a great teaching customer and a great way to -- great customer to collaborate with to really go make this happen. So it's very exciting to see Celestial come into Marvell because -- just as a reminder, we have an incredibly strong internal silicon photonics organization. I mean this team that we got from Inphi pioneered this technology. We drove it into high-volume production and market leadership in the 100 and the 400-gig ZR business, now 800 gig. It's all the DCI stuff they enable. That's on the back of very, very innovative silicon photonics technology. So we have a ton of know-how here. So then when we bring in Celestial, we're all going to benefit from that common set of learnings and then having a lead customer pulling us through is very exciting. So I think this is just the first phase here, Chris, that you're going to see. But certainly, the interest is there across the industry. And this is just so evident to us over the last couple of years and really pronounced at the last strategic review. It's a process I run every year where we do our capital allocation review with everybody, and I was a couple of months back, and we're just staring at this just giant transition to photonics inside the data center in that TAM. And we looked at could we do it ourselves, could we do it with the partner and we just concluded this is a home run to bring them in with our internal team plus really a lead customer to pull us through the first wave. But beyond that, we see very broad adoption beyond that time frame. Yes.

Our next question is from Harsh Kumar with Piper Sandler.

Congratulations, Matt and the team and a lot of good things happening. Matt I had two. So I'll just ask them both together. The first one is on custom. You specifically mentioned at least 20% growth for next year. Seems like you've got a lot of good things happening, a lot of good stuff ramping. And I was hoping that you could book in that number for us. So at least 20% to me seems like that's what's in the bag. But maybe you could give us a sense of what -- if everything worked out right for you, what could be a normalized 2027 FY growth rate for custom? And then the second question was you gave us a lot of color all the way to FY 2028. And this is kind of not the norm on Wall Street, most companies go out of a quarter, as you well know. So my question is, and this we've been getting from a lot of investors already is what is your comfort level? And what is the visibility on some of this long-term revenue materializing that you're talking about?

Yes, sure. A lot in there, but let me unpack it. I think the first is, you should definitely model the 20%, and I think that's -- that is a good number. That's a good safe base case number. And remember, it's a handful of programs today. So I think I would just go for that now. Clearly, we're going to be ready if people want to do more, but I am also mindful of some of the history on this custom business where either people got ahead of themselves or there's a lot of noise in the system and it's just -- look, we've got strong backlog, we got that covered. That's great. If we do better later, we'll definitely update you guys. But we got a very, very good outlook for next year in that business. And then obviously, the year after it really builds. And kind of part of your question is it's not 20% sort of linearly, meaning the second half, especially in custom and the exit rate is much higher than where we are now. And that's simply because we're building all this momentum into fiscal '28.

And then yes, on the sort of guide quarter at a time, that's definitely been RMO historically. I think given the multiyear cycles we're looking at now in the AI infrastructure build and also, quite frankly, just being sensitive to investor feedback about looking at some pretty ambitious targets we've set for the company several years out. And then wondering, well, how do you see yourselves getting there? And I just view it as it's incumbent on me to paint the picture for folks and share what I see as kind of base case assumptions, as I mentioned. I mean these are what I view as the base case, not dream the dream, not putting anything in that we -- is sort of a maybe. I think it's very rational how we came up with our fiscal '27 outlook and even our fiscal '28 if you just look at it bottoms up on the custom, we should definitely be able to do that because we know those programs. I mean if I say CapEx is growing 20%, it certainly may do better, and our optics business is growing way above CapEx, so that has some legs to it.

And then look, even in a quarter, that other category in switching and storage and other has also floated up. So there's still a lot of goodness out there. But I think just from a modeling perspective, that's what you should think about. And I would say, finally, it really though is tied to our customer planning. I mean they're realizing to go get all this build to happen. They've got to provide visibility several years out. That's giving us increased confidence in our outlook because we're having to plan around that. Our R&D schedules, our capacity, our ramps, all that stuff has got to go kind of 6 to 8 quarters out now to really be ready. So that's also part of why I felt it was worth sharing this perspective for investors. And again, I think it's timed well as well with Celestial because then that thing just kicks in beyond the time frame I'm talking about fiscal '29 -- I mean some of it 20 -- in fiscal '28 but then early '29, '30 and beyond. So hopefully, that was helpful. Thank you so much for the question.

Our next question is from Blayne Curtis with Jefferies.

I want to ask you, Matt, I know it's always tough to talk about customers, but you did file an 8-K and it says you graded Amazon a warrant for 1 million shares to buy photonic fabric. So I guess, one, very simply, is that your lead customer? And if you can maybe talk about that expanding relationship, obviously, there's been a lot of back and forth your status to the customer, but this seems like a positive in terms of your engagement for the next generation?

Yes. And great job checking out the EDGAR website, Blayne, you're always one step ahead. Yes, I think it's great. So first, we did file Form 8-K talking about really an extension, if you think about it to the warrant agreement we have, which is effectively adding a new swim lane. I mean if you can believe it it was only 1 year ago that we announced with AWS a warrant and strategic arrangement with them. Back then, a year ago, it was really bucketed between AI, custom products and then networking products. And so think of this as just adding another swim lane of photonic fabric products to the mix. And all in, the potential of each of these is quite significant. So that's a positive on the first one, that's out there. Also pleased to get a very, very strong support in our press release for the acquisition from AWS that was positive. And so when you sort of look at all the things we're saying and all the data that's out there, you can sort of figure out where we're headed with this and who is helping drive this technology forward. We're very, very excited about where we can go with this technology, especially with our lead Tier 1 hyperscaler but then also the rest of the market, which I think will be shortly behind them once we can roll the full force of Marvell behind this. So thanks for catching that and giving me a chance to talk about it. It's pretty exciting to see. Really, if you think about it, we've got the warrant agreement, we got an aggressive earn out that the team is driving, which is a $2 billion number, by the way, through the end of fiscal '29. So all this adds up to just a great sort of set of incentives for everybody to go execute like crazy and bring this into production.

Our next question is from Vivek Arya with Bank of America.

Matt, I had 2 kind of questions on the data center, one on optics and one on custom. So on optics, what correlated to cloud CapEx? Why not do growth of AI accelerators, which is expected to be much faster than the 30%? And then on the customer side, you mentioned 20% as the baseline growth. Is that because the second customer is supposed to come on board? Or is it because you will grow with that first customer? And I ask about the second customer because they don't have a history of ramping, big ASIC program. So I just wanted to get your overall views on optics and [indiscernible] related to CapEx. And then on the custom side, just kind of puts and takes of the second customer that's supposed to come on board.

Yes. Thanks, Vivek. Yes. So on your first question, I mean, just the rationale was just to give just a broad proxy to the investment community about how to think about our business of sort of a very common metric, which is CapEx. I agree with you. The optics business is fundamentally driven by AI and AI acceleration. And that's why it's been growing so far above CapEx each year. And -- but I think this was just in the interest of providing some pretty broad brush strokes, but you should certainly assume that the optics piece is tied to AI, which is growing faster than CapEx. So that's a fair assumption. But just this was really in the spirit of making this a proxy. And then on the growth next year, yes, that's all still driven by -- the way to think about it is mostly our current business today. So there's a product transition with our lead customer from one generation to another in there. There's some XPU attached that's kicking in the second half that's going to lead to the '28 and '29 revenues I was talking about. But the next bigger XPU customer we have, not much of that really you should count on for next year. It's really the year after and that's captured in the doubling from fiscal '27 to '28. But I think we've got a very rational base plan for that, that we think is achievable. And then obviously, if things improve or get better, then we would up those numbers. But I think what we're trying to do right now, Vivek, has just paint a very rational, clear picture for people, and then we'll obviously update along the way more on a granular basis as we go forward. .

Our next question is from Christopher Rolland with Susquehanna.

Congrats on the results. So I think your main competitor in ASIC has moved to providing racks, not just silicon. And then with this acquisition and kind of given the increasing complexity we're seeing out there, might you be moving to systems and then perhaps even rack-level solutions as well? And do you have the capabilities to do that?

Yes. Thanks, Chris, for the question. And I would sort of answer the second part, which is, we are very much looking at this as a rack-level solution in totality and that is all the various flavors of optical interconnect. I rattled through a bunch of those. We're the one-stop shop, right, from AECs, traditional DSPs, retimers, LPOs, photonic fabrics and then scale up and scale out switching and XPU attached sockets and circuits to make all this work. And then working closely with our customers with that vision on how we enable that entire end-to-end and enable them to do that. That's our current strategy. And most of the people we're working with have that capability themselves today. They're very good at it, but we also add quite a bit of value in how to think about how to pull it all together. So we absolutely have a rack scale vision, and this is where Celestial AI really fits in. But we don't have any system-level revenues comprehended in anything I've talked to you about over the next 2 years. But certainly, from a strategic standpoint, it's imperative, right, that we go to market in a very comprehensive way, Chris, and not in a point solution way, but rather be able to provide all the fundamental pieces right from the biggest XPU chip all the way down to a retimer on the board.

This was our final question for the evening. I'd now like to hand the call back over to management for any closing comments.

Yes. Fantastic, and thanks for all the great questions, and I appreciate everybody listening. I know it was a lot. I think this was the world record for the longest prepared remarks I've ever done. But like I said at the beginning, it's [indiscernible] because there's a lot of good stuff. And there really is a lot of great things happening with Marvell. I'm just really pleased with how our team has executed. I want to say thank you to all of them. we have a phenomenal setup for next year, as I indicated, and even through the next year, we just have very good visibility. Programs are on track. We are playing offense in this company, okay? We're out doing strategic acquisitions like Celestial AI, and we're thrilled to welcome them to the Marvell team. And I think our future is very bright where we're headed. So appreciate everybody's interest. I look forward to follow-up conversations and appreciate all the investor interest in following Marvell and our journey. Thank you, everybody. Have a great day.

Thank you. This concludes today's conference. You may disconnect your lines at this time. Thank you again for your participation.

Welcome, and thank you for standing by. I would like to inform all participants that this conference call as well as any Q&A may be recorded. Where a company is presenting, any recording may also be posted on their website. Views and opinions expressed by any external speakers on this call are those of the speakers and not of JPMorgan. Parts of this conference call may be reproduced in JPMorgan Research. If you have any objections, you may disconnect at this time. Unless otherwise permitted by internal JPMorgan policy, members of JPMorgan Investment and Corporate Banking are not permitted on this call and to disconnect now. I would now like to turn the call over to your host.

Great. Thank you. Good morning, good afternoon, everyone. My name is Harlan Sur. I'm the semiconductor and semiconductor capital equipment analyst here at JPMorgan. Very pleased to be hosting Matt Murphy, Chairman and Chief Executive Officer of Marvell here with us today. We also have Ashish Saran, Senior Vice President of Investor Relations at Marvell here with us as well. Gentlemen, thank you for joining us this morning.

Thank you, Harlan.

Thanks, Harlan. Thanks for hosting. We appreciate it.

No, I thank you very much for the support. It's been -- really, it's been an incredible transformation of the business over the past 36 months, right, as you are benefiting from the strong demand pull, from accelerated compute and AI, your overall data center business is up 3x since calendar '23. It's 75% of your total revenues. Your AI franchise is up greater than 6x by our estimates over that same period of time, right? You've blown through your AI target set in April of last year for calendar '24 and on track to do the same for this year's targets as well.

So I thought that we would just sort of jump straight into the biggest question on investors' minds right now, which is the controversy around your lead customers' AI XPU program, right? Investors I think they're concerned that you're going to have a revenue hole in your custom XPU business next year because another company is claiming to have won the follow-on design to the program you're currently ramping. I understand that there are customer confidentiality issues at play here. But Matt, maybe you can just help us clarify your plans, what do you expect next year for this program? Do you expect revenue growth air pocket next year for this program?

Great. Harlan. And again, thanks for hosting. So yes, before I answer the question, maybe just to set the stage. And I think we're going to have a great discussion today. You alluded to some of the progress we made in data center and kind of the explosive growth we've seen. We're going to talk more about that and I want to cover the entire spectrum of the Marvell opportunity within data center. And then actually, I want to get to also talking about our communications and enterprise business after we get through that.

I've had the opportunity since the earnings call to meet with a number of investors, attended a number of investor events and gotten a perspective on what people are looking for, what people are thinking about or concerned about. So one of the opportunities in this call is to sort of coalesce all that and address some of those things in a coordinated way. On top of that, which is very encouraging, since we had our call, we've been able to firm up a lot of our plans now for next year. We've seen very strong demand signals, Harlan and we'll talk about that throughout the conversation today.

But very excited about where things are headed for next year, and not only that, but the firming and the schedules that we now see relative to what we think we can go do for next year is very encouraging. And so actually, we'll be talking today about what some scenarios can look like for next year. Think of it as a floor Harlan and a base case. We're still -- just had the earnings call, right? We just printed our Q2. We've got 2 more quarters to go this year, and then we're looking out into next year. But given the feedback I've gotten and also the information we now have, I think it's a good opportunity today to share how we're thinking about the business next year and then also how that goes forward. That sounds like a good plan?

No that's perfect. That's exactly what we're looking for Matt.

Ready to go, Harlan?

I'm ready to go. Ready to go.

Okay. So let's start with the first question. So yes, through -- for some time and certainly since the earnings call, we've gotten questions about our engagement and our relationship and our partnership with AWS. In December of last year, we announced a 5-year agreement with them, which covered a variety of product categories, actually, a whole bunch of things in networking that we can go off and do together, which was a portion of it. And then another portion of that was in custom AI silicon. And then also as part of this arrangement, we were leaning in together on the EDA side, and that's gone really well, and there's a whole bunch of activity over there.

So the partnership is going extremely well, and we're really happy to see the progress of our customer in the marketplace. We ramp numerous programs together successfully and into production, and we are collaborating closely with them across the board, including the custom AI silicon road map, okay? So that's all intact. That's going well. And based on the plans that we see today and our outlook for next year, to answer your question from the very beginning, just to tackle it head on, we do not see a revenue hole next year from custom AI silicon from this customer, okay? And on the contrary, and we're going to get into it as we talk about the Marvell opportunity on the custom business overall, when we look through next year and actually into '27 and '28, where you have even some new programs kicking in, the short term, medium term and long term also look very encouraging for our custom silicon business, which, by the way, is still in its early innings, and we're going to talk about that later because I do want to provide some context on where we were and where we're going.

But relative to what people care about right now, which is where we're going. Next year, like I said, we expect growth, and we don't expect a hole. And I think the setup is great. So I hope that helps address some concerns. Obviously, we have customer confidentiality to think about and the level of detail, as I've said a number of times, we can go into is somewhat limited. But I think given the dialogue we're having and where things are looking, this is where we stand today. Now let's -- I want to go through the kind of overall view on data center for a moment, and then we can circle back on a few questions, okay? So the first is very encouraged by the demand signals we're seeing. You can see this reflected in the overall market with big kind of episodic announcements on all kinds of investments, programs, et cetera. So -- and of course, or -- so as a result, the expectation for growth again next year in terms of CapEx is -- looks very solid.

Now we don't know exactly where that's going to land. But right now -- and I'm going to use this today, Harlan, just as a baseline concept. But basically, CapEx has grown a lot this year. All the projections we're looking at right now have a peg at about 18% for next year. I think there's probably upward bias there. But for the purposes of this discussion today, when we look at the overall Marvell opportunity set, why don't we just go and index it to 18%. And then if it does better, great. But that's a way to help contextualize and think about our business. And again, I would -- given what we see, I want to just remind everybody here, we're -- what we're talking about today is what we really see as a floor and kind of a baseline for growth I think we have opportunities to overachieve and we'll talk about each of those. But just for right now, I think that's a -- just use that as a base assumption, and then we'll update as we go along. Okay so...

That's fair.

Yes. Perfect. So let's talk about data center. So this business overall, we've given some ratios in the past, and they're still largely consistent, which is if you bucket it half the business this year is in our optics franchise, which there's actually a number of subcategories, which we're going to talk about today with different growth drivers and opportunities. And I would just say for a moment, this has been completely ignored by investors, quite frankly. The overwhelming discussion points and all the oxygen has been out of the room talking about really one socket, on one custom piece of business. And so that's why I think it was good to hit that one upfront. And we can circle back if you have a follow-up. But basically, let's -- I want to have the real conversation today, Harlan, about our business.

So half the business is in optics. 25% is in custom, which has come up a lot as we've ramped these programs. And then the other 25% is actually -- I mean, it's got its own set of growth opportunities, by the way, but that includes our storage franchise, our emerging cloud switching product line and then also our security business and then just a variety of processors, management switches, things like that, that are just general purpose used in data center. So those are 3 buckets. 50% optics, 25% custom and then 25% in this emerging category.

So let's talk about custom first. So '26 has been a strong year. If you take the 25%, you apply it to what Street things we're going to do this year, you just get big round numbers, you get about $1.5 billion. And that's effectively doubled from a year ago, okay? So in 1 year, we basically have doubled our custom business. And when we look at the next year, and again, this is for the purposes of just setting a baseline and indexing, we see the custom business growing in line with CapEx. It's just a base case scenario, okay, for next year.

Yes.

Due to the nature of this business, which is today concentrated with a few sockets, more are coming, and we're going to talk about that it's -- that's probably the best bet we can use, obviously, as we get closer, if things are biasing up, we can go adjust that. But that's a safe assumption. And within our growth for next year, we do have XPU attach kicking in. That's probably some of these new sockets are later in the year. So there'll be -- that part is actually growing quite well, but it's smaller as a percent of the total. So it will help on the growth. But the assumptions we're using right now and the model I'm giving you assumes the current projects that are running continue to the extent there's any generational changes on those products from one generation to another. That's comprehended as well, any product transitions.

And any new major programs on the XPU side, may have some small contribution, but you should think of those as we said at the AI Day, Harlan, is a '27, '28 type of inflection. Okay. So that's -- that's the custom. Now I'm just going to talk about XPU attach for a moment because I would say on my investor meetings, once we got through kind of the big story, so a lot of interest in this in unpacking XPU attach in more detail. So let me give you some thoughts on it. The first is the way I would bucket that one is today, there's really kind of 3 major categories we're looking at. One is customized NICs or SmartNICs, so that could be foundational or that could actually be with a lot of CPU cores and extra offload special capability. Memory expansion is another one for disaggregated memory and this was kind of on the basis of a lot of our early CXL investments, which have now created kind of a platform for memory expansion that we can use in our custom platform. And then accelerators and security coprocessors where you want to basically move some of the processing capability that would be inside an XPU save that valuable real estate for GPU cycles and move those functions on an offload type of chips. So think of those as kind of the 3 big buckets.

And I would say even since June, when we outlined what we thought that business could do, which was about a $15 billion TAM with a huge quantity of opportunities, which ranged in the kind of a couple of hundred million dollar lifetime revenue range. I would say even since that Investor Day, when we look at the business we're closing or the -- actually the products in development as we look to the long-term forecast, I would just say that the lifetime revenue numbers we gave were probably low. They're definitely creeping up to where -- that's -- we said in a couple of hundred million range. But I mean some of these probably are $1 billion lifetime, Harlan. So the average has probably come up, but some of these can end up being quite a bit larger.

And I think the reason that this is happening is because as the XPUs are now going into production and the AI spend is getting so great, there's just a greater need in the economic support doing more and more customization. And you go back to 2021, we talked about this, Harlan, called cloud-optimized silicon, but basically the idea that every cloud is unique. Everybody is going to want their own optimized TCO network and optimize TCO for compute. And that's really what this provides. This business is attractive to us because it leverages a lot of the Marvell IP we already have. And we can do a lot of the design work actually for the customer.

And so that makes it asset light on their side, and they can get what they want and then we tend to get paid a little bit better margin on these because they're not as big as an XPU as an example in lifetime revenues, but they contain Marvell IP. And they also tend to -- they're not tied to the XPU platform. So we have programs right now with design wins where it's somebody else's XPU. We're not the provider of that. Somebody else is doing that or it's a merchant platform. We can attach to those too. So that actually, I think, over time, I'm not calling the year, but I think if you look at it, that 20% target we set as a long-term target for custom market share, I think the XPU portion over time could probably be better than that just given the traction we're seeing. Let me move on to the optics business. Let me just walk you through data center first, and then we'll take a pause, okay?

Yes.

So on optics, so this business has done phenomenal. I mean we have blown past any kind of CapEx growth rates for the last 4, 5 years in this business. And I don't see this stopping anytime soon. And so if you look at next year, and you look at the kind of 18% kind of baseline number, optics is clearly going to do better than that, clearly going to do better than that. 800 gig is very, very strong for next year. We were early and we were in the lead position there. 1.6T is now ramping. We're shipping it now. That's going to also layer in next year, but not at the expense of 800 gig. And then on top of that, within our optics, which we haven't talked about as much, but the -- because the traditional optics we have is really for scale-out, Harlan. And as scale up is something we haven't really talked about, but I'm going to cover it a little bit in this call because in scale up, actually, we have retimers today, Ethernet retimers. We have AECs that are ramping.

We have PCIe products that are coming, that are sampling. So that's actually a sub-product category that I'm going to talk about it a little bit today, but that business is now getting to be important, and it's going to actually be significant revenue growth for us next year, the scale up portion. This doesn't include any switching at this point. This is just retime products and AECs. So we'll dive into that more later. I'm going to finish this out. So that's 3/4 of it, right? Custom, you got optics. And then the remaining business in data center is in -- of the 25% left. Switching is really where the action is, that, and we'll go over that later. But basically, we acquired a company called Innovium back in 2021 to get us into this business. At that time, it had very de minimis revenue. I mean it was doing like $5 million, $10 million a quarter, but it was going to do $150 million, and that's kind of what the deal model was based on. And we already hit that.

We actually achieved that very quickly. I mean we've actually -- I was looking back and it feels like forever ago, but really, we had basically just a few years, right, it's really how long we've had this business in terms of when it's actually ramped and going. So this year, it's already double, even over double where it was -- where we were on that baseline $150 million. So it's already grown a bunch. It's going to grow again next year as we continue to ship 12.8T, but then 51.2T goes into production. And I'm not willing to call it just yet, but we see very strong growth forecast for our switching business next year. Well above CapEx, obviously, you can just imagine the growth rate we're talking about here. I mean maybe this thing could be triple at some point where it was on that baseline assumption as we head forward. And then within that, you still got storage recovering, which is great.

By the way, that was a wall of worry a few years ago. That thing is back up to that $200 million a quarter. It's come back and probably goes from there. And then you've got the other products and they're like merchant processors and the security stuff. So just to be super conservative on that one, you got the switch growing and the other stuff may be a little slower growth. Just assume that, that one goes double digits, right? As again when you walk through the buckets. So you got first one at basically baseline of CapEx, optics above, this one put it at double digits. And again, just as a floor, just as a model to work with, you get very healthy growth if you do this. You end up above CapEx relative to where things are at the moment with really, I think, a lot of upside potential. But I wouldn't update the model just yet Harland on that. Let's go get some time in front of us. But I'm not saying to update the model because, we don't think we can do it. I'm saying wait till we actually can start really getting closer and we can see the opportunity in front of us. So that's on the data center.

Let me just finish on comms real quick and enterprise because that just finishes the whole picture of the year. And then we can dive in. On comms and enterprise or what we're calling communications and other now, we bucketed consumer in there, industrial, enterprise and communications. That whole business, like comms and enterprise, we had guided up like 30% sequentially. That was a question I got from investors. Is that a onetime thing? Is it going to go down after that? Did you guys just kind of overship or is there something weird going on? No, sure. It's recovered and it's inflected, which is great. And that's on the back of new products actually ramping plus the inventory recovery, but that business will continue to grow. And I think it's safe number to pencil in for the whole thing comes and other with like just -- and assume consumer stays at $300 million a year, assume industrial stays at $100 million within there, just kind of keep those flat, the overall should baseline grow like double digits next year.

So everything is growing next year, custom, optics, storage, switching, other, comms, enterprise, the whole thing. And that's as a floor. That's the baseline. All right. Take a pause for a second. You can fire off another question. Thanks, everybody, for listening for the long opening, but I think this was an important way for me to let investors know what was in my head and how we're thinking about the business.

No. I think that was a great overview, very articulate because as you know, that's been one of the concerns by the market is how do we think about the shape of Marvell's growth profile for next year. So I think that was hugely insightful. And like you said, I think part of the reason for you being here today is maybe some of the incremental confidence and visibility on when a lot of these programs are going to fire. So I'm just going to kick it off with a number of questions. And we'll get into more details on a lot of the segments that you talked about.

But let's just go back to the initial controversy and misconceptions around the business. And so based on your comments about not expecting any hole in revenue from your lead XPU program next year strong confidence in growing custom revenues in the 18% high teens range as a floor next year. Also, most of that coming from your existing XPU socket wins, I think that the only reasonable interpretation conclusion is that you have to be participating in your lead customers next-generation program which has always been my assumption and a topic that we've written on extensively. I understand that you can't comment specifically, but I think you've kind of laid out the breadcrumbs very clearly, which investors and I should appreciate. Is that fair?

Well, Harlan, you're a smart analyst. So I think the way to think about it is, like I said, we have a very good relationship with our lead customer. I'm really pleased on the outlook that we see for next year overall data center and then within custom, as I mentioned, they've got a very safe assumption to see that strong growth continue from this year to next year. And like I said, I think short term, medium term, long term, our custom business is in good shape.

That's great. And then at the June Custom AI event, the team talked about this notion of concurrent design programs, right? In other words, at the same time that the team is at the tail end of your customer's 3-nanometer design, you're already working with your customers on their next-generation 2-nanometer designs. I specifically asked Will Chu. He's the head of your custom cloud programs, right? Is the Marvell team already working on customers' next-generation 2-nanometer designs. He said, oh, absolutely, right? So my question to you is has anything changed from the June event as it relates to 2-nanometer engagements? Or is the Marvell team still very much engaged on follow-on 2-nanometer programs with all of your major customers?

Yes. Thanks. No, I think Will said it right when he said absolutely. I would put an exclamation point on that. Things have only heated up, quite frankly, since we had our June event. We are heads down on multiple programs in terms of driving execution and tape outs into next year. I don't even -- at some point, it's not even going to be a 2-nanometer discussion, honestly, Harlan, we're already moving on to the next, okay? So we're going to see multiple products across 2-nanometer across the various franchises, absolutely heads down. And again, the concept of multi-generational having visibility to the road map, figuring out where we can intersect, where we can add value, how the customers are thinking about their road map, which is -- I've said this a few times, I think there's going to be lots of different opportunities. You've got XPU, you've got XPU attach and even within XPU, I think you're going to see multiple opportunities and some bifurcation occur.

So the design activity right now is off the charts, but I don't think it's even a 2-nanometer discussion anymore. At some point next year, I think the discussion is going to be around A14 and A16 and how do you think about that? And how does the technology evolve? And we're still early there, but I think we'll have some exciting updates next year around where we see the technology going as kind of a follow-on to what we did at the June event. And I mean, we just announced last month, for example, we did a press release on our 2-nanometer 64-gig die-to-die IP, which is being very well received for these different XPU architectures. So anyway, you'll see it in the press and -- but behind the scenes, the team is driving strong execution in this area.

Great. I mean let's shift from the sort of focus on customers and near-term programs, like let's just shift and you brought this up in your commentary part to this. But let's just shift to the overall pipeline, right, the custom XPU attach pipeline, right? And as you rightly pointed out, at the June AI investor event, the team outlined 18 multigenerational custom XPU and custom XPU attach wins, right? And you mentioned on the recent earnings call, additional design wins, the same thing that you mentioned today for custom taking your sockets to something greater than 18, right? I mean can you just -- any more color that you can give in terms of the incremental revenue that these design wins are expected to drive? And more importantly, sort of the time frame which these 18 programs will start to contribute more incrementally, more significantly to revenue profile of your AI business as we move forward in time.

Yes. Well, we've been on fire since the June event. I mean I did a review last week, we had a Board meeting and our sales leader was presenting, Dean, the whole team's kind of outlook for the year. I mean, we're blowing past some of our targets for the whole year right now on some of these different product opportunities, okay? So the number is not 18 anymore. It's 20 plus that we've now closed, locked, SOWed, staffing, et cetera. That's since June. And just to contextualize it a little bit, obviously, we're early here, and -- but I'll give you a framework. This is multiple design wins, right? Obviously, they go from 18 to 20 plus. And it's one incremental XPU opportunity at an emerging hyperscaler. It's multiple XPU attach sockets that we were chasing and working on, have been in the pipeline. Those are now closed. And the way to think about it is, which I'm really encouraged by is, we put out a $75 billion funnel at the AI Day that we were tracking in terms of opportunities, which look, just enormous.

It was almost hard to even put the number on the screen because that's like real stuff in our Salesforce system and database. These are like real sockets we're tracking like 50 of them. Just to give you a sense, we've now knocked off like 10% of that in basically 2, 3 months. So now we're going to refill it because I think we're already finding new opportunities, but take 75x 10%, and that's the lifetime value baseline estimate we have for those wins. So that's really encouraging. Now some of those will probably hit, we could call it '28. Let's just -- but I think this is -- think of this as '28 and beyond. So it would be not to raise our '28 kind of goal, but maybe to give incremental confidence that we can get there because programs, you have a baseline, and it's some are going to be higher, some are going to be lower. So I think the worse the better, okay?

But these are very solid programs because most of them are follow-ons to things that were already being done, like in a merchant fashion as an example. So these are all very high-quality wins that will contribute in '28 and beyond. And obviously, if we can go faster, then these would contribute sooner. And I think one of the challenges in this market is that once you win something, then it's all hands on deck and can we have it yesterday. But just for our planning purposes, assume these new wins are just adding into the '28 and beyond time frame.

And Harlan, in terms of how some investors may want to think about how to put that in kind of an annual kind of bucket, right? So think about that 10% as $75 billion. On average, like we said, these XPU sockets tend to be only a few years versus the XPU attaches on average, let's call it 4 years, right? So if you just want to use roughly 4 years and just take that number divided that's kind of the incremental revenue on an annualized basis, which is obviously very, very meaningful for us. So that's kind of the other thing I wanted to point out.

No, that's great. And I've got a couple of questions from investors here. And I think you did a great job of outlining a baseline sort of floor/baseline view the data center business in calendar '26, right? And I interpret that as somewhere at the baseline floor and sort of that high teens, the plus 20% growth in the data center business looking into calendar '26 kind of rough numbers, right? We all know that you have a goal of 20% share of the total data center market by calendar '28, which would imply off of the 26% of baseline that you gave us that Marvell team should see acceleration in their data center business in calendar '27 and calendar '28. With the pipeline that you see today and the update that you gave us today your confidence level on acceleration of data center in calendar '27 and calendar '28?

Well, yes, it's a great question from probably a great investor because you guys just stole my punchline at the end. But basically, yes, the way to think about it is, we've got this outlook for '26, but it only accelerates, Harlan, in '27 and '28. We even said this at the June AI Day, but I feel even more incrementally confident in that. We have additional custom programs, new ones kicking in, in that time frame on XPU. We have a number of the XPU attaches starting to layer in. We start to get a more meaningful contribution from 1.6T optics. We start to see 800-gig ZR taking off. I mean I can go through the switching thing keeps going. The scale-up keeps going. I mean, we're going to actually talk about this, I think. I think we should spend some time on it. But the bottom line answer is, yes. The growth -- our view right now will accelerate in '27 and '28 above what we're looking at for next year. I mean, this year, think about it, we did basically based on Street, we'll do about 40%, right. Actually, a little bit more, right, Ashish?

Yes, 40% would be -- exactly.

And then if you look at the TAM and the opportunity set, so yes, that's the current thinking.

Okay. Perfect. Let's move over to the electro-optics business, right? You talked about it at the beginning of our conversation. And you've said it over the past number of earnings season. But to your point, I'm not sure that the market truly understands the growth and the scale of your optics business, right? We used to cover Inphi. When you guys acquired Inphi, I think they were driving $800 million a year in electro-optics. This year, 50% of your data center business, that's a $3 billion optical franchise. So just huge growth in your optical business. And obviously, the Marvell team continues to remain the leader in that space, right?

And so with that in mind, like the team kind of near term, you guided your optical business to grow double digits sequentially here in fiscal Q3. It seems like the market isn't, like I said, valuing this business appropriately. Do you see this continuing to grow for Marvell in the long term? I know you gave us your view on optical for next year. But how do you think about the growth in optical more longer term? And what are you seeing on the competitive side as investors are concerned about potential share loss on your core DSP business, right?

And to put things into context, it feels like from 200 gig to 400 gig to 800 gig to 1.6T, like every transition, there always seems to be a lot of noise on the competition. From my perspective, Marvell always ends up coming out the predominant share leader in this space. But what are you also seeing on the competitive side as we move to the 1.6T generation?

Yes. That's great. No, I'm actually glad I have the floor here to talk about this. I'm going to get a little fired up here, Harlan, okay? So we acquired Inphi. We announced it in 2020. We closed it in 2021. We closed it very quickly, executed really well on the transaction. The whole company had done like 680, the 2020 year and Street had them at like 820 or something, 830 for -- that was the whole Holdco, which included some other things. By the way, that was a very high multiple we paid back then, and it was unclear. I mean, investors were concerned, what's the accretion going to be on this, and it was certainly thin at the headline value, but we saw the strategic value, and I think the market recognized that pretty quickly.

This business has absolutely ripped since we bought it, okay? The DSP franchise alone has basically 5x-ed since that time frame. So it was actually about $600 million, just to give you the number that was in the DSP area. That's 5x. So that's been growing at 50% CAGR per year since we bought it. And it was already at a decent size, by the way. This wasn't going from $50 million a year to $50 million a year or $100 million a year. It was already like had ramp, but it had more legs in front of it. Absolutely, our team has kicked a** in this area, okay? And when we're putting up record numbers like right now, I would say that -- and we can talk about it later. It's not really worth rehashing the past, but you've actually -- I appreciate the measured tone you and the JPM team have taken because and we'll talk about competition after I give you a couple of thoughts first. But I don't know how many times there's been a scare, there's been a concern. There's been a market rumor, something that's going on.

And it looks like the world is coming to an end for us. Not the case. Not the case. On the transitions, like you said, we absolutely crushed it on the 100 gig per lane, 200 gig per lane transition, enabling 1.6T. We're out in the lead on 400 gig. I'm very, very proud of this team, okay, and what they've been able to accomplish. And now we've gotten this thing, as you mentioned, like a $3 billion franchise. And so if you think that -- if anybody thinks that all of a sudden, I know that was great, but now it's just going to all go to hell. Not true. I mean even if you look at the TAM growth we outlined, right? Just the TAM alone at the AI Day was like 25% to 35%. Let's just call it, 30%, right?

So if you just take the growth and you just assume we can grow at the market, I would put this business up against any other benchmark out there in terms of interconnect franchises. It's healthy from a margin standpoint, very healthy. In fact, we have capacity to continue to invest here aggressively because of the scale, and we're doing that. We are staffing new programs. We are staffing new opportunities. We've got an absolutely, I think, best-in-class road map. So this business has done absolutely phenomenal. Within that as well, so you got the DSP franchise, you also have DCI, right? And DCI was always at Inphi about a $100 million business. That's how people thought about it. They invented the category. they fundamentally brought to market the first pluggable long-haul optics with 100 gig ZR, led the transition to 400. We were first to market announcing 800.

And by the way, that's still all in front of us. And that business has grown at the same rate, by the way. That thing is also like 5x-ed, and going to keep going, by the way. So that's also been a real shining star for us. And so what drives this business ultimately, you can't sit around. I mean you have to be in the lead. And it's not just nanometers, it's architecture, it's the BVA circuits around it. It's the module partnerships. It's a robust supply chain. It's the flexibility. It's the unique feature sets that we're able to incorporate into the chips that no one else knows about because we're first. So when people come to follow, sometimes they don't even know what they don't know. So we're just focused in this team, head down, execute, be the big winner.

And as I look at beyond 1.6T and we're going to 3.2T at some point, PAM still has legs. We've demonstrated the technology. We're deep in development there. The other thing that we pioneered was this concept of coherent light, which DCI uses coherent, inside data center uses PAM. Coherent light was a solution to do like 20-kilometer kind of ranges like for campus environments. We pioneered that. Standard in that product line, that is ramping into production. That's going to be a whole market segment for us. And that coherent light investment will also benefit us back long term at some point when PAM does run out of legs. We have a great solution there in terms of the capability we have.

So honestly, I'll just give you the punchline. This business should be worth more than all of Marvell right now. If you just take like a SoCs multiple on this thing. Don't even put the sky high multiples on this. Just put a premium multiple, SoCs plus, it's an incredibly valuable asset. So we're going to drive value in this company with all the tools we have, but do not count out this optics business and do not undervalue it. And I just want to give a major shout out to the team that's been involved in driving this incredible performance, the Inphi folks, the Marvell folks, everybody involved, it's really been a stunning performance and great job and the future is bright. I'll pause there.

Yes, just to just remind everybody, I mean, just to be clear, this is really still the PAM business, which he talked about is essentially of scale out. So our revenue today really is vast majority scale out and DCI, right? There's a tiny bit of scale up, which we'll get to next, but just to say that the scale-up thing is entirely in front of us. As you know, most of that's on copper on the interconnect side. So that's another big opportunity and also on the switching side. So I know you're going to get to that, but I want to make sure people understand the context that -- this core business on its own is going to keep growing to the rates which Matt talked about, but there's an entirely new growth paradigm, which is going to come from scale up, and I'm sure we can get to that in a few minutes.

Yes, exactly. And let's pivot...

And maybe Harlan, but I want to get to scale up next because that's going to layer into the growth story.

Yes, exactly. And let's pivot over to that because I do think that the Marvell -- core Marvell team's historical leadership in networking and then you overlay that with the optical capabilities that the Inphi team brought to the Marvell team, I think it sets you really, really well for this incremental opportunity that the market is starting to figure out, which is this scale-up networking connectivity architecture, right? Scale up or what we call within rack connectivity this opportunity driven by the continued deployment or expansion of deployment of rack-scale based AI systems, right? And the Marvell team offers a wide range of solutions here, your switching architecture, AOCs, AECs, retimers and more. Maybe, Matt, walk us through the opportunity for Marvell, highlight some of the new products you're developing to capture this growth. But additionally, like how does the opportunity evolve across the different technology protocols for scale-up such as Ethernet, UALink and PCIe Express.

Yes. No, happy to cover all that. So yes, so a couple of things. So one is the scale-up thing is totally in front of us. And we have participated so far not at the scale we're going to, okay? And one of the reasons for that is that the current generation, both on -- and I'll just give 2 examples, PCIe in terms of retiming and then on the AEC side, both of those first sort of generation products that are now being deployed, they were based on an NRZ modulation, okay? So this was -- these have not yet made the transition to PAM, right? So our strategy has always been when that transition happens across the board, retiming, AECs, all these different product categories you mentioned, that would be where Marvell could really, I think, add value.

Now we built a business already. I mean when we acquired Inphi, Marvell had some retimer products for Ethernet. They had some small business, $30 million a year or something. We've got AECs ramping now this year even. And so revenues are going to be approaching $100 million this year. Maybe we don't know. We can see where Q4 lands, but it's -- we're growing into it and probably that easily doubles in the near term. And then over time, this is going to be off to the races, okay, from an AEC perspective, retimer perspective. And it's perfect because we are announcing multiple different products based on our PAM technology, leveraging our latest -- and so we've gone from optical to electrical here, but we have excellent SerDes, okay, excellent SerDes in both of these and leadership position here. So just -- so that's one bucket, right?

It's just scale up, we're going to compete. We've got revenue. We've got a bunch of opportunities and design wins. We're engaged across the board. I think that's going to be a broad-based business where we can compete in that area. Scale-up switch is, I think -- I don't even know how to size the TAM yet, quite frankly. It's -- and we have to see how everything plays out. But scale up switch, I just want to be very, very clear on this call, we are absolutely investing here. We are engaged with multiple customers and multiple different platforms at the moment. Some of these will include opportunities we also have with XPUs, but some of these will be independent of that.

And the reason that this is, I think, a perfect sweet spot for Marvell is you actually have to have switching IP. That is going to be one of the key differentiating things. And what we're finding from an ecosystem perspective, right, whether it's people that -- let's just pick people in the photonics world, right, that are saying, "Hey, we're going to drive our own CPO road map. We need a switch partner, that's one aspect." The other is the hyperscalers themselves, right? We're trying to decide on our architecture. And it might be UAL. We're investing there. We're on the -- we're involved in that standards body. It could be Ethernet as well. It could be some combination. This is a lot in flux here, Harlan. But the first point is you have to actually know how to design a switch.

And the Innovium IP, which is best-in-class in terms of low latency, low power, we can scale that to a very effective scale-up switch. We obviously have the connectivity that sits around it. But the key on the connectivity is you have to have the SerDes as well. Homegrown SerDes, not third party, not third-party IP for switching, not third-party IP. We think controlling the whole stack is really, really important. And then obviously, having kind of that broad engagement now with 20-plus XPU and XPU attach sockets, all the connectivity we're doing, we are very much in the mix on -- and these are not small products. These are large, high ASP, high complexity chips. So you have to be also on the leading process node to do these.

So it's not for the faint of heart kind of derivative product. I'm just going to slap a team together, go make this happen. And this is going to be fundamental to how rack-scale architecture ultimately is defined. And we can have shots on goal, like I said, to connect our connectivity around it. We have our own photonics effort. We can actually work with partners, too. I mean we're very flexible in how we approach this whole ecosystem. We're trying to enable our customers to have a lot of choice in this area, and we're going to be pretty flexible and nimble on how we go to market. So there's more to come on this one, Harlan, but there's been some question, are you guys even going to be in this business? Or what are you guys doing? And we are heads down in opportunities right now. Makes sense?

That makes a lot of sense. And that's a good segue into maybe diving a little bit deeper into your switching franchise, right? Because as you mentioned, you bought Innovium, quite a bit of success with the first generation 12.8T switching platform, you're now ramping your 51.2T what they call Teralynx 10 switching platform, which is comparable to Broadcom's Tomahawk 5 switching chip. Maybe just give us an update on current revenue levels. I think you did throw out some revenue levels, but could you just reiterate that customer design wins. We know that you do have a primary customer, but customer design wins beyond the primary customer. And maybe share just any insights on the team's product cadence and road map over the next few years.

Yes. Yes. No, we're -- we successfully integrated Innovium. The product that we inherited, if you will, or that they had designed was at 12.8T. That was Teralynx 7. We were able to get that into high-volume manufacturing and production. It was really a joint effort with the Marvell team and Innovium team. I think that was a combo to actually get it to where we could meet the thermal requirements, the power requirements, manufacturability scale, got that done, executed 51.2T first pass success. SerDes performance is excellent. And then we're now working on our 100T product. And then there's actually an interesting road map beyond that, which we have not announced. But we are definitely considering -- we are definitely following the road map and the cadence of the industry, and we're following our customers here.

Nothing yet to announce, but I would just say that broadly speaking, there's just tremendous interest now that we've proven ourselves with our own product in 51.2. And just given the sheer market size and kind of how some of these different opportunities are now bifurcating, I think there's going to be a lot of opportunity for Marvell here in the switching area. And I said it earlier, I've given a lot of information, but I'm just trying to give some more context for people to understand how much progress we've made. So yes, when we bought it, the deal model said $150 million was kind of the revenue level we would get to once we sort of got the first products into production.

We're running at twice that now. And like I said, in the next as 51.2T ramps, you could sort of think about it being a 3x on that plus. So start to think about a business that gets to like $0.5 billion sometime in the next year plus. And then maybe I don't know where it goes from there, right? It depends on the traction with existing and future customers and then also how the 100T layers in eventually. So yes, very excited on how that's done. I mean it's only been a couple of years, and I think this is one where we just need to be patient, but it's great to see the results so far.

No. And I agree with you. And I think as you rightly pointed out, this is a strong differentiation and springboard which to then enter the scale-up segment of the market, which is now just emerging. Let's pivot over to because you...

Everyone knows our technology on scale up and switching. Every single customer. They know the architecture team. They know our people. And so it's a logical extension on something that's newer where we're not -- there's not a -- outside the one big merchant guy, there's not a strong incumbent yet and the standards are being developed. So it's actually a really good sweet spot for us to enter.

That's -- you brought up as you were talking about how you think about growth in your data center business next year, did break out the other part of the data center business, which included switching, but it included storage. And I feel like a lot of times, the market sort of misses the fact that Marvell for a long period of time has been a powerhouse in storage controller silicon capabilities, right? And the good news about storage and storage controllers is that we are seeing a lot of strong demand pull for high-capacity HDD. We're seeing the strong demand pull for enterprise SSD, right? The market has been looking for what is that correlation between AI demand and workloads and the impact that it has on the high-capacity HDD and enterprise SSD segments of market. That's now finally starting to emerge, right?

And so the good news is Marvell team, as most of us that have followed Marvell for a long time, knows that you have a very strong position in both nearline HDD controllers and enterprise SSDs like Seagate, Micron, Kioxia, right, all customers. There's a strong demand pull, like I said, these solutions for AI now. Storage was at $1,800 million a year business for Marvell, as you pointed out. What's the revenue run rate today? I think you said we're getting pretty close there. And can you discuss your design win pipeline? And is there more growth beyond this $800 million floor as you move forward in time?

Yes, great question. And I'm having a back to the future moment here. I feel like I'm back in 2016 when I became CEO when Marvell was only a storage company. And our goal was actually to diversify away from that. We never said it was a bad business. We just said it was really kind of oversized in terms of our scale. We still are very active in this area. The business has recovered. It was at about a $200 million a quarter run rate kind of as a baseline a few years back in data center, overshot like crazy during the pandemic. I mean we had some kind of bonkers quarters in there and then really inflected hard down and everyone say, oh my God, we're going to come back. And it's basically there now. I can't remember Ashish now Q2 or Q3...

It's about there. It's about there.

It's like we're back. And I think I agree with your assessment that there absolutely are these growth pockets now that we're going to see into '26 coming from AI. I would just say that in our business, our storage, not a small portion is actually in fiber channel. So that's enterprise on-prem. And we've kind of told investors even going back to last year, like just kind of keep that guy flat or maybe it's down a little, it's not a grower, okay? So even if we see a little of these green shoots within storage, Harlan, that's why the growth rate of that probably is not going to be as strong as you would just sort of mentally model because you got a fiber channel piece. But we see what you're saying for sure. We're seeing upsides on certain parts of our storage business. So that will grow a little bit next year, but the real driver of that sort of third category of data center is really the growth in the switching area.

Yes. I think, Harlan, I mean, it is a bit of a call option, I think, on storage from AI. So far, AI hasn't really -- because it's been mostly training, but I think as you move into inferencing and more complex referencing and you start to store training data sets, I think there could certainly be a call option. That's obviously not what we figured into what we talked about. But that certainly, I think, is something to be very much aware of. And you're hearing -- starting to hear that. So I think that is something we should all keep an eye on it, and it clearly could be an upside scenario.

And as we near the top of the hour, and I really appreciate the update from the team here, I have to end it with this, which was it was great to see the $1 billion ASR that you announced earlier today in addition to the $300 million that you've already bought back in this current quarter, right? You also announced the new larger $5 billion stock buyback authorization. You're driving, I don't know, close to like $400 million of free cash flow per quarter. So how should we think about -- when we think about Marvell, how should we think about the mix between buybacks and M&A going forward?

Yes. Yes. So set aside the ASR for a moment, you're right. I mean we have been driving very strong free cash flow in this company. We had an investment mode a couple of years back, told everybody we got to go put some capacity stuff in place. We were doing all kinds of expansion. We're reaping the benefits of that now. So like you said, we've already done $300 million this quarter, right? And now we did the ASR. Look, we got a lot of flexibility, right, when we did the automotive divestiture. Closed it early, team executed well, ended up with a lot of cash, a lot earlier than we thought. So we had a Board meeting last week. We discussed it. We looked at kind of where we thought Marvell was. They're very aware. A lot of them have been on the Board a long time. They've seen the journey. Very confident. Me, the management team and the Board in the outlook.

So the $1 billion was really to signify that to investors and also to return capital in a very clear way to investors in the form of the ASR. And that was kind of -- that was a big -- it's like half the proceeds or something, but it was -- I appreciate everybody letting me invest in that business because nobody wanted to invest in that business 9 years ago, and we turned it into a $2.5 billion asset. So that's positive. So when you look at the $1 billion, you look at the $300 million and kind of what we were going to go do for the next few quarters, we were going to run out on the $3 billion authorization that we had. By the way, we did the $3 billion authorization, Harlan, in March '24. We've blown through it, okay, like -- which I mean, that was a big number for Marvell, right? So we've achieved that very quickly.

And we discussed it, and we just -- the Board approved on my recommendation, the increase to $5 billion in terms of the open off. So that's out there. And so -- but even after the ASR, like you kind of referred to it upfront of M&A, just I want to close on this and then give a few closing comments. The first is we still have flexibility with the cash balance we've got, with the free cash flow that we're going to be generating, we can still be opportunistic. And I said this when I was on the road with investors, look, we're at a historic moment, and there may be properties or there may be teams or there may be things that we want to go do to really set us up for the next wave of growth.

We've been pretty much all organic, Harlan, for like the last 4 years. We just did it on our own. We invested in ourselves. That's gone great. But I want to stay opportunistic. And I think as I was preparing for this, I was getting a little fired up, like look at this Inphi thing. It was doing -- it was $600 million to $3 billion, right? The overall -- so the overall data center has grown, but it's on the back of the M&A. So Inphi has crushed it. Think about Avera, okay? You just baseline like what was the data center business in custom back then when we bought it, it's like $100 million a year. It's 15x-ed okay? And like 4 years, 15x-ed. And again, I know everyone is worried about the future and what's going to happen. But this is like just ramped. I mean it doubled year-over-year. It's going to grow next year. So that's really positive.

And then Innovium, I gave you some of those numbers just to contextualize it. So I really -- when you do the right deal into the right market and you got a team like Marvell, which is like a great place to work. We know how to integrate. We know how to execute, how to bring everybody together in one solution. I think it's powerful. So I was reflecting on all this when I was like looking at the data and got my own spreadsheet going, like this has been an absolute home run, and I appreciate investor support over the years to go do that. So -- but look, that's all potential. Right now, we're returning capital to shareholders. We've got a big growth opportunity in front of us. And let me -- can I just close now, Harlan? I just got a few more comments.

Absolutely. Absolutely. Absolutely.

So final thing, I would say to everybody, I told you I was going to be fired up. Didn't I Harlan?

You did. And you are.

I am because we had a Board meeting, as I mentioned, last week. And one of the slides I do for the Board in my private session is I've just shown some peer benchmarking. I calibrate everybody on the big picture. I've had the same set of slides for like 8 years. And the first slide always shows the benchmarking, Marvell's performance year-over-year, so it would be like Q2. And then we do that vis-a-vis all the large semiconductor companies or just to give the Board a landscape. And every -- in the last few years, we're kind of moving up and our growth has been really good. In Q2, our Q2, on that chart, the fastest-growing semiconductor company of large semiconductor companies in terms of percent growth was Marvell. It's the first time it's ever showed up on the chart. It's the first one, and we're all staring at it like, so is it really the end of the world? I think it's just the beginning. It's not going to all fall apart.

All these rumors, all these market noise over the years. I'm talking over the years, right? I'm not picking on anybody in particular, but just stocks up, stocks down. I'm just saying if I just look at the last quarter we printed and how did we do? We're doing pretty well. And like you said at the beginning, we 4x the data center business in 4 years. And we got a huge market tailwind in front of us in terms of the TAM. We have these product cycles kicking in. We have inflections coming in '27 and '28. And I think the future is very bright. I don't think -- like I said, it's not the end. It's the beginning. I'm in my 10th year at this company, man. We took 10 years to get here. Now it's time to reap the rewards.

So my final comment I want to thank -- I want to say something. I want to thank my team. All the Marvell people that are listening or if you're going to read the transcript, I want you to listen to me right now. This is a group that knows how to win. These people are tireless, they're fearless. They've been relentless. We have been the underdog, the scrapper. We're humble, we're hungry, okay? We never take anything for granted. And this is a team here that has no excuses. They just get it done. And I appreciate everybody's patience in the Marvell employee base for riding the waves and the things that go on that aren't in their control. Our engineers don't know what's going on and why the stock is up one day and why it's down another day.

But what they should know is you guys continue to deliver the value and you will be rewarded with the value. I want everybody in this company to be successful, make money and build your career here. So with that, I couldn't be more proud of the team. I appreciate, Harlan, the opportunity to have more of a free-form podcast, let it rip type of a discussion. So you guys could all hear the passion I have for this, the confidence I have in the business and that the future is very bright for Marvell. So I appreciate everybody's interest in the company.

I appreciate that, Matt. And Ashish as well. Thank you very much for the great insights today. I think you really did a great job of clearing the air on some of the controversies, misconceptions. But more importantly, laying out the strong forward growth opportunities, right, as you just highlighted. And I also want to thank everyone on the call for listening today as well. So hopefully, everybody has a great week, and thank you for tuning in.

Thank you, Harlan.

Thanks, everybody.

Thank you.

My name is Atif Malik. I cover U.S. semiconductors, semiconductor equipment and networking equipment stocks at Citi. It's my absolute pleasure to welcome Matt Murphy, Chairman and CEO, Marvell as well as Willem Meintjes, CFO. I always have to remind Matt, that he started his Marvell journey at Citi conference many years ago with Rick Hill on the stage. That's why he owes one to me every time.

That's right.

Exactly. So thank you for being here, Matt. Matt, coming out of your recent earnings call, investors are a bit concerned that something changed on the status of the two key XPU programs and that you didn't appear as confident as before on the growth from customer revenue next year. Can you clarify what you were trying to communicate to investors?

Sure. Well, first of all, good morning, everybody. It's great to see all the folks here and also on the call. just quite a couple of words upfront, and then I'll get to your question. I think everybody is curious about my answer to that. But to start off, I appreciate you hosting us here. This is Citi 10 for me. So yes, I mean -- if you can believe it back then, it was a whole different world. It was my first conference actually as CEO. And I mean, we were just happy to get our financials on file with the SEC at that point. And we had a consumer-oriented business with some enterprise and that was about it. And by the way, multiple cell ratings on the stock. Nobody wants to point the finger. But yes, anyway. It's okay. I don't blame you. I don't blame you. There's no -- and I had the pleasure of being up with Rick Hill, who is my Chairman at the time.

So anyway, fast forward, here we are 2025, it couldn't be a more different picture, right, in terms of where Marvell is positioned. I've been in the semiconductor industry for 31 years. I've been through all these major sort of product cycles going back to the early to mid-90s. And the AI opportunity in front of us is just massive. And we found ourselves at Marvell right in the middle of it. So it's still early innings for us and there's still a lot more to go. So maybe by Citi 20, it will even be a different story.

So look, a couple of things on the call. I think the first is that, yes, I want to just acknowledge and recognize concerns from investors around the way -- the answer to some of the questions were communicated. So I want to try to be helpful here this morning so we can clear the air. So the first is, as you saw, I rotated and we rotated a lot harder on the customer confidentiality side. And I want to take a moment just to explain the why, so it's not just like a canned response, but it's a real thing. The strategic nature now of these custom programs is such that it really is what defines the future architecture of these systems. And it's -- these are fundamentally trade secrets of these companies. And being in a business like this, especially with a concentrated customer base where the stakes are so high, it's a business fundamentally about people. It's about trust, it's about long-term relationships, and it's about multi-generational investments. And so you need to have that trust there. So that is the reason for the rotation to be a little bit harder on that front because of the concerns from our customers there. But that being said, it was not our intent to signal a change in our business when we said that.

Now a couple of things on that. What I was asked to do was provide guidance for fiscal '27, which is next year, I know Marvell has got this one year off, but still, I think fiscal '27 to me sounds like a long time away. I mean it's -- we just finished our Q2 of this year. We just reported those results. So we're a ways out. But the time will come when it's appropriate to do that. And once we have the visibility on the next year, we'll provide that. I mean if you look back when we started this AI journey, really back in 2023, we were -- we've been signaling and giving information to investors to be helpful, right? If you remember, the first call we had was -- I think it was May of '23 and we said, hey, we think AI is going to be like $200 million this year and $400 million the next year. That was kind of overall, it was mostly optics. And then along the way, we had an AI day and then we had another AI day and one of the latest updates we had given was that we would -- we had a target of like $1 billion plus for custom this year. And we've been tracking, and we're actually way ahead of that. So we'll continue to be helpful there, and we've been planning for growth in this business, and we still are. We still are. But I need the time. And once those plans, by the way, shore up and we have better visibility, then we'll communicate that. We'll communicate any change, if it's up, if it's down, we'll communicate that. And that's basically how we've always been doing it, and that's how I will continue to do it.

So that's some of the background on the why, but I mean I couldn't be more excited about where this business is going. I mean we had the AI day. We laid out a $55 billion custom TAM, which was up from just a year prior, and we actually provide the additional clarity, which maybe you'll have some questions on that $40 billion of that was for XPU, which gets a lot of attention, obviously, and it's very critical. And then $15 billion of XPU attached, which is really a growing and emerging market where there's a different set of dynamics where Marvell is also participating. And we're now up to like 18 -- I call it really 18-plus sockets because even since the AI day, we've won additional programs. I mean it's -- there's a lot happening right now. So I guess, when I just look at the trajectory and the fact that we're in the early innings of this business, it's relatively concentrated today because these are the first set of initial design wins we had that are ramping. This is -- we're in the early years here. But if I go out to 2028 and beyond and where the TAM is, and you see these 18-plus programs ramping, we'll also have better diversity in the business going forward. It will be easier to plan and communicate it.

So those are some of the dynamics and challenges that we're facing. But I hope that this is helpful to investors to understand the why and a little bit more of the background.

Yes, super helpful on the sensitivity around hyperscalers. Matt, most investors I talked to, they were super impressed by your June Investor Day on your IP portfolio and everything on the custom and ASIC side. Now design cycles are long on these custom silicon projects. I tell my clients that Marvell is where Broadcom was a year ago we're mostly focused on sockets right now, but there is a whole big pipeline that's coming and diversification will happen.

Now you've talked about this 20% share goal in the custom market. Can you just walk us through your assumptions on TAM? You introduced XPU attach sockets as well. Just kind of walk us through how do you get to this 20% goal?

Sure. Yes. And we have a 20% goal on the overall TAM opportunity. And that's the full boat Marvell, right, all the different product segments. But we've also said that we believe that the custom piece can also get there, albeit we're starting a little bit later. And the bulk of that is really based on the 18 wins that we outlined today.

Now there's optionality there. I mean, we said there was another 50 sockets that we're tracking worth $75 billion plus of lifetime revenue. Like I said, subsequent to the call, we've actually closed several of those. So think of that $75 billion has kind of come down because we've closed some of them. Now I think it's going to come back up because there's just a tremendous wave of opportunity right now in terms of the design activity with these customers.

And so the 20% share is really a combination of, one, our bottoms-up view of the individual opportunities. We obviously try to risk adjust and size those. And guys, let's be clear, it's a very dynamic environment, okay? I mean we're trying to call the ball obviously, three years out. We're trying to call the ball for next year and things are changing. But directionally, the good news is it's only been up. And so I think we'll see where it shakes out. But I think the bottoms-up certainly supports it.

And then also just from a top-down perspective, when we look at the TAM and the opportunity, and as you mentioned, and we can go through it if you want. The key tenets of the AI day we did in June around Marvell's leadership in process and package technology, our IP portfolio, our ability to stitch the full solution together, our ability to have a world-class, robust supply chain, deep customer partnerships and the investment profile to invest for the future, it's resonating really well with these customers.

So the other angle, which is more subjective is like, hey, tops down, why can't you guys get to this 20%? I mean the pushback I got was why can't it be bigger if you just look at the law of market share and if you're successful. So we'll see. But I think we set a very achievable goal. Now it's a big number because the TAM is growing so fast. So a big part of that equation on where the net revenue lands is, obviously, does the TAM keep growing at the rate it does and how much converts to custom, and we have those assumptions in there. But I feel like those are pretty safe assumptions at this point. And certainly, if you look at the cycle we're in, CapEx spend, design win activity. It certainly suggests that, that TAM profile going forward is going to happen. And certainly, our win rate suggests that we're continuing to knock down these wins across a wide variety of opportunities. So I still feel very good, very bullish on where this business is going.

Okay. Matt, given how vocal some Asian companies are being about their involvement with the key customers' design. Do you see, given the level of CapEx that these hyperscalers are planning that multiple companies can either work on multiple projects or perhaps even share the same ASIC projects? Or is it a winner takes all?

Yes. I think what we -- we started this a little bit at the first AI era that we did in 2024 and certainly followed it up just a few months ago, is -- when we outlined the TAM, I mean, we did talk about, and I believe there is going to be diversity in there, right? The spend is so big that I think you're going to see some variety of business models that emerge. And I think -- and actually, the presentation we talked about in June, these different models, right, which were more on a spectrum, right, how much customer IP do they want from a company like Marvell? What kind of manufacturing services do you want? How important is other kind of bespoke unique technologies in the road map, like PIVR, or CPO or some of these other things, right, which may cause some of those dynamics to change.

And so -- and just again, given the sheer spend, I think you're going to continue to see more SKUs, more variants and more opportunities and diversity as we go forward. But I still believe that the vast majority of the shipments long term out of the custom TAM, the majority market share will still come from what I would call full service or full turnkey providers where if you have all the pieces and you can actually manufacture -- design it together, manufacture it, yield it, ship it, move to the next generation, hit the beat rate, hit the cadence. That's where the bulk of the market will land.

But certainly, there's going to be room for different models, and there should be. It's -- the spend on these things is -- I mean, for example, the custom spend alone is projected to be bigger than the entire x86 CPU market by 2028. So you got to imagine there's going to be, I think, a lot of opportunities. And I think that's a good thing for Marvell. But it's not a winner take all. It's not going to be two sockets, and that's it and whoever gets it, gets it and the other one is going to be -- it's not a binary market anymore. But you're right, it's in the early innings even for the large, more established providers.

All right. Willem, going back to you though in terms of gross margins, you guys have been kind of clear that the nature of the beast is that the ASIC projects have an impact on your gross margins, maybe 58 to 59 kind of range. But as you look into next year and just kind of in terms of the operating margins, we've talked about the revenue growth rate being 2x the OpEx run rate. So how do you kind of position the company to expand the operating margins where you do benefit from the NRE payments from some of your customers?

Sure. I think maybe let's start with -- If you just go back over the last year, right, and we've been signaling very consistently that as these custom programs are ramping, and you've seen that in our gross margins, right? But at the same time, you can see this leverage that we've driven an operating margin, right? And we just guided a quarter to like 36.2 OM which, that's really starting to approach that 38%, 40% long-term target that we have, right? And so that model that we set out has really played out in the numbers that we've executed on so far.

And the one thing that we've been very consistent on is investing for growth, right? And so the way that we've managed the portfolio is that we had all these other products, right, where we call that the mass market or carrier enterprise where we did a refresh and that -- those refresh products have started ramping, and you can kind of see that and how those sort of other end markets has really recovered pretty well. And so the intensity of investment there has reduced quite significantly, and we've sort of reallocated that to our data center portfolio.

So that, in combination with the NRE that's called nonrecurring engineering. We recognize that as a contra OpEx, it reduces our OpEx and think of that as an investment that our customers are making in these programs. And so those two dynamics have really allowed us to continue to manage OpEx very tightly, and we'll continue to do that through next year.

Now we do recognize that this is sort of a -- as Matt was just saying, like a historic sort of growth opportunity, and we want to make sure that we invest. I mean these resources to go have this differentiated technology is very scarce, right? And you're seeing significant competition on those resources and we're sort of in a battle to maintain and attract those resources. And so we'll do what we need to do to continue to invest to drive that growth.

I would just to put a point on it, I mean, the R&D position that AI and cloud has just been absolutely turbocharged over the last few years because, one, we've increased R&D spending, and we've been consistent on that despite the up years and down years. We've always grown our R&D. The NRE on the new programs is kicking in. So that's an additional lift that we're getting. And then as Willem said, the strong kind of capital allocation framework that we drive in Marvell through our strategic planning process and sort of our check-ins and really looking at where our precious R&D dollars are going. By the way, we just completed our tenth annual strategic planning review, right, which is the first one was August 2016, about 6 weeks after I became CEO which was our capital allocation framework that we continue today. And so now you see well over 80-plus percent of our R&D spend all in data center. And so -- and we're able to do that and still reap the benefits of the investments we made in the past on the other broad market opportunities. So I think we're in a great place relative to the invest and the spend. And to be able to do that at scale, which is only a handful of companies really possess the team to be able to go do this.

Great. Let's talk about the other piece of your AI sales, the optics or the DSP part of the portfolio. You talked about double-digit growth in the next quarter. Can you just share with us what you're seeing in that part of the market, why there was a slowdown in the reported quarter and then the reacceleration in the next quarter? Is it just timing?

I think some of it is timing. It's not -- it's never been a completely linear business either. That one is a little bit different always by design because just for the investors to understand in the optics area for us, there's one other step in the supply chain, which you normally don't see, which is the module manufacturing in the middle. So sometimes you can get like a quarterly fluctuation there. But in general, if you just look, I mean we've been worried about down quarters and inventory digestion and resets for, I don't know, coming up on two years. And I remember back, I think it was the end of '23, I mean, I got a little cautious and everybody kind of freaked out because I was just saying "Hey, this thing has grown at like 80%. Is it going to slow down?" and it hasn't. And it's been a very, very strong performer for us. And some of this goes through the channel, that sell-through has gone great. And all the signals we're getting, which I think is a great proxy for where the AI spend is happening. If I just look at kind of what the lineup is on the go forward, a lot of strong demand signals coming through about continued future growth in that business, especially for next year. So I think there's a lot of optimism. And you see that in the numbers, everybody is kind of talking about where things can go.

But we -- and I think the other thing in that business is that we've -- since we acquired Inphi and integrated it, I mean, we've just absolutely maxed out on the resources assigned to that business. The team has done an excellent job on technology leadership in terms of moving ourselves and our customers to the next generation. And we see that cadence continuing, whether it was our move from 100 gig per lane to 200 gig. We're now 200 gig to 400 gig in the future, which translates to these other cycles of 800 gig, 1.6T, 3.2T. All those we're just absolutely heads down driving that. And that's going to be key to really enabling this next wave of AI growth to happen is sort of releasing the communication and the IO bottleneck, which we play a key role in. But it's a very, very strategic business for us. The team has done an amazing job, super proud of them. And I think we've got just a great future ahead there, given our investment profile and the quality of the team we have.

Great. Let me pause here and see if there's any questions in the audience. If you have a question, please raise your hand. Yes, this one in the back.

-- Some capital with the sale. Given your bullish forecast the next three years, gating might be the factor. Why aren't you doing an ARS?

Willem, do you want to start with that?

Yes, sure. Look, we've always had a balanced capital return plan, right? And so if you go back, I've been personally very, very focused on getting our free cash flow to be a lot more consistent. And you've seen that consistency here over the last year. And that's really resulted in a step-up in our capital return. This is, as you point out, a nice capital infusion here, and it gives us a lot of optionality. The other discussion we were just having is around the size of the opportunity, and we're very, very pleased with our IP portfolio, but this is an extremely dynamic environment, right? And so as we look across the ecosystem where there's potentially bolt-ons that can accelerate our growth and our portfolio, we're going to be looking at that. And so you should expect us to sort of step up in terms of capital returns. But at the same time, want to keep that flexibility on potential bolt-ons.

All right. And Matt, sticking with the hyperscalers. We have slowly seen entities like Sovereign and Enterprise is starting to ramp up on AI investments. And I understand your focus mostly on hyperscalers right now, but do you see your chances of getting a piece of the non-hyperscaler AI CapEx in the future?

Yes. I think the way we broke it down -- and there's different ways you can slice and dice this market. We talked about the two kind of big buckets at the AI day. One was kind of the traditional top four hyperscalers and then we had another category called emerging, which includes Sovereign by the way, if that's your specific question, but also companies now building their own potential cloud infrastructure and kind of that second but very important tier of cloud players that are vying for share.

So we have an opportunity set in all of them. Certainly, from an optics perspective, that in networking, that cuts across everywhere. And we have direct engagements on almost all those companies because we have great relationships with our module partners. They're out marketing our solutions, but we're also directly engaged and have sales teams and business development and business unit coverage of those other accounts and opportunities. And increasingly, those are also driving custom silicon opportunities as well. And I think more and more you're going to see that as those -- as their sort of CapEx and their spend rate goes up and their recognition, all of them that their need to have some differentiation. And some of it's not just differentiate to be differentiated, but their apps are different. Their workloads are different. Their use cases are different, how they construct their data center footprint is different. And so there's -- given the sheer spend, it actually opens up a lot of opportunity for us.

So that was one we called out at the AI day is kind of a big change over the last year which used to look more like a standard product kind of by the generic solution business. It looks like more and more, those are going to get a little bit more bespoke and a little bit more unique, and that's where we can really come in and add value.

[ One of the ] networking that we're seeing upsiding for some of your peers is on kind of scale out and even scale-up efforts. And you guys don't really break this out. You have in your AI sales. But can you just talk about the momentum you're seeing on the scale up side?

Sure. Yes. And today, we've given kind of these big buckets of our optics business, our custom business and then the rest of it. Over time, should any of those other areas like networking, for example, emerge to be big enough, we would start to call those out. But by the way, just on the scale out, which is really where most of that spend is today, we acquired a company called Innovium back in 2021 to get us into the higher layer, high-speed cloud networking market, and that's gone really well. That business has grown quite a bit since we acquired it. It's set for very nice growth over the next couple of years as product transitions happen. When we bought the asset, we got a position in the footprint and 12.8T switching. We put the whole team and the resources on 51.2T and that, from a market standpoint, that adoption is starting to happen. And so we're going to see that business inflect as a result. The reason I bring up the scale out to start is that the same kind of fundamental architecture team we have for that is the team that's driving the scale up solutions, which is really a combination of high-speed, low-latency low-power networking with our IO and SerDes technology. And so that -- the scale-out is -- there will be absolutely more to come there. We're very involved in the standards around that area. We certainly are getting pulled in through our connectivity products and our XPU shots on goal as well because it all fits together.

So there's more to come on that one, but that will be a large driver of TAM growth for us in terms of where the scale-out is going to go. And I think it represents a big opportunity in the industry, and we're very well positioned. We're very focused there. So more to come on that one. But it's -- I think it's going to be a big opportunity that really plays well to Marvell's strengths because we own our own SerDes IP. We have our own -- really best-in-class architecture team on high-performance switching and a very good system level view of how the connectivity and the XPU kind of road maps all play together.

Willem, on the non-AI markets, the enterprise networking and care infrastructure, they've recovered nicely. You guys have talked about a $2 billion kind of normalized revenue run rate. How strategic are those two businesses? Are there opportunities to maybe divest other areas after the auto Ethernet sale?

Sure. Yes. Look, those businesses go back -- we've got a very long history with those businesses, right? And we've really optimized them. And so they're a core part of our profitability growth engine, [ what these ] profitability driver. And so over time, once we get to that $2 billion, we don't -- we see no reason why that can't continue to grow at a similar rate to those markets. And so Matt mentioned, our capital allocation exercise that we do every year, and we look at everything. So we don't have any sort of holy cows that we're not going to touch. But that business has been sort of the core profitability driver for us. And we're actually very pleased with that recovery, as you point out, right? I think when we are sitting here, when it hit the low of, call it, a $900 million run rate, and we were saying, "Hey, it's going to get back to $2 billion". I think there was a lot of skepticism, right, whether that even comes back. And so now we've just guided $1.7 billion, right? And we do expect that growth to continue. We've seen bookings be very strong and continue to be strong in those businesses.

Yes. Just to add. I think first, yes, very pleasantly surprised with the reinflection on that business. I mean it obviously inflected down hard and then we kept wondering what's the slope coming back, and it's not been linear. I mean it started coming back and coming back. But I think I'm very pleased to see that the recovery in the booking shipments and kind of the demand outlook for our customers in those areas. And it's not just complete inventory recovery. I mean it's kind of tied into what I said earlier about how we've been able to start to reap some of the benefit of the prior investments in enterprise networking and carrier end markets. We'll hold 5-nanometer portfolio refresh. The fleet is like state-of-the-art. And so those new products are also kicking in, too. And so that's giving us some of the additional lift.

So I mean we've got a whole business group team, division that's focused on this. And they just presented out at our strategic review. They've got a growth plan, they've got ways they're going to go drive the revenue. We're giving them investment. It's skinnier than it was. But again, that's just because we put in a big lift upfront. And then automotive specifically, that was just a very different kind of business. It had its own sales team, its own quality team, its own processes. And so given kind of our scale in automotive relative to what that business is going to do, I mean it's on a great trajectory. It's going to -- Infineon will do great with that business. But for us, when we looked at the big picture, even if it doubled from where it was, it probably wouldn't move the needle net-net over the next couple of years on where we're going on data center and it was a competitive process and the economics were right. So that was the thought process on that. That was a little different.

We do want to have some balance in our portfolio. But quite frankly, this historic AI thing, we just got to go for it. I mean it's just -- it's where the TAM is going. It's a big seismic shift in the industry, and I think we would be kicking ourselves if we didn't go after it and to the question earlier, which was a great one on capital return, just maintaining some flexibility at the moment because we're in this inflection to figure out what we want to do. But certainly, if that -- if we just conclude the organic stuff we're doing is great, then we'll be much more aggressive on returning capital to shareholders. We're very focused on that, not only driving our own internal free cash flow metrics, but obviously using the proceeds. So more to come on that one.

Great. Then just coming back to the optics, there's a kind of nonstop discussion on technology trends in CPU and OCS and 1.60. And NVIDIA talked about the use of co-packaged optics at the GTC earlier this year. Can you just share with us your view on where Photonics market is going to go and when are we going to see volume adoption?

Yes. Well, it's the age-old question that we've been chasing for some time. And for those of you that attend the OFC conference, this is kind of an annual discussion, right? Like when is high-volume silicon photonics really going to hit, i.e. in mass volume inside data center for things like scale-out networking or scale-up networking. But it has been -- there's been a big shift in the last year or so relative to some of the big players really putting serious capital investments here committing to road maps, technology developments getting driven and the technology has come along. I mean, we're -- we have a strong skill set there. We're shipping relatively high volume in SIFO today, but that's for between data centers, right? Data centers so -- our DCI products. I mean it's Marvell's owned silicon photonics solution. It's our own pick, it's our own design. And by the way, we designed the entire chipset around it. We actually designed the entire module. So we have a lot of experience in manufacturing and volume here. It's a bigger knee of the curve when you go inside data center. And we've been showing off for the last couple of years, our own integrated light engine solutions, which ultimately we can plug in and integrate with either our networking products or our ASIC products.

So the change I would say is that it looks a lot more positive that this kind of an impact can happen. But I don't know if it's '28 or '29 or 230 or -- and how long it really takes to get mass adoption, I really still think that's a ways out. I think we shouldn't all get completely over our skis there. But we're going to be ready for both, okay? And that's a long-term thing that we're investing in. But just to be like crystal clear, when we look at the next several years from our customers, it's all pluggables and it's how fast can we move to the next one. And that's then augmented by some more unique solutions like AOCs or AECs, which were in both of those, by the way. We have solutions for both linear pluggable optics, which was -- had failed to get traction for the last couple of years. That will start seeing some adoption. We actually have some great solutions there. We have design wins there. We're going to be an LPO. It's really -- like we're trying to just be the one-stop shop. If you want to have a range of solutions, we don't just have religion that it has to be pluggables, it has to be the latest greatest, and we're just going to hold our nose. I think we can leverage our technology across all of these. And I think more and more, you're going to see that. But in the end, it's going to be proven, pluggable, swappable kind of high-volume scalable solutions out there. And so some of these things like -- a lot needs to be ironed out on the CPO side first, but we're not pooling it at all.

Great. We're almost out of time. Matt and Willem, thanks for being here.

Yes, I appreciate it. Thank you. Thanks, everybody. We'll see you today.

Good afternoon, and welcome to the Marvell Technology Inc. Second Quarter of fiscal year 2026 Earnings Conference Call. [Operator Instructions] Please note this event is being recorded. I will now turn the conference over to Mr. Ashish Saran, Senior Vice President of Investor Relations. Thank you. You may begin.

Thank you, and good afternoon, everyone. Welcome to Marvell's Second Quarter Fiscal Year 2026 Earnings Call. Joining me today are Matt Murphy, Marvell's Chairman and CEO; Willem Meintjes CFO; Chris Koopmans, President and COO; and Sandeep Bharathi, President, Data Center Group.

Let me remind everyone that certain comments made today include forward-looking statements, which are subject to significant risks and uncertainties and that could cause our actual results to differ materially from management's current expectations. Please review the cautionary statements and risk factors contained in our earnings press release, which we filed with the SEC today and posted on our website. as well as our most recent 10-K and 10-Q filings, we do not intend to update our forward-looking statements.

During our call today, we will refer to certain non-GAAP financial measures. A reconciliation between our GAAP and non-GAAP financial measures is available in our earnings press release. Let me now turn the call over to Matt for his comments on the quarter. Matt?

Thanks, Ashish, and good afternoon, everyone. For the second quarter of fiscal 2026, Marvell delivered record revenue of $2.006 billion, reflecting a 6% sequential increase and strong 58% year-over-year growth. Our data center end market continued its strong momentum, growing 69% year-over-year, fueled by robust AI demand. We also saw a solid recovery in our enterprise networking and carrier infrastructure end markets. which collectively grew 43% year-over-year. We expanded our non-GAAP operating margin by 870 basis points year-over-year to 34.8% and delivered record non-GAAP earnings per share of $0.67. And up 123% year-over-year. We also delivered $462 million in operating cash flow, up significantly from the $333 million in the first quarter. Robust cash flow generation is enabling us to continue to return significant capital to our stockholders. We have repurchased $540 million of stock through the first half of the fiscal year with approximately $2 billion remaining in our authorization.

At the beginning of the third quarter, we completed the divestiture of our automotive ethernet business and a $2.5 billion all-cash transaction at a very compelling valuation I'm pleased with our team's execution in closing this transaction ahead of schedule. The proceeds from this transaction provide us flexibility to continue to drive our ongoing stock repurchase program, and deploy capital to further bolster our technology platform. The auto divestiture aligns with our strategy to focus the company on what we expect to continue to be a massive AI opportunity in front of us by purposely redirecting our investments towards data center relative to our other end markets. That strategy has been very successful with data center alone now driving 3/4 of our total revenue. The auto divestiture further reduces the relative proportion of revenue from our non-data center end markets. As a result, starting in the third quarter, we will consolidate our non-data center end markets into a new single communications and other end markets. Willem will cover this in more detail in his prepared remarks.

During the quarter, we hosted a highly successful custom silicon investor event in June, where we outlined an expanded $94 billion data center TAM for calendar 2028, a 26% increase from our prior view. We also unveiled a new fast-growing custom silicon product category of XPU attach, updated our custom design win Board to 18 multigenerational XPU and XPU attach sockets and highlighted over 50 new pipeline opportunities with an estimated $75 billion of lifetime revenue potential. Based on the sockets we have already won, we concluded with our plan to grow our data center market share from 13% of a $33 billion TAM in calendar '24 to 20% of a $94 million TAM in calendar '28.

Let me now take a moment to share how we are enhancing our leadership structure to further capitalize on significant opportunities in the large and fast moving AI and cloud markets. We have promoted 2 exceptional leaders. Chris Koopmans, to President and COO; and Sandeep Bharati to President, Data Center Group and consolidated substantial parts of the organization under their leadership. The proven track record of innovation, execution and results positions them to accelerate Marvell's growth. Chris joined Marvell 9 years ago, and he has been a key enabler of our transformation to a leader in the data center market. He successfully led sales, our networking business and most recently, Global Business Operations and Marketing. I'm pleased to see Chris take on sales again along with managing our non-data center businesses in corporate development. This expanded role now encompasses end-to-end revenue execution from go-to-market strategy and customer engagement to operations and long-term strategic planning.

Sandeep joined us in early 2019 to lead our central engineering team and accelerate development of our technology platform. He was instrumental in driving Marvell's leap to 5-nanometer process technology leadership and integrating Avera, the custom business we acquired that has since become our largest growth opportunity. Under Sandy's guidance, our engineering teams have successfully delivered multiple highly complex custom XPU and XPU attach projects in the high-volume production with first-time silicon success. With this promotion, Sandeep now has overall responsibility for our data center business in addition to its continued leadership of data center engineering and central engineering. This unifies full ownership of our largest and most important business under a single leader, spanning the entire product life cycle, from technology platform, IP and road map to customer engagement, product definition and chip development.

Let me now discuss our results and expectations for each of our end markets. In our data center end market, we achieved record revenue of $1.49 billion in the second quarter, growing 3% sequentially and 69% year-over-year. The strong performance was led by our custom XPU and XPU attached products as well as our electro-optics interconnect portfolio. AI and cloud continue to be the primary drivers accounting for over 90% of our data center revenue with the remainder coming from the on-premise portion of our data center end market. We expect on-premise revenue to remain stable at an annualized revenue run rate of approximately $500 million. Looking ahead to the third quarter, we expect revenue from our Electro-Optics products to grow double digits sequentially on a percentage basis as we continue to benefit from our market-leading position in AI interconnect. Our custom business is also performing well and remains on track to grow in the second half of the fiscal year compared to the first.

However, we expect growth to be nonlinear in the custom business, with the fourth quarter substantially stronger than the third. As a result, we expect overall data center revenue in the third quarter to be flat sequentially with electrooptics strength offset by lower custom revenue. On a year-over-year basis, we expect data center revenue to continue to deliver strong growth in the mid-30% range in the third quarter. We are very pleased with the progress of our 18 XPU and XPU attach sockets, several of which are already in volume production. We are making excellent progress on development of the remaining sockets with all of them expected to ramp over the next couple of years. The success of our initial wave of custom programs, combined with rapidly growing industry interest in custom silicon has expanded our design win pipeline to over 50 new opportunities. As next-generation XPU and XPU attach products increasing complexity, we believe it will become even more critical for customers to partner with a full-service custom silicon provider like Marvell.

Since our event in June, our team has won additional sockets, adding to the 18 sockets we had already discussed. Collectively, these new wins represent multibillion-dollar lifetime revenue potential and we remain deeply engaged in advanced architectural discussions of many of the opportunities still in the funnel. As next-generation AI data centers evolve, scale-up networks are becoming essential to tightly interconnect tens, hundreds and eventually thousands of XPUs within and across racks. These require ultra-low latency and multi-terabit bandwidth to meet the demands of training and inference workloads. Marvell's multigenerational custom engagements with the hyperscalers gives us unique visibility into upcoming XPU architectures enabling us to design scale-up switches supporting both open standard ethernet and UA link fabrics purpose-built for AI. Combined with Marvell's leadership in Ethernet switching and proprietary high-speed, low power, low latency SerDes IP, we are strongly positioned to lead this market inflection. We are investing in developing scale up switches tailored to each customer's protocol of choice and look forward to updating you on our progress.

Beyond switching, our interconnect portfolio extends the opportunity. While copper dominates the scale-up links today as networks expand and bandwidth grows, optics adoption will follow. This represents a large opportunity for Marvell's full suite of interconnect products and technologies, including for active electrical cables or ADCs and active optical cables or AOCs. Retimers for PCI, Ethernet and UA Link, and silicon photonics for near packaged and co-packaged XPU optics. Our AEC and AOC DSPs are already in the market and our retimers are in customer evaluation. We have demonstrated our 6.4 silicon photonics light engines and expect our technology to be a key enabler of NPO and CPO implementations once the industry is ready to adopt. Collectively, between switching and interconnect, we see a massive scale-up opportunity for Marvell over time.

Turning to our electro-optics interconnect portfolio. Our PAM and DCI franchises continue to lead the industry in enabling the build-out of AI and cloud infrastructure. Demand for our 800 gig PAM DSPs remain strong with a long life cycle still ahead. We have also begun volume shipments of our next-generation 200 gig per lane 1.6 TPAM DSPs to multiple customers. We expect adoption to accelerate in the next several quarters. Looking further ahead, we are driving the next optical technology transition. At this year's Optical Fiber Conference, we demonstrated our 400 gig per lane PAM technology, a critical innovation and step towards enabling 3.2G optical interconnects. This milestone underscores Marvell's leadership and pushing the boundaries of next-generation optical connectivity.

Our data center interconnect business also continues to expand with adoption proliferating across large hyperscalers. Collectively, the custom and electro-optics product lines I just described, now account for over 3/4 of our total data center revenue. The balance comes primarily from our data center storage, switching and security portfolios, each of which is showing solid progress. Our data center storage revenue has improved significantly, reflecting a return to health in both the SSD and HDD markets. In AI and cloud switching, our 12.8T products continue to ship in high volume, while our next-generation 51.2T switches are now ramping. Adoption is accelerating, and we expect these products to be a major driver of Switch revenue growth in the next fiscal year. In the security market, we recently expanded our collaboration with Microsoft Azure on our hardware security modules building on a long-standing and trusted relationship with this customer.

Now let me turn to our enterprise networking and carrier infrastructure end markets. In the second quarter, Enterprise Networking revenue was $194 million, and carrier infrastructure revenue totaled $130 million. Combined revenue for these end markets grew 2% sequentially and 43% year-over-year. Looking ahead to the third quarter of fiscal 2026, we expect aggregate revenue from enterprise networking and carrier infrastructure to grow sequentially by approximately 30%. This growth is driven by normalizing customer inventory levels and strong adoption of our refreshed product portfolio. As a reminder, we recently migrated these products to advanced process nodes and investment we expect to yield benefits for many years to come given the long product life cycles in these markets.

In the consumer end market, second quarter revenue was $116 million, up 84% sequentially and 30% year-over-year. Gaming demand and its seasonality continues to be the primary driver of this business. For the third quarter, we expect consumer revenue to be down sequentially in the low single digits on a percentage basis.

Turning to our automotive and industrial end markets. Second quarter revenue was $76 million, flat both sequentially and year-over-year. For the third quarter of fiscal 2026, reflecting the divestiture of our automotive Ethernet business, we anticipate overall revenue of approximately $35 million from this end market. This includes a mid-single-digit million dollar contribution from our automotive Ethernet business prior to the transaction closing.

In summary, in the second quarter of fiscal 2026, we continued to deliver operating margin expansion, earnings per share growth and new revenue records. Looking ahead, we expect momentum to continue in the third quarter, with total company revenue forecast at $2.06 billion at the midpoint, representing 36% year-over-year growth. Excluding revenue from automotive Ethernet, the implied revenue growth for Marvell's go-forward business would be closer to 40% year-over-year at the midpoint of our forecast for the third quarter. We also expect to continue driving operating leverage with non-GAAP earnings per share forecast to grow 10% sequentially at the midpoint of guidance, more than double our projected revenue growth rate. Our second quarter results and third quarter guidance reflect robust contributions from our AI-driven data center end market, complemented by strong recovery in our enterprise networking and carrier infrastructure end markets. At the same time, our custom AI design engagements at are an all-time high, with customers showing very strong interest in our broad range of differentiated technologies.

As I discussed earlier, our team continues to accumulate new wins, we are pleased with the strong progress across both current and next-generation custom programs, which reinforces our confidence that we can achieve our long-term customer revenue goals. In addition, our market-leading electro-optics franchises continue to see strong demand for both current and next-generation solutions, and our scale-out switching platforms are positioned for strong growth. Over time, the emergence of scale of networking for AI infrastructure should provide another strong tailwind for Marvell. With that, I'll turn the call over to Willem for more detail on our recent results and outlook.

Thank you, Matt, and good afternoon, everyone. Let me start with a summary of financial results for the second quarter of fiscal 2026. Revenue in the second quarter was $2.06 billion, growing 58% year-over-year and 6% sequentially. Data center was our largest end market, contributing 74% of total revenue. GAAP gross margin was 50.4%, non-GAAP gross margin was 59.4%.

Moving on to operating expenses. GAAP operating expenses were $721 million, including stock-based compensation amortization of acquired intangible assets, restructuring costs and acquisition-related costs. Non-GAAP operating expenses came in at $493 million, slightly below our guidance. Our GAAP operating margin was 14.5%, while non-GAAP operating margin was 34.8%. For the second quarter, GAAP earnings per diluted share was $0.22. Non-GAAP earnings per diluted share was $0.67, reflecting year-over-year growth of 123%, which is more than double the pace of revenue growth, demonstrating the significant operating leverage in our model.

Now turning to our cash flow and balance sheet. Cash flow from operations in the second quarter was approximately $462 million, growing by $129 million from the prior quarter. Our inventory at the end of the second quarter was $1.05 billion, a decrease of $20 million from the prior quarter. We returned $52 million to shareholders through cash dividends. In addition, we repurchased $200 million of our stock in the second quarter. In June, we completed the public offering of notes totaling $1 billion and used most of the proceeds to repay existing debt. As of the end of the second quarter, our total debt was $4.5 billion with our gross debt-to-EBITDA ratio of 1.63x and a net debt-to-EBITDA ratio of 1.9x. Our debt ratios have continued to improve as we have driven an increase in our EBITDA. As of the end of the second fiscal quarter, our cash and cash equivalents were $1.2 billion.

We recently completed the divestiture of our automotive Ethernet business in $2.5 billion all-cash transaction. Proceeds from this sale give us flexibility to continue to drive our ongoing stock repurchase program as well as invest further in our technology capabilities. Turning to our guidance for the third quarter of fiscal 2026. We are forecasting revenue to be in the range of $2.06 billion, plus or minus 5%. As a reminder, this forecast includes revenue in the mid-single-digit millions of dollars from the automotive Ethernet business before the completion of the divestiture. If the divestiture has not taken place, and we had operated the automotive Ethernet business for the full quarter, we would have added approximately $60 million to our guidance. We expect our GAAP gross margin to be between 51.5% and 52%. We expect our non-GAAP gross margin to be between 59.5% and 60%.

Looking forward, we anticipate that the overall level of revenue and product mix will remain key determinants of our gross margin in any given quarter. For the third quarter, we project our GAAP operating expenses to be approximately $719 million. We anticipate our non-GAAP operating expenses to be approximately $485 million. For the third quarter, we expect GAAP other income and expense, including interest on our debt and the gain from the divestiture of our automotive Ethernet business to be an income of approximately $1.8 billion. Non-GAAP other income and expense, including interest on our debt, is expected to be an expense of approximately $33 million. We expect a non-GAAP tax rate of 10% for the third quarter. We do not expect the recently passed tax bill act to have a material effect on our current year's non-GAAP tax rate. We expect our basic weighted average shares outstanding to be $863 million and our diluted weighted average shares outstanding to be $870 million. We anticipate GAAP earnings per diluted share in the range of $1.98 to $2.08. We expect non-GAAP earnings per diluted share in the range of $0.69 to $0.79.

As Matt mentioned, we plan on updating our revenue by end market classification beginning next quarter. Over the past several years, our strategic focus on expanding revenue in the data center market has delivered strong results, driving significant growth in this end market. On a relative basis, Data center revenue has more than doubled as a percentage of total company revenue from 34% in the second quarter of fiscal 2024 to 74% in the second quarter of fiscal 2026. As a result, in our most recent quarter, our 4 other end markets collectively represented only 26% of total company revenue. The divestiture of our automotive Ethernet business further reduces the relative contribution of our non-data center end markets. Looking ahead, we expect data center to continue outpacing all other end markets in both size and growth rate. As a result, our fiscal Q3 results will be the last quarter with the current classification and our Q4 guide will reflect the streamlined revenue reporting. Results will be reported in 2 categories: data center and communications and other. The composition of our data center end market will remain unchanged. The new communications and other end market will consolidate revenue currently reported separately from our enterprise networking, carrier infrastructure, consumer and auto industrial end markets.

We will continue to provide qualitative commentary in our earnings discussions, highlighting notable developments within submarkets in the consolidated communications and other end markets. We expect most of the revenue in the new communications and other end markets come from our current enterprise networking and carrier infrastructure end markets, which have both continued to recover. On a combined basis for these 2 end markets, our guidance for the third quarter of this fiscal year implies an annualized revenue run rate of approximately $1.7 billion compared to the low point we saw in the first quarter of fiscal 2025 of approximately $900 million. Consistent with prior comments, we expect these 2 end markets to collectively generate approximately $2 billion in annual revenue over time. Additionally, as we have stated previously, we anticipate annual revenue of approximately $300 million from our consumer end market. And following the divestiture of our automotive Ethernet business, approximately $100 million from our industrial end market.

In conclusion, we are executing on our strategy, driving strong revenue growth and expanding our operating margins towards our long-term target. In addition, our balance sheet has continued to strengthen and provides a solid foundation to support our growth opportunities. With that, we are ready to start our Q&A session. Operator, please open the line and announce Q&A instructions. Thank you.

[Operator Instructions] Our first question is from Ross Seymore with Deutsche Bank.

I want to dive into the guidance for the custom business, Matt. I appreciate the lumpiness of it, but could you give any more color on the headwinds are in the third quarter? And then what gives you the confidence and any sort of magnitude on the increase in the fiscal fourth quarter?

Yes. Thanks, Ross. And I think you captured the right phrase, which is lumpiness. I think this is normal to see, particularly with the large hyperscale builds that happen and especially as you ramp them into production, which we've done this year on a number of programs. So this is not unusual. Fortunately, our optics business is quite strong in the coming quarter, and that's growing double digits. And then as we said -- as I said in the prepared remarks, we see a demand increase again in custom. So yes, there's nothing unique there, Ross, other than we've spent the last couple of years ramping these into production, and we've got kind of a 1 quarter digestion with the recovery in Q4. I will say that, overall, we expect custom to be up in the second half over the first half. And so you should expect a strong fourth quarter for custom

Our next question is from Tore Svanberg with Stifel.

Yes. This is Jeremy calling for Tore. Maybe if you could provide a little bit more clarity on the design wins that you're seeing, how much of your custom products revenue that you expect in the second half is coming from some of these new programs? And how much is coming from some of your existing design wins. Any kind of color or clarity you can provide would be very helpful.

Yes. Thanks, Jeremy. Good to hear from you. And yes, I'll actually turn this one over to Chris to talk about the design win momentum we're seeing in the opportunity set as it relates to your question. Thanks. Go ahead, Chris.

Yes. Thanks, Matt. So yes, it's truly an exciting time to be in the custom silicon business for data center. We have a tremendous amount of design activity more than I've ever seen in my 9 years at Marvell. And ultimately, we're seeing that across XPU, XPU attach emerging and existing hyperscalers. And I should say that even since our event in June, where we said that the XPU attach opportunities were in the sort of several hundred million dollar design win lifetime that's grown from there. Some of the ones we're chasing now are much, much larger than that just as the hyperscalers out these rack scale infrastructure. So yes, since June, the design wins that we've added, these are very meaningful thinking the billions of dollars for the new design wins. If you put it all together, it just gives us even more confidence in our 20% share target and it's an incredibly fast-growing market.

And maybe a follow-up in terms of is there any impact that you're seeing from supply constraints anywhere along the supply chain -- any impact from tariffs that you can see from your end?

Yes, great. I'll let Chris runs our global operations, so I'll have him cover that, and then Willa, maybe you can make a quick comment on the tariffs, and I'll add.

Sure. Yes. Certainly, the supply chain is very tight. It requires very tight coordinations with our customers and very strong execution by our team. I'm very proud of our team to have met this ramp over the last year and very confident in our ability going forward. We've really been able to meet everything that our customers have needed. But it is tight and we have very strong coordination and execution.

And Jeremy, on tariffs, it remains a very dynamic environment. But really, we haven't seen any impact on our business to date. We keep tracking it very, very closely. But as we look across all the different end markets that we're addressing really haven't seen any significant impact.

Our next question is from Aaron Rakers with Wells Fargo.

Yes. Kind of building on the earlier question, just to level set as we think about the lumpiness in the custom XPU business, I'm curious, you've had obviously a very talked about lead customer. I'm curious, as you're looking at the business today, how concentrated are you amongst your lead customer? And if we look out, let's say, 6 months or even 12 months from now, how do we expect to see some of these additional design wins start to fold into the XPU revenue stream? I'm just -- I'm trying to gauge the timing of some of these additional opportunities.

Yes. Thanks, Aaron. And I think you said it right. We had started a few -- just a few years back, a couple of AI days ago, really talking about a handful of sockets that we're going to be kind of our initial lead and those have now ramped and are ramping, albeit lumpiness we're seeing in the short term, but those are happening. And then on top of it, the 18 we talked about just a couple of months ago at the AI day, those are all either starting now, next year, really in the next sort of, I'd call it, between now and the next 18 to 24 months, those will all start to layer in. And then as Chris mentioned, we've actually secured some incremental wins. So think of it as kind of 18-plus. So there's a journey there from a handful to 18-plus to beyond and that's really what we're focused on is driving our market share from where we were just a couple of years back at 10% share in '23, 13% the year after and driving to 20% over time, and we're certainly getting a lot of confidence in that with just the recent design activity that Chris talked about. It's kind of unprecedented. It's almost episodic right now.

And then as a quick follow-up, I know NVIDIA this week talked about scale out or scale across networks. I'm curious on how Marvell sees this opportunity moving from just not scale out and scale up but scale across DCI. Any framing of how big of an opportunity that might represent for Marvell?

Yes. Maybe I'll have -- I'll comment and Sandeep will comment. It's certainly something we're aware of. There's a number of different opportunities that keep layering in that would leverage our networking and our connectivity technology. Sandeep, I don't know if you have any additional thoughts, but this is something that is relatively new. But Sandeep, go ahead if you've got some thoughts.

Yes. Thank you, Matt. So in terms of scale-up opportunities, there is certainly, aside from the lead GPO player who has its own proprietary scale of fabric. There's a huge demand for Ethernet and purpose-built fabrics such as the UA link. And we see a lot of traction over the next couple of years for the scale-up requirements, and we are investing heavily to bring our scale up switches to the market, and we see momentum in the next couple of years. So we will have standard products using our state-of-the-art, low latency scale-up switching IP portfolio, some of which we acquired from Innovium, which has been a great asset for us. So we are very confident of scale-up switches being a key growth driver for us in the next couple of years.

Yes. And then more to come in the future, Aaron, on the other type of opportunities, but thanks for covering that Sandeep. appreciate it.

Our next question is from Vivek Arya with Bank of America.

Just a near and longer-term question on your custom business. So just near term, Matt, do you think Q4, your data center growth can accelerate year-on-year from the Q3 levels that you gave, just so that we got levels that are our models. And then as we look at 2026, one of your XPU competitors has suggested their business can grow 60% I think yesterday, Jensen kind of threw a 50% or so. So whatever industry growth rate seems to be in this 50% or so Zipcode for next year? Do you think Marvell has the visibility today around timing and magnitude of your large projects to kind of say that your business can sort of grow in line with what industry expectations are? Or are there other puts and takes we should keep in mind.

Yes. Thanks, Vivek. Yes, a couple of things. I think one is our custom. We don't do an annual guide, and we typically just guide a quarter at a time. I'll get to Q4 in a minute, but just as a baseline. And then I would say on the annual stuff, we've only done that very, very rarely, and that's typically been later in the year as we have more visibility. So just to set the stage. Look, I think the overall momentum in the business has been very strong for several quarters now. And I think I gave you some of the data points, which is that custom would be up in the second half versus the first half. You can look at our optics performance, Q2, Q3, especially the Q3 up double digits.

And then obviously, when you look at the big picture, we're very pleased, which no one has asked the question about yet, but it layers into the big picture on overall Marvell performance is the very strong recovery in the core business and enterprise networking and carrier I mean just for reference on that business, we hit a low point during the inventory recovery cycle at about a $900 million annualized run rate and implied in our Q3 guide, this business goes back up at like a 17 run rate. So very, very strong recovery both on inventory as well is on new programs that are kicking in and new products that are in the next technology node. So that's all a positive in terms of the setup for Q3 and Q4.

Our next question is from Tom O'Malley with Barclays.

Mat, I'm going to hit on the ASIC topic again. So apologies just one to dive in for a little more clarity. But when you look at what the digestion that's occurring in the October quarter is, is that 1 project winding down, while another is then winding back up in the fourth quarter? Is that just a temporary pause? Like any color on what's happening there? Is that just traditionally, you see certain pockets where customers take product and then they stop. But is there anything to do with the product transition there as well? Any help there would be useful.

Yes. Thanks, Tom. No problem for the question. Yes. No at a high level, these are existing programs. And it's really a it's really just a timing issue in terms of how we deliver the product and when the customers builds are occurring and when they want the product from us. So given that this is -- as I said earlier, we're in the early stages of custom. This is really our first big first year with the handful of sockets that will translate over to many more it's really just a timing issue between the quarters. So it's just more apparent. Now over time, we do see a lot more diversity in this part of the business in Marvell as additional programs ramp. But obviously, we're starting from a pretty low base just a couple of years back.

Helpful. And then just as a follow-up on the optical business, you're guiding to double-digit growth in the October quarter. You've heard others during this earnings period, you talked about supply constraints, particularly on the laser side. You're obviously a component provider that's going into these modules. But in terms of the ecosystem. Are you seeing any stops and starts there in terms of product ramps as well? Or are you hearing about any component issues? Or are you relatively immune from that in your ramp?

Yes. I'll lead off and I'll let Chris comment if it's appropriate. I mean, look, I think we've ramped this optics business just massively, okay, over the last few years. And very successfully, by the way. So I wanted to just echo what Chris said, our business unit team, sales team and operations team have done a great job. We have deep partnerships up and down the supply chain and with the key module companies to really plan our business together. So there seems like there's always something going on, but I think we've been able to just manage through it and continue to grow quite dramatically if you look at the ramp over the last few years. So I think there's always noise in the system, Tom, relative to different pieces of it. But I'd say, overall, we're tracking really well. Chris, do you have anything to add or do I capture that?

I think you captured it. Just very strong partnerships with our customers and trying to stay one step ahead of all the changes and executing very well.

Our next question is from Timothy Arcuri with UBS.

Matt, so you're guiding data center flat and OpEx is up double digits. Since you're guiding optics up double digits, can you give us a sense of sort of what the baseline is for the optics business? I know you did provide that the AI revenue would cross over half of the total company revenue. Is that happening as soon as fiscal Q3, so is optics plus custom at 50% of the total company revenue. I'm just kind of wondering because you're guiding optical, I'd like to see if you can give us some sense of what the baseline was coming off fiscal Q2.

Yes. Let me just start off real quick and I'll see if Willem wants to add. I mean we haven't updated that number. I don't think since Q4, where optics was about half, custom was about a quarter and then other was about 25%. Obviously, optics and customer both come up since then, but we haven't exactly put a beat on that and updated that exact mix. Willem anything -- any commentary that would be helpful. It's obviously something to we're probably not going to update on a quarterly basis. I totally get the question, but Will, anything to add?

No, that's the right framework and to exactly what Matt said, we take that guidance we gave in Q4, and you can apply your growth rates. And it's just not a number we're going to be sharing every quarter, but that should give you a good sense of what it is.

Okay. But is the total -- just because, Matt, you did say last quarter that the total AI number would be half the company before the end of the fiscal year. I mean, can you at least provide a sort of a mile post that? Is that happening in fiscal Q3? Or will that happen more in fiscal Q4?

Yes. I think I'd have to give you a follow-up, Jim. I don't have the spreadsheet right in front of me on that. But it's clearly -- I mean, the -- yes, I think with the puts and takes, custom up second half over first, strong optics, yes, I don't have that number right at the tip of my fingers. But it's definitely trending the same way. I wouldn't say there's any directional change there.

Our next question is from Harsh Kumar with Piper Sandler.

Yes. I had a question on the scale of the AI business. I think you mentioned you had 18 wins before you might have picked up, I think you suggested a couple of more wins. So I wanted to understand of all the custom and attach chips that Marvell is working on many of them are actually producing revenues today? I want to understand kind of like where we are today because we understand that you are aiming for 20% of $94 billion by 2028. So I'm trying to understand where we stand today and kind of knowing where we're headed to.

Yes. Thanks, Harsh. Chris, do you want to give some commentary on that one?

Sure. Yes. So certainly, there are multiple -- several that are in production today and have been since late last year. And of those 18, they're all either sort of -- they're either going to production now or have gone to production this year or into next year. So what you're seeing is pretty much every quarter, you're seeing new parts of those programs moving into production. And ultimately, we see that will continue to grow over time.

Okay. And then just maybe broadly, very broadly, help us understand, and I'm not asking for any customer -- specific customer, but if most of your wins or all of your wins largely speaking are on track. And the reason why I'm asking is when we talk to clients, investors, there's just a lot of controversies. So any kind of statement that you can make would be helpful.

Well, there's always controversy, Harsh, I mean, I think that's why at the end of the day, that was a big motivation for us with respect to you guys and the broader investor community around, our AI Day was really trying to frame where we're driving the business, what the technology differentiation is what the opportunity set is, breaking it down actually and a lot more granularity than I think we've done before relative to hyperscale versus emerging XPU, XPU attach, the relative size of those opportunities. And so that's -- and we gave some commentary today just that we're tracking against those and have now closed some. So that's going to be the focus of how we think about the go forward. But given the design win momentum we're seeing, clearly, we're continuing to garner new incremental business from really across the board, the traditional big hyperscalers as well as the emerging generation. So that's -- hopefully, that's helpful.

Our next question is from Jim Schneider with Goldman Sachs.

I was wondering if you could maybe address capital allocation from a high level for a moment. If you look at your automotive Ethernet business, that's a very attractive price you're able to get from that. So maybe you can maybe talk a little bit about the intended use of the proceeds with your bias is more towards tuck-in acquisitions that allowed you to pursue the AI strategy even faster or buybacks? And then more broadly, are you open to potential sale of other components of the business at the right price, whether that be carrier consumer or otherwise?

Yes. Thanks, Jim. It's a thoughtful question, and I'll up-level it just for a second. So yes, this is a driving force our capital allocation framework on how we run the company, and it the automotive divestiture and then the proceeds are just kind of an output of that. And I'll -- in a minute, I'll just have Willem comment on that. But as a background, we have run since August 2016, so basically 6 weeks after I became CEO, we implemented a strategy process, which was really our capital allocation framework on how we think about investing our R&D dollars primarily. At that time, too, there were some opportunities around buybacks and so forth given where Marvell was at that time. But primarily, we drive it from the strategy first. And our vision is -- and we -- and by the way, we just completed our tenth strategic review a couple of weeks ago.

So we've been doing this kind of year in and year out. And what you've hopefully seen is that over time, we've continued to evolve the company from really a consumer enterprise kind of focused company to a data center, AI, first company. And I'd say even in the last few years as we made the pivot, we've now got our R&D spending well north of 80% of our total spending in AI and data center. And that number has come up probably from, I don't know, 60% just a few years ago and then way back when it was almost nothing because we had no business there. So that -- and so around your question then, and when we looked at automotive as an example, it was just -- it was a great business. We had built it up from scratch. It was -- it continued to stay a small portion of our total revenue. And as the AI thing took off, it became even smaller, and then we had this opportunity to give it a great new home and Infineon, which they found -- and for Marvell, we obviously got significant and compelling valuation for that. We have the proceeds. And so now we're looking at how to deploy those. And I think it's -- it's going to be some -- it's not decided at the moment. We just closed this. And by the way, great job to the team. I think this was -- we had said end of the year, and we closed it in early August. So it was a huge, huge win for everybody involved.

And I'll let Willem comment in a moment, but I think the answer is probably some of both. I mean we're definitely going to keep looking at our organic investments to figure out how to differentiate and win in AI. If there are tuck-ins and things we can do, that's obviously on the table. But we're -- we've been consistent really since -- for probably the last 4 or 5 years, which was we invested early and heavy in M&A to build the portfolio we wanted. We've done a lot organically to build up our capabilities. So we're in great shape there, but we're always going to look. Willem, maybe a little bit to add to this? I know it's a long answer, but it's I think helpful for investors to know how we think. But Willem, anything to add.

Yes, I'll just add a couple and just also call out to the team for doing a phenomenal job on getting this deal closed in particularly 4 months a deal of this size and complexity is a phenomenal job. So Jim, when you look back at the last couple of quarters here, we've really driven an increased level of buybacks really through much more consistent execution on our free cash flow. And so as a basis, you should expect for us to continue to driving that and having a focus on very consistent free cash flow execution driving a higher level of buybacks and then as Matt mentioned, I think this additional capital really gives us a lot of flexibility around being opportunistic on doing more buybacks. But at the same time, we're at this historic moment in terms of the size of this AI market and where we do see tuck-ins that can accelerate our book -- our road map towards addressing that, we'll take advantage of that.

Our next question is from Harlan Sur with JPMorgan.

This goes back to one of the previous questions. You know the noise level out of Asia on your lead customers follow-on 3-nanometer XTU program continues at this deafening pace, right? With your Asia competitor they're essentially claiming victories on 3-nanometer. So what's the update with Marvell's 3-nanometer XPU follow-on program with your lead customer? I think last earnings call, Matt, you talked about securing 3-nanometer wafer capacity, packaging capacity production in calendar '26. Is this program still tracking? What's the confidence level on this program still driving growth next year? And then maybe just an update on your third XPU customer win at 3-nanometer, which was supposed to wrap back half of calendar '26. How is this program tracking as well?

Yes. Thanks, Harlan. I appreciate the question. I understand the noise. As I said, though, in my answer to the earlier question, we're at a point where the initial programs and wins we have are ramping. We've increased our opportunity set pretty significantly to -- from a handful of sockets to this 18-plus. And we're really driving to the market share targets in the future. And just given the massive sort of focus in this area and sensitivity, commenting on just the individual sockets at this point is only probably increasing the noise level. And what we're really focused on is winning incremental designs, executing on the ones we've got and driving the business forward and ultimately trying to get 20% of a $90-plus billion TAM in the future. That's where that's at.

Our last question is from Quinn Bolton with Needham & Company.

I wanted to follow up just kind of on the scale-up, switch fabric opportunity. It seems like it's bigger part of the XPU attach market, and they're different flavors, Ethernet, UA Link. Just wondering, can you give us a sense when Marvell may have its first products ramping to revenue? Is that a calendar '16 event? Or is it going to be more UA link based and more likely a calendar 2017 event? And then I've got a follow-up.

Yes. Thanks, Quinn. I think maybe I'll have Sandeep add a little bit more, but I think he did a good job framing it. I will just say though to maybe up level for a second. I think on the scale up, it really is a great combination of key Marvell IPs all into one, especially our low-latency switching IP, our SerDes and then just the ecosystem we're living in relative to XPUs. And then this being a key XPU attached that's fundamentally almost a chipset type of a decision. So Sandeep, anything else to add on that. I think we haven't really articulated a lot yet publicly on what we're doing, but there's a huge amount of momentum here, and we're engaged very broadly in this area. Sandeep, any thoughts and closing remarks on this one.

Yes. Thank you, Matt. So definitely, we are investing to bring UA link and Ethernet-based products as we engage with our customers and working very closely with our customers' time line, what I would say is product introductions in the UL Link and Ethernet space for scale-up specifically will be in the next 2 years. And certainly, with the assets that we have -- not only are we looking at link-based products, in interconnect, we're already starting to see the use of ACs in the near term and AOCs all for -- which is active electrical cables and active optical cables positioning us to participate in these markets. So for UA Link and Ethernet-based specifically, it will be in the next 2 years.

Got it. And then I just wanted to ask how are you guys is have a very substantial business in the DSP-based optical modules. A few of your peers have started to note that I believe 3 hyperscalers are beginning to ramp LPO modules. Can you just kind of frame for us? Do you think that the -- are there substantial LPO developments beginning to occur? Are they pretty niche applications? I mean just any sense of LPO penetration of the overall optical transceiver market? Is it likely to stay in the low single-digit percentages. Do you see it getting bigger than that over a couple of years? Just like to hear your thoughts since you guys are obviously the incumbent.

Yes. Thanks, Quinn. Yes. No, it's happening at a smaller scale. And by the way, we're in some of those, too. We have active wins, and we're going to production and those types of modules as well. But just given the sheer scale of the DSP-based pluggables it just ends up being a very small number and more of a niche use case, but a valuable one, if the customer really needs it and can implement it and get it working in a production, it can be a benefit. But the vast majority we see today still is -- and for the foreseeable future is pluggable.

All right. Operator, I think that's it. I'm just going to make some closing remarks. Okay. So anyway, thanks, everyone, for joining. I appreciate all of your interest in Marvell and joining the call and listening in. Just a couple of points. I think the first is, as I indicated, and Chris and also Sandeep talked about, I mean the design win momentum in custom has been very, very strong. even since the AI day, I feel really good about that $75 billion pipeline that we're really bringing to close some of those key opportunities within that. I think that pipeline, by the way, from what we can see is probably going to just keep growing. And by the way, this is across XPU, XPU attach. It's at the large hyperscalers and then continuing to increase around the emerging. Optics continues to be very strong. We're managing the execution quite well and growing the business there. And then the core business, which was a point of consternation in the past about when would that come back and what would that ramp look like? It's nice to see in Q3, the strong sequential in enterprise networking and carrier I think it's like 30% sequential and 80-plus percent year-over-year.

So very strong recovery there. And finally, it's showing up in the numbers. I mean, we're getting a lot of leverage here looked at Q2 EPS, it's up like 123% year-over-year. In Q3, at the -- if you look at the guide, EPS would be up like 70%, so much faster than revenue. Same thing on the sequential. So overall, we're very pleased with the performance of the company. We see a massive opportunity ahead, and I appreciate everybody's interest in Marvell, and we'll talk to you all soon. Thank you so much.

Ladies and gentlemen, thank you for your participation. This does conclude today's conference. Please disconnect your lines, and have a wonderful day.

Welcome to Marvell's custom AI Investor Event. Please welcome Senior Vice President, Investor Relations, Ashish Saran.

Good morning, and welcome to Marvell's custom silicon investor event. I'd like to draw your attention to our forward-looking statements. As a reminder, this presentation contains projections and other forward-looking statements regarding future events and financial performance of the company. Such statements are predictions and subject to significant risks and uncertainties, which could cause our actual results to differ materially. Please consider the risk factors in our SEC filings, which could potentially affect our business and financial performance. These filings are available from the SEC and on our website. During this presentation, we may be mentioning certain non-GAAP financial measures. A reconciliation to GAAP is available on our website.

Let me walk you through our plan for today. While Marvell has 5 end markets, today, we will focus exclusively on the custom silicon opportunity in our data center end market. Matt will kick things off with an update on our custom silicon progress since our AI Day last year. Chris is going to focus on the key factors driving demand for custom silicon. Nick will discuss our full-service custom cloud platform, which is driving our success. Sandeep will do a deep dive into our technology platform. will bring us home with how his team is engaged and winning in custom silicon. We will conclude our event with a Q&A session.

With that, I would like to invite Matt Murphy, Marvell's Chairman and CEO on to the stage. Matt?

Please welcome Chairman and CEO, Matt Murphy.

Awesome. Thank you. All right. Good morning, everyone. It's great to see all of you. Welcome to Marvell's 2025 Custom AI Investor Event. And to our audience joining via webcast, welcome and thank you for being with us. And for those of you here in the room, Marvell's most senior distinguished engineers and fellows. Thank you for the incredible work you're doing to drive this business and this company forward. Okay. So let's get started.

What if I were to tell you that there's a revolution happening inside cloud data centers. And it's all around the silicon in which the data infrastructure is built. We call it cloud optimized silicon. Now if that sounds familiar, it's because I stood on this stage almost 4 years ago and talked about exactly where this was headed. Check out the data on this thing. This was my first slide actually from our 2021 Investor Day. It's truly incredible, isn't that how this has played out. Custom silicon has become one of the largest growth drivers in the entire semiconductor industry.

Here's another slide from that event. Back then, we said that emerging applications weren't going to work running software solely on x86 processors. Do you believe that was a debate back then New workloads we said required new compute like GPUs, and that cloud providers would start customizing their machine learning ships, their CPUs, their DPUs and ultimately, the entire cloud infrastructure. Now at Marvell, we've been on a mission since I became CEO in 2016 to lead in hyperscale data center infrastructure. And in 2018, my team and I made the decision that the future of cloud was going to be custom, and we needed to build a strategy and a team to lead. And you're going to hear from many of those leaders here today.

So in 2021, we laid out the strategy for you, including our initial set of design wins. And at that time, if you remember, there was a lot of debate in the industry on whether custom silicon could actually take off and ramp at scale. I don't think anyone questions that anymore. Custom is happening. It's happening in every cloud, and it's here today. Now look, we all know we've come a long way, and it certainly has not been easy, but we've been investing for nearly a decade to get to where we are today. So if you're just waking up now and you want to be in the business of building custom silicon for the cloud, you're too late. So why has this trend accelerated over the past few years?

Back in 2023, the top 4 U.S. hyperscalers were spending about $150 billion in CapEx, which was already a huge amount back then. But then it grew to over $200 billion in 2024 and now over $300 billion in 2025. It's an incredible level of investment. And if you think about that CapEx, a huge portion of it actually goes to the silicon. And that's why these were the first companies, the big 4 to first customize silicon for their data centers. They clearly saw the future back then on where this was heading and how they could benefit from this trend. And it just made sense to build optimized solutions for their individual use cases from top to bottom.

But it's not just the top 4 anymore. We see a whole new wave of companies investing in their own data infrastructure, and we call them emerging hyperscalers. Now the first group are companies building the foundational models, and they've realized that the value of controlling their own infrastructure and that they're building their own data centers and beginning to build their own data centers. Now take xAI, for example. They built a 200,000-unit AI cluster in just 1 year, and they've already produced a very powerful model in Grok. And then there are those that are building the end applications. They're also building highly specialized infrastructure for AI. For example, Tesla built its own Dojo based data center to power the AI behind full self-driving. And then there's media coverage suggesting many other companies are heading in the same direction.

And recently, we're seeing the rise of what's called sovereign AI, which is nations around the world who are also launching major investments to build local AI infrastructure. So all of this is driving even more demand requires more innovation, and it creates more opportunity for Marvell. All right. So let's go back to the CapEx numbers. So the top 4 U.S. hyperscalers have grown, they're spending at a 46% CAGR over this period. But if you zoom out and you bring it -- and you look at total data center CapEx, you see it's growing even faster at a 51% CAGR. And that's because historically, the rest of the data center CapEx was coming from Tier 2 cloud providers and on-prem data center applications. But more recently, we're seeing these emerging hyperscalers contribute at an increasingly significant rate.

So if you fast forward to 2028, analysts are forecasting data center CapEx exceeding $1 trillion. And who knows? Today, they're saying $1 trillion, but it could be more. So where is the lion's share of that spending going to be in that time frame? Clearly, these 4 hyperscalers are not slowing down anytime soon. But I wouldn't be surprised if the emerging hyperscalers grow to be a significant portion of themselves. So either way, it's clear from the trend line and what we see from our customers that both are going to grow substantially in this time frame.

So what does all this mean for Marvell? At our AI event last year, we outlined a $75 billion TAM, and it was growing at an almost 30% CAGR across custom silicon, switching, interconnect and storage. So we stand here a year later, everything has gotten bigger. Our overall estimate has grown by about 25%, and we're now seeing forecasts for a $94 billion TAM in 2028. If you look underneath that, the 2 fastest-growing markets last year have grown even more. So compute is almost 30% larger than what we projected last year, and interconnect is up about 37%. So both of these are right in Marvell's wheelhouse and they're key focus areas for us.

So now Marvell's total data center opportunities, $94 billion, growing at a 35% CAGR. Custom compute is the largest and fastest-growing portion, followed by interconnect, then switching and then storage. At today's event, we're going to focus on custom compute. And over the past year, this is where we've seen the biggest change, both in terms of the size and diversity of the opportunity. And at the same time, we continue to drive strong execution across interconnect, switching and storage. But compute remains by far the largest incremental opportunity in front of all of us. And that's why we've chosen to focus our today's event entirely on custom. We've made tremendous progress in the portfolio in the past year, and you'll hear from our team today, they're going to walk you through what's happening inside this market and why Marvell is winning.

So first, let's take a look at what's included in the market. Clearly, they are the XPUs themselves. And this is the biggest part of the compute TAM. They're the largest and most complex chips in the world. And the technology required to compete in this market continues to accelerate at an astonishing rate. You'll hear more from my team today on what we're doing to win in this market. What's also become clear over the past year is that the number of XPU opportunities continues to expand, both within the top 4 hyperscalers and also with the emerging hyperscalers we talked about. And we're going to spend more time on this exciting set of opportunities in a few minutes. But first, let's take a look at what else is in the TAM.

These AI compute platforms are comprised of more than just XPUs. Modern AI infrastructure requires complete systems packed with silicon to run AI workloads at scale. So within these platforms, there's a multitude of companion chips that help support and scale the XPUs. And that's what we call XPU attach. And just to be clear, XPU attach is all custom silicon. This is independent and distinct from the other product areas of interconnect switching and the other parts of our broader data center opportunity. By the way every platform

And that's what we call XPU attach. And just to be clear, XPU attach is all custom silicon. This is independent and distinct from the other product areas of interconnect, switching and the other parts of our broader data center opportunity. By the way, every platform is different. There are common elements like network interface controllers or NICs, power management ICs, the scale-up fabric, just to name a few. In other cases, there's other customized solutions, specialized coprocessors for security and other functions or memory or storage poolers and expanders. What we're seeing is an explosion of different sockets inside these AI systems, but they're all attaching to the XPUs. This represents an incremental custom opportunity for Marvell on top of what we discussed last year. So if you turn back to the $55 billion in TAM, that breaks down to about $40 billion in XPU, and that's growing at a 47% CAGR and $15 billion in XPU attach, which is growing at an incredible 90% CAGR, nearly doubling every year from a pretty small base in 2023, but still doubling. And we'll explain why that's growing so fast. But essentially, it's because of the increasing complexity of these custom systems. And if you just put this in context, when you look out to 2028, the custom XPU attach market is of the same magnitude of the entire custom silicon market today for the cloud. So both XPU and XPU attach are incredibly important parts of this market, and we're going to spend a lot of time today walking you through the dynamics in each and how we're positioned to win. All right. So let me update you now on our growing customer traction in this market. So the goal here is to win what we call sockets. Now a socket represents a multigenerational opportunity inside a customer's architecture. And as the architecture evolves from generation to generation, these sockets tend to sustain. And so once you've won the socket, as long as you do a good job and you execute and you deliver, you're in the pole position for the next generation. Now last year, we talked to you about these 3 custom compute sockets at 3 different U.S. hyperscalers, and they're all on track. Two, we've taken to production and they're driving substantial revenue for us today. The third is well into its development, and it's also on track, and we're fully engaged on the next generation on all 3. But here's the interesting thing. We've also won 9 additional custom sockets at the top 4 hyperscalers. Some of these we had already won last year and some are brand new this year. And this initial wave is in production today, but the remaining are in design execution and they're set to deliver revenue over the next couple of years. These are the XPU attached sockets. So as I mentioned before, as architectures evolve, sockets tend to sustain. But when the architecture changes and it's changing rapidly, new sockets emerge that weren't part of the prior architecture.

And that's what we're seeing in the custom AI infrastructure now, a rapidly evolving architecture with more custom sockets each generation. This means more shots on goal to the Marvell team, and we are winning. So that gives us 12 total custom sockets with the U.S. hyperscalers, the 3 we talked about last year, up from the 3 we talked about last year. But that's just at the top 4. Now remember, I talked about the emerging class of hyperscalers. We've been very active in this segment as well. And I'm excited to announce to all of you today that we have won 2 XPUs and 4 XPU attach sockets in the emerging category. These are either in production or on their way to production today. And this is going to continue to build over time, and it's an incredibly active part of the market.

So when you add it all up, Marvell's won 18 different sockets in the custom compute market. Those 18 give us line of sight to achieve our market share goals. If you take a step back, there's been this perception that the entire custom silicon market for data center really comes down to just a handful of sockets. And to be fair, that used to be true. If I look back to 2023, the largest single socket probably made up or more of the TAM, 1 socket. But that's clearly no longer the case. And by the time we get to 2028, I would expect that no individual socket is going to be more than 10% or 15% of the TAM. And that's because we're seeing more XPU opportunities at existing and emerging hyperscalers and then the XPU attach opportunities are expanding even faster as these architectures evolve.

In Marvell, we've won 18 of them already, but we're not stopping there. The pipeline of opportunities in front of us continues to grow rapidly with an incredible number of active engagements. In fact, today, we are tracking more than 50 additional opportunities in our pipeline. More than 50. And again, it's not just driven by 1 or 2 customers. There are over 10 different customers now that we're engaged with across this range of opportunities. So if you look at the magnitude of this and you add it all up, there's a $75 billion lifetime revenue potential for Marvell hanging out there in front of us. And these opportunities represent growth above and beyond the ones we've already won [indiscernible] So this is the future.

Now some of these could turn into revenue by 2028, which then would be incremental to our plan. And let me just put the $75 billion in context to help investors on the line think through what this means. And so I'll start with the opportunities first. So when you break down the opportunities, the [ 50 ], about 1/3 are XPUs and about 2/3 are XPU attach in terms of count. Now these XPU programs are monsters. They're typically multibillion dollars in lifetime revenue each, and they run between an 18- and 24-month period in terms of the lifespan of the program. So that's what we refer to by lifetime revenue. You just take the total revenue over the lifetime of the project. Now when you look at the XPU attach programs, these are also very significant. They are incredible opportunities. These are in the several hundred million dollars in lifetime revenue, but they span over a 2- to 4-year period. And so for the financial community that's on the call, let me just give you some context on how you might want to think about this.

Let's just take the 18 programs we talked about, 5 XPUs, 13 are XPU attach. And now we've given you a range for the lifetime of the programs and the scale of the revenue that's potential. So you're all smart on the line. I'll let you guys go do the math, and you can actually go and build a model and start to imagine what this could look like. But you could easily see how we could get to the revenue scale we've been discussing. And if you take our 18 existing programs and you think about the 50-plus opportunities we're chasing, you can start to imagine like what does that opportunity look like long term.

Okay. So why are we winning? How do we get to this place where we now have 18 different sockets that we're ramping into production. And that's what my team is going to spend the day talking to you about. We're going to explain what we do for these customers and why Marvell is so well positioned. But fundamentally, it comes down to this. We are unique as an end-to-end full service custom silicon provider. What we do is bring together the best system architecture, design IP silicon services, packaging expertise and full manufacturing and logistics support to enable our customers to realize their silicon ambitions. This means the customer does not need to cobble together IP from a variety of third parties and then go off and hire a design house to complement their in-house team and then find another vendor to go manage their supply chain.

And while this may have worked in the past from time to time, it simply is not going to work in the future as the technology landscape accelerates. You need pretested, pre-integrated IP and architecture support coupled with best-in-class design. That's what we do. And the underlying technology platform we've established at Marvell is truly astonishing. It's what allows us to continue to win these multigenerational programs. We have a proven track record of delivering on leading-edge process nodes. And today, we're in volume production in both 5-nanometer and 3-nanometer, but we're not stopping there. We already have test chips on 2-nanometer, which will enable our first 2-nanometer products. And as we are leading the charge into the angstrom era with development already on A16 and A14 nodes for future products. And you'll hear from Sandeep and his team on how we're executing there.

At the IP level, we built one of the broadest portfolios of analog mixed signal IP in the industry, focused on high-performance, low-power, low-latency SerDes along with our high bandwidth die-to-die integration. And just to give you 1 example, Marvell demonstrated the world's first 448 gig SerDes running at OFC a few months ago. And this kind of capability is critical, both for scale-up and for scale-out networks. And then finally, advanced packaging. This has become just as important as the silicon itself. It's not an afterthought anymore. It's fundamental to enabling AI technology at scale. But don't just take it from me.

Let's hear from 2 of the largest cloud computing companies in the world and 2 of Marvell's most important partners. We built an incredible partnership with AWS, and Marvell has been a thought leader on implementing EDA in the cloud for our products. It's helped us scale our design capability and move at hyperscale speed for our customers. And then we're building chips for their cloud, which is helping them scale. And you can see a quote here from AWS CEO, Matt Garman, which really -- it speaks to the strength of the partnership and the joint work that our teams have been doing together.

All right. Next, I'd like to introduce Ronnie Borkar, who has graciously agreed to speak with all of you today about our partnership. Ronnie is Corporate Vice President of Azure Hardware Systems and Infrastructure at Microsoft. So let's turn it over to Ronnie.

Thanks, Matt. It's great to be here. We are in a pivotal moment in our industry. And it's partnerships like ours that enable us to innovate boldly and deliver at scale. At Microsoft, we are architecting our entire cloud stack to deliver the highest performance, lowest cost and most secure infrastructure. And as the demand for compute and model innovation outpaces the traditional rate of hardware innovation. We are reimagining every layer of the stack to meet the needs of next-generation cloud and AI workloads. This requires an end-to-end approach, optimizing across the data center system silicon and the serving and application layers.

By doing so, we are able to bend the cost curve and accelerate innovation, unlocking compounding gains that go beyond what's possible in any 1 layer. Sitting at the foundation of our stack is silicon, where deep hardware, software core design is what truly unlocks breakthrough capabilities. By combining Microsoft's software expertise with purpose-built silicon, we deliver the performance and efficiency needed to scale compute and AI infrastructure for the next wave of transformative applications. But we don't do this alone. These gains, these breakthroughs are only possible through deep, long-standing partnerships across the ecosystem.

For more than a decade, Microsoft has partnered with Marvell as we advance our infrastructure. As a leader in silicon, this includes various aspects of our custom silicon journey, and we benefit from Marvell's ongoing technology innovation. As we look ahead to the next wave of cloud and AI innovation, we are excited to continue our relationship with Marvell as a trusted partner on this journey.

Wow. Very cool. Awesome. Awesome. Thank you, Ronnie. I think you actually -- she captured it really well. I mean how companies are working like Microsoft to optimize their infrastructure from top to bottom.

Okay. So let's dive back into Marvell's data center business. And today, I want to help investors get a better sense of all the moving pieces inside the data center end market. So last year, we did about $4.2 billion in data center revenue and about $500 million of that, you can see it on the bottom there, came from on-prem data centers. Now that part of our business has been for around for a while, and it's pretty stable. And so going forward, for modeling purposes, you want to think about that business generally being in that ZIP code going forward, which means the rest of our data center revenue is now coming from cloud and AI.

Now if you go back a few years when ChatGPT launched, we went through this exercise to break out AI as a separate category. And at that time, it was new and it was emerging, and it was really just tied to a handful of programs. So after some work, we were able to get a pretty good read on that revenue. And we made some projections which turned out to be actually wrong. But the good news is they were wrong in a good way because we've vast exceeded them from just a few years ago. But as I sit here today, we've seen this just dramatic transition where AI is now in everything. I mean, all the applications we use every day and the tools that we use, AI is embedded, and it's touching every aspect of technology including everything in the cloud. And now most of the applications running in the cloud are using AI. And all the cloud infrastructure is really becoming an AI factory. And so going forward, when I look at the investment and the CapEx, all of it's moved to AI. And so from our perspective, our cloud revenue will be AI revenue going forward. It's just all combined.

Now let's take a deeper dive inside that AI business. So in our last earnings call, we said that in Q4 of fiscal '25, about 25% of data center revenue was custom. So if you just take out the on-prem piece, it would actually be greater than 25% of the cloud AI revenue. And with the wins we have in place, we expect that custom will grow to greater than 50% of Cloud AI revenue over time. And based on the size of the opportunity, it probably just keeps going from there. So if you go back to the $55 billion custom compute TAM, Marvell had very little share back in 2023. It was less than 5% in our custom silicon programs at that time were just getting started.

Now going forward with the 18 sockets that we talked about, we're targeting 20% share by 2028, and we're actually well on track to achieve it. And in fact, with the 5 XPU sockets we've won, I would expect that we could achieve about 20% of the $40 billion market. And even though it's been -- it started a little bit later and it's newer, I actually don't see any reason why we can achieve 20% share on the $15 billion XPU attach market in that same time frame. So we've already won 13 -- because we've already won 13 sockets in that portion of the market. And our revenue and XPU attach by the way, from last year to this year is doubling, and it looks like it's going to keep doubling going forward. So if you put it all together, we're well on track to achieve our 20% share.

Now let's take that $55 billion TAM, okay? And then we're going to add it back to our total data center TAM, see it there. Okay. Now let me tell you about our progress. Last year, we had said that we had about 10% share of a $21 billion market in calendar '23, and we're targeting 20% share in the 2028 time frame for data center. In 2024, the market actually grew almost 60%, but our business at Marvell almost doubled. So if you take the $4.2 billion we did last year in data center revenue, we actually had about 13% of the market. So -- and obviously, this year, we're still growing quite strongly. So while we expect to grow continuously, we actually see even a larger step-up coming in calendar '27 as several new programs, including a major XPU socket, hit mass production. So this all positions us very well to achieve our 20% target by 2028 and on a market that's now grown to almost $95 billion. So let me just take a moment to summarize everything we've talked about this morning.

Last year, we told you we had a $75 billion opportunity in 2028. This year, we're saying a balloon to $94 billion. Last year, we talked about the custom XPU opportunity. This year, we unveiled an entirely new market for custom XPU attach. Last year, we said we had 10% share driving to 20%. This year, we showed we're already at 13% in 2024 and growing strongly again in 2025. Last year, we said we had these 3 existing custom XPU sockets. And this year, we revealed and talked about 18 multigenerational sockets. Last year, we talked just about the opportunity at the top 4 U.S. hyperscalers. This year, we see a brand-new customer set with emerging hyperscalers. And on top of that, we're actively pursuing over 50 additional opportunities worth $75 billion in potential lifetime revenue for Marvell. That's the opportunity in front of us.

Now I'm going to hand it over to my very talented leadership team. And once you hear from all of them, you'll see why I'm so excited and why I'm so confident that we can make this happen. Thank you very much.

Welcome to Operating Officer, Chris Koopmans.

All right. Good morning, everyone. I'm excited to be here this morning and talk to you all about our custom silicon opportunity. And I'd like to start out by taking a step 1 level deeper into that market and trying to talk about what's driving that move towards customization.

So let's start with that $1 trillion that Matt outlined at the beginning of his presentation. Now this is analysts forecast for total data center CapEx in 2028. And that's really based on a fairly modest growth rate from here of only 20% per year. So if we start with that $1 trillion and we go 1 level down and take out the physical infrastructure, that's about $800 billion in equipment. And inside that is silicon, about $500 billion in semiconductors. $500 billion, that's bigger than the entire global semiconductor market in 2020. It's half of the total CapEx is being spent on silicon. And the largest part of that is the accelerated computing TAM. And that's at about $350 billion.

Now that $350 million is based on the current analyst forecasts and some folks think it's going to be even larger than that. And that's fair. A year ago, at our AI event, we said [indiscernible] be $172 billion. So it's actually -- and that's based on analyst forecast at the time. So it's actually doubled just in the last year. So clearly, there's a lot of room for growth on this accelerated computing TAM. But if we take that $350 billion and we break it down into its component parts, we see that a large portion of it is the XPU itself. Think of this as the XPU tile. And there's a big attach market as well, as Matt outlined, the biggest part of the attach is actually the HBM or the high bandwidth memory that's packaged inside the module. And the second part is the XPU attach that Matt outlined, which is the other networking coprocessors, network interfaces, memory and storage interfaces controllers. So if you put it all together, there's a very large TAM made up of these component parts.

Now remember, this is the total accelerated computing TAM. At Marvell, we focus only on the custom portion. And for custom, when we work with the very largest hyperscalers in the world, what they tell us is that they want to buy the HBM directly. They have a very large relationship with these memory vendors, and they buy a lot of DRAM today, and so they want to purchase it directly. Now emerging hyperscalers, they might want to deliver -- get a turnkey service from Marvell, and we can deliver that for them. But just to be conservative, we're going to take that out of our TAM, leaving us with a $220 billion TAM.

Now that's just the DRAM that I pulled out of the TAM. Ultimately, if we deliver a custom HCM solution that we'll talk about later today, and we're building the bottom die with our proprietary diode interfaces, that will be part of the XPU attach. So if we take this $220 billion TAM and then ask ourselves, what percentage of it is going to be custom by 2028. So last year, we said we thought about 1/4 of the market would get to be custom. And based on the progress over the last year, and the opportunities we see in front of us, we're well on the way to achieving that and ultimately potentially exceeding that number. So this is where we get the $55 billion custom compute TAM that Matt talked about. And you can see that it's growing very fast as customization is increasingly taking on the market. And XPU is growing almost 50% a year, the custom XPO, I should say.

But the custom XPU attach is almost doubling every year. So this really leaves us with a question of why? why the move towards customization? And why would the custom XPU attach even go faster than the custom excuse themselves. So the answer lies in the workloads. No longer do we just have a simple singular AI workload where we're just training a large model? The workloads have now diversified significantly. You have pre-chaining and post training. You have a whole variety of inference workloads. And as the workloads diversify, specialization is on the rise. You build specialized infrastructure to have superior total cost of ownership and superior performance for those diverse workloads. And that in turn drives customization. So let's look at those workloads.

First, we absolutely have the original workload of building the largest model possible. And we see every company out there racing to build the biggest possible cluster to train the biggest possible model with the largest amount of data. And you can see what progress has been made over the past couple of years, and this trend has emerged of a race to be building the first 1 million XPU cluster. So that's one of the critical workloads that's driving this infrastructure trend in the market. But we've also seen now inference emerge as one of the most important applications in AI. And inference is also diversified. You have the traditional large language inference, which is chat interfaces, content writing, search. This is a fairly compute light but memory-intensive workload. It's memory intensive because you need to be able to access the entire model for every query, but you're not doing that much computation on each query.

On the other side, we have chain of thought based reasoning inference. This is a much different application, much more compute-intensive as you're recursively asking the infrastructure to produce a better and better result. It allows you to produce much more complex answers, deep research, solve complex puzzles and problems but it's much more compute-intensive. So the trend is clear. If you're in the business of building an AI factory, one size does not fit all. Ultimately, you need to be able to have clusters that are very large with heavy with compute to train the biggest possible models. You also need to be able to have millions of instances of smaller clusters with less compute to be able to answer inference queries, and you need to have a chain of thought optimized infrastructure to be able to deliver the best performance in that particular part of the market.

So as workloads in AI diversify, so too must be infrastructure -- and we're seeing it today. We've seen the largest hyperscalers to deploy huge clusters of GPUs. You've also seen them deploy very large custom XPU fleets. In some cases, we've seen them deploy performance optimize large cluster custom XPU fleets. And we've seen them deploy separate efficiency optimized smaller cluster custom fleets. We've even seen emerging hyperscalers realize the value of deploying optimized inference infrastructure for their model. And the data set matters. If you're training your model based on millions of hours of video for self-driving, that's a very different workload than if you're training it on billions of tweets, and so you've seen companies build separate custom infrastructure for those 2 types of training modes. So it's pretty clear that diversification is driving specialization of the infrastructure, which is driving customization of the infrastructure.

Let's take a quick look at what that means, and let's start with the XPU itself. Now I get a lot of questions about this. What is the custom XPU? Is it a hard-coded ASIC. The answer is no. Our customers have designed these as highly programmable processors. They all have a multitude of compute cores on chip SRAM and then IO interfaces to the outside world. So where is the customization come in? Starts in the compute cores themselves. Do you have more matrix math optimized cores or more scaler and floating point courses? What level of precision have you optimized it for? What's the ratio of cores to memory. It's really important to get this right because if you end up with the wrong ratio, you end up with idle compute waiting for memory or you end up with idle memory or empty memory while the compute constraint kicks in.

So having the right ratio of these 2 is critical to building an application optimized XPU, and that continues in the module. So again, getting the on-chip SRAM versus the in module HBM memory right for your application is critically important and can really affect the performance and the total cost of ownership. And that continues in the interfaces. This is where we build out into the fabric. Is it a direct attach to copper fabric? Is it a co-package copper or a co-packaged optics fabric. And as you customize all of this, you build a scale up logical XPU that looks like 1 XPU to the software. And even in this architecture, we see specialization. Is that fabric in any to any star fabric or a nearest neighbor connected tourist fabric. Ultimately, depending on the workload and the application, a specialized infrastructure makes the difference.

So let's take a step back. If you're a hyperscaler or an emerging hyperscaler today, and you've made the decision to customize your XPO. The next question is, what platform do you deploy it in? On the left, we put the general-purpose platform. This is where everybody starts, deployed in a platform that was originally designed for racks of servers. So when you deploy it there and the emerging -- the hyperscalers started there, you quickly find out that it becomes a bottleneck. You can't really unlock the value of your XPU in a general-purpose platform. So the largest hyperscalers have already built full custom platforms to house their custom XPUs, optimized for their applications. Recently, we've seen new ideas emerge.

So for example, a third party accelerated infrastructure optimized platform available for others to drop in their custom XPUs or it will a standards-based accelerated infrastructure platform emerge, just in the last couple of months, Marvell has made a few announcements in this area. At Computex last month, we announced with NVIDIA and our ability to help our customers build an XPU and integrate NVLink fusion, which would allow them to deploy within NVIDIA platforms. Separately, we've announced our UA Link platform, which would drive a move towards a standards-based accelerated infrastructure platform.

Whatever the case, it's clear that the platform needs to be optimized for accelerated infrastructure, which is driving customization. And so as you go from the left-hand side towards the right-hand side, the platform becomes increasingly customized. And that's what's really driving the move towards custom XPUs, optimized for the different workloads and the even larger increase in custom platform as it got started later, but it's moving very fast. And in Marvell, with our 18 current sockets, we're well on our way to achieving our 20% share in this market. And with the 50 opportunities that we're driving, we're just getting started. Thank you.

Welcome Senior Vice President and General Manager, Cloud Platform. [ Nick Kochoreski ].

Good morning. So thank you all for being here today. I've been building semiconductors for hyperscale data center applications since 2007. It's been about 18 years. But what we're seeing today is really unprecedented in terms of the market opportunity for advanced semiconductor design. But also the game has changed in terms of what's required to build a leading product for this market. We're seeing all new levels of vertical integration system innovation, a lot of custom silicon and the pace of component new technology is moving faster than ever. This requires a different category of company to participate in AI silicon.

Marvell has built that company. We have the team and we have the product strategy for this next generation of custom AI. We have 30 years of experience building high-performance ASICs with a track record of doing it right the first time. We have built a comprehensive portfolio of technologies, specifically for data center and AI applications, many of which we've developed over multiple generations. And we have built an engagement model specifically designed to work with hyperscale customers to manage the entire life cycle of the design process and that's what you're going to hear about over the continuation of today's section, the Marvell custom cloud platform. Now in my talk, you'll hear about our product strategy and our engagement model. And then you'll hear from Sandeep and the engineering leadership who is building these leading portfolio technologies and bringing our customers solutions to market. And from Will, you hear more about the opportunities and why our customers choose to work with Marvell. Now there's going to be a lot of information here, and we wanted to give you a framework for how to think about this. It really comes down to 1 very simple question. What do our customers look for in choosing a custom silicon partner.

For our customers, this question is becoming more critical with every passing day. You see the customer's final product, which is the AI data center itself, it is a composite of silicon system and software technologies implemented in the right combination and then instantiated thousands of times across the full AI deployment. Inside every chip is a long list of key enabling technologies, which enable those next-generation systems to reach the highest level of performance. But each one of those technology elements are innovating continuously. And the key is to grab and choose the right pieces in the right combination for that next-generation product. But as Chris mentioned, the applications in AI are evolving so fast that there is an almost continuous demand for new system-level approaches to achieve the next level of optimization or customization for those new workloads. And all of this is happening at industrial scale. That means exaflops of compute performance, petabits per second of networking bandwidth, gigawatts of power.

When you're talking at that scale, 1 day of delay in bringing our product to market can mean millions of dollars in lost revenue. So the stakes are higher than ever before, and it's critical to work with the right silicon partner. Now in the world of custom silicon, not all companies are the same. In fact, it seems like every month, I hear about a new company that's announced their intentions to get involved in custom silicon, put the sign on the building and announce that you're ready to do custom chips. It's not that simple. Marvell's engagement model is based on a full life cycle engagement that anticipates the needs of the market multiple years before design even begins. We build the core IP and the design methodology well in advance.

And we go through a process with the customer to choose the right technologies. And we work with customers on their product designs that integrate the latest leading technology for silicon package systems and manufacturing. But this is not the only way to do it. In fact, there are other models within the industry for building custom silicon. In the second row, you see physical design services. Now this is what most people would refer to as an ASIC provider. It's a business model that's been around since the early 1980s.

Now in this case, the custom partner may participate in a portion of the design process and then handle the manufacturing, but largely, the customer is doing all the heavy lifting. They have to find all of the key different IPs. They have to make them sure that they integrate together and then they take on a lot of the work in terms of making sure that's ready for volume manufacturing. And there is also an option out there for the do-it-yourselfers. This is known as COT or customer-owned tool set. In this case, the end customer is responsible for procuring all of the IP, doing all of the design and designing their own design methodology in-house and ensuring that everything comes together in combination. And then the custom partner is just handling the manufacturing.

So the core technology, our design methodology, our manufacturing standards and our logistics systems have been developed over years by some of the most talented engineers in the industry. The technologies and services we provide can't be found just by opening the yellow pages or looking on Yelp for an ASIC services partner. It's unique and differentiated IP that's not generally available on the open market. And this is what customers look for when they want to reduce risk and go to market on time. So I'll be talking about each one of these elements in terms of what Marvell provides, which is different and differentiated within the market.

So first, talk about the system architecture. Marvell starts years ahead to see what technologies will be available for building high-performance silicon, multiple years ahead of the project. We work with our customers interactively to assess those technologies and to explore what kind of systems can be built with them. This enables our customers to build new system concepts and to define the silicon components based on the most advanced technology with a full view of the system. Now from there, we work cooperatively with our customers to do interactive system definition. We look at power consumption performance, cost trade-offs so that the product and the component silicon is truly market leading. Now as you'll see through our sessions today, the key to building these high-performance AI cloud silicon products is the enabling technology, the key design IP.

Now often, the first thing we talk about is the SerDes. And that's for a good reason. SerDes is the single most differentiated element in building these products. But this is 1 of 10 or 20 different core IPs that go into building these advanced chips. Everything from high-performance die to die to digital cores for compute and networking, integrated silicon photonics, high-bandwidth memory, on-chip memory, software and firmware. All of these elements come together to build the highest performance product for the next generation. Marvell builds these technologies in-house. So our customers see a one-stop shop for many of the technologies they need to build their products. But not only that, we work closely with the ecosystem. We collaborate with start-ups to incubate new technologies and bring them to market and then integrate them into our product platform.

We also work with some of the largest IP providers in the industry to ensure that our customers have a very broad portfolio with a choice of all the elements they need to build their products. Next, we'll talk about the silicon design itself. Now here, you often hear about process technology, 5-nanometer, 3-nanometer and so on. And that's definitely critical. But the key point to note is that there's a lot of preparation that goes into every one of those process technology transitions. For example, all of the IP I talked about on the previous slide has to be brought to be supported in that next-generation process generation.

In some cases, we build test silicon to characterize the performance of those IPs. Based on that data, we'll create models so that customers can use the models to build their products. Sometimes, we even have to evolve the ADA workflow. We work with our partners to ensure the tool set used to build the chip is updated to reflect the latest technologies and the latest generation of the process node. Not every services provider does this. So that those that do like Marvell are years ahead of the curve in being ready for next-generation designs and working with customers and moving very quickly.

Next, when we talk about design execution, we focus on the logic design and the physical design, but then we go beyond that. It's also about design for testability and design for manufacturing. We introduced new innovations for DFT and DFM, which translates into higher quality and better cost efficiency. And we also innovate on power consumption. With every new process generation, there is an opportunity to do new innovations, which are unique to that process for how you get the most in terms of the lowest power and the best efficiency. And we also focus on yield enhancement. When you're talking about very large products built for AI all of those innovations that go into optimizing for yield translates into superior cost at scale.

Now as you'll see today, packaging is one of the most critical innovations in building high-performance AI silicon. In fact, this is the key to scaling beyond Moore's Law. And we have innovated on multiple dimensions to enable the latest generation of packaging technology for these AI products. We have a multi-die platform to enable the advanced integration of multiple silicon components. This is based on a Marvell-developed approach, and we also work with TSMC Coos. So customers have the option for how they build multi-die packages.

Now for high-bandwidth memory, which means integrating a large number of DRAM right into the package, Marvell has its own platform solution, and we also support industry standard platform. So here also, we provide customers with the flexibility depending on the approach they want to take. But it's not just about packaging for high integration, it's about pulling system innovation into the package, allowing for better power consumption, lower cost and better performance through system-level innovation right in the package. One really good example of this is co-packaged copper.

Now in the past, high-speed signaling was done at the system level through the printed circuit board as we've been doing this for decades. But more recently, we've actually pulled those high-speed signals out and above the package, and we run thousands of wires directly out of the package across the system. This gives you longer reach at lower power [indiscernible] and can allow you to achieve higher performance levels in the context of an AI rack. And then going beyond copper, Marvell has announced its platform for co-packaged optics, bringing fiber optic signaling right into the package, which can give you much longer reach than copper at very high bandwidth at lower latency than you could do before. This is a key enabling technology for AI scale-up systems and those platforms, which Chris was talking about.

Now all of these innovations work together to enable products with more functionality and performance. And it gives you better generational performance than ever thought possible. Now once the product is defined, designed and implemented in the silicon and the packaging, we move to volume manufacturing. Here, too, our customers need to work with a top partner. Marvell has the scale to deliver. In fact, we're one of the largest semiconductor suppliers in the world. And we have decades of experience delivering large-scale semiconductors, logging billions of device hours in the field. And we've developed a methodology that ensures we're ready to move to volume production on time and in line with customer schedule and volume expectations.

Now this is based on a multiphase operations implementation that runs in parallel with the design team, covering product engineering, test engineering, quality assurance and reliability. This operations engineering team works in parallel with the chip design team and to make sure that we are ready to go to high-volume manufacturing when the customer is. So now you see looking at the big picture, it's clear why working with a full-service custom partner offers major advantages. It's about having access to the latest technology that is not even available in the open market and then discussing a plan for those technologies with the customer years in advance so they can define their systems based on what's going to be available out in the future. And then it's building products with the newest packaging, the latest silicon technology and using a very broad portfolio of that IP. We have built the capability for custom cloud silicon that is best in class.

Now over time, this will become even more critical as AI trends continue to accelerate. Now when you consider all of the different technology components that go into AI semiconductors, each one of those components is refreshing approximately once every 2 years and you're getting an improvement in speed on each one of these components. But when you look at all those parts coming together in components and then systems, the net result is a 10x increase in performance every 2 years. That's what we see as possible moving forward through the confluence of all these technologies.

So these -- as these AI applications diversify and customization becomes critical, -- it's really important that you work with a full-service partner who can provide that kind of breadth of offering. The customer do-it-yourself model is possible, but it will increasingly be challenged. And it poses risks to the customer where working with a partner like us can really improve their odds. So now as you look at the presentations today, keep in mind the 3 most critical points for our customers when choosing a custom partner. One is a partnership to select the right enabling technology years in advance of the product design. Second is access to an exclusive set of design IP that is not generally available in the open market; and three, methodology and experience to enable total product quality manufactured at volume scale, plus the ability to do this across the full spectrum of AI cloud applications. This is the real advantage that our customers experience when they work with Marvell. Thanks very much for your time today.

Welcome, Chief Development Officer, Sandeep Bharathi

Good morning, everyone. I'm very excited to share our journey of technology and engineering leadership. First, before I get started, I wanted to share the journey of how I got to Marvell. In 2018, Matt was interviewing me and he threw down the gauntlet. At the time, Marvell was in 28- and 16-nanometer, and he asked me, how do we get to leadership? And the answer was to say, how do we get to leadership. And the answer was to skip all those nodes and go straight to 5, but that was easier said than done. Today, I will share with you the journey of how we got here and how we got to industry leadership, and you will hear from many of our senior technical leadership across multiple domains who each have 25-plus years of impeccable innovation and track record of outstanding execution. And I represent thousands of engineers of very talented teams who have been delivering industry-leading products across custom cloud and silicon.

So in the past 5 decades, the compute capability has doubled at a very predictable pace every 2 years. But something remarkable happened in the last few years, the advent of deep learning. And what happened is the large-scale AI needs shattered the cadence of 2 years and now compute capability doubles less than a year, but it's going to less than 6 months. That's a remarkable acceleration in scale. And in order for us to take a look at it, it requires a new level of thinking of system level, architecture, packaging, device design, et cetera, because AI compute doubling less than a year is something we have tuned ourselves to do it differently.

And what does that mean? The insatiable demand for scaling means we have to pack a large number of transistors what used to be easy, tens of millions of transistors at the time is now 100 billion transistors in monolithic die, but that is just not enough. In order to meet the demands of AI workloads, trillion plus transistor is the order of the day. And that can only be done through innovations in heterogenous integration of dies, multi-die systems and also vertically with innovations in advanced packaging. And even that is still not enough. You have to go from trillion transistor systems to 100 trillion-plus for rack-scale integration. Now let's ask ourselves, how do we do this at Marvell.

This can only be done with a very proven technology stack and the technology stack is comprising of superior innovations in process technology, a very comprehensive IP portfolio and advances in packaging for multi-die trillion plus transistor integration. And all these technologies that you see in the stack would just not be enough without a very production-ready integrated design flow. What does that really mean? It's really taking all these components from an architecture definition through logic design, through physical design, through tapeout and all the manufacturing test and volume ramp through an integrated design flow that we can also scale with our EDA on the cloud. And we have been pioneering that for end-to-end, and we are one of the few companies that can do an end-to-end flow, and you heard about all the design wins, and that's how we scale concurrent development of multiple chips. Now let's break down each one of this, and we have a multigenerational execution track record.

Let's start with the process technology. We have more than 20 products in volume production in 5-nanometer and 3-nanometer. And Matt said this, we are not resting there, and we are advancing our innovations in 2-nanometer with many test chips that are working. Ken will walk you through that. as well as the move from nanometers to andstroms. But how do we do that? How do we achieve this? I just want to tell you it's analog is to -- if you want to launch a Saturn rocket, the moon mission. They didn't just launch a Saturn fire rocket on day 1. They had years of development of electronics, launch systems, booster rockets multiple stages before they launched the rocket to the moon. Precisely, advanced process technologies are as complex. In order for the product to be launched at A0 to production, which means that the very first try, process technologies have to be started 3 to 4 years, codeveloping these with the foundries and with our partners and putting complex analog mixed signal IP on test chips, validate that before we can put on the product. That's a 3- to 4-year cycle lifetime.

And Marvell is a leader in advancing this shifting left such that when we put the products at A0 to production, our customers can ramp it in volume on the very first try into millions of units in a very short order. What is key to that is validated critical IP ahead of product development. So we have all heard about CMOS and nanometers, but CMOS is not the only technology that is needed for heterogenous integration. You will hear from rate about silicon photonics. Silicon germanium and silicon photonics are an equal important capability in addition to CMOS without all these technologies, the product just doesn't happen. And we will talk about what is the anatomy of an XPU shortly.

So the important advantage Marvell has is being process agnostic, meaning having leadership in development of all these process technologies, foundry ready so that we have future proof. Now let's talk about the IP layer, the IP portfolio. It means that we need to have analog mixed signal IP, which are very important and for data movement, custom CPU cores, complex digital IP. And we will dissect each one of this in what you're seeing is the anatomy of a modern XPU. What does that contain? It contains 1 very critical IP called SerDes, which is serializer, deserializer realizer. It is really the component that transfers data at gigabit rates across centimeters to meters. And not only that, you need die-to-die, die-to-die stands for exactly communication between 2 dies and the chiplets. And this is different than SerDes because it has to transfer data in very short segment at high data rate and low power. Next, you have Custom high-bandwidth memory. High-bandwidth memory is a critical component for the AI workloads.

And what Marvell has innovated is to have a base logic die at the bottom of the HBM stack and Mark will walk you through what is the advantages of that custom HBM stack. Then you have custom SRAM, standard randomstatic random access memory. And the reason SRAMs are very important is to pack as much of data in a very small area so that the rest of the diarrhea can be devoted to compute and having high-performing low-power S&Ms is a defining competitive advantage. Last but not the least is also the co-packaged optics basically making sure that all the way from electrons to photons, we can integrate all of this on the modern XPU. That's an equally important capability that rate will go through. And then if you take a look at how this is integrated at an advanced packaging, it's multilayers, whether we call it 2.5D, 3.5D, 4D, et cetera. There are innovations in materials thermal and power management as well as signal integrity and power integrity in order to make sure the entire system works in the power envelope that you see in the rack scale AI systems.

Now Matt talked about XPU and XPU attach. So there are 2 different varieties. What is important to see in the XPU attach, there may be certain IPs that are not necessary, for example, CPU. But what it means that for each of this, the power and performance per watt requirements are different, which means you need to optimize different SerDes or different die-for-die for each one of this to meet the needs of the workloads. So customization to achieve the highest performance per watt is a Marvell specialty. And you can see there a lot of content in each of these designs. Now SerDes is something that Ken will talk a lot about in the next session.

Why is SerDes so difficult? You can think about an analogy saying, if you want to transport millions of passengers, you can do it through multi-line highway with many cars going at 60 miles an hour. But what if you want to send it faster, you would have it on a high-speed bullet train traveling at 300 miles an hour faster but reliably, meaning SerDes is exactly the same thing. It has to have multiple bits at gigabit transfer speeds for reaching a longer distance across a cable with the lowest power, lowest latency, you need to get from the data movement from one chip to another at a very fast pace. And you cannot lose bits, meaning you have to have the lowest bit error rate.

Marvell has multi-generation SerDes leadership on 56 gig, 112 gig, 224 gig and you will see the demo of the 400 gig as well. And this takes a very talented team and a longitudinal experience of seeing silicon results to continue the innovations on all these figures of merit. What I wanted to discuss now is each of the technical leaders who are going to come up on stage, will talk to you about SerDes and die-to-die, custom HBM and custom SRAM innovations, co-packaged optics and silicon photonics as well as advanced packaging. And Marvell is a technology powerhouse because we have focused investments and the technical leadership is second to none. This enables first-to-market advantage, which is critical for deploying AI silicon at scale.

Now I would like to welcome Dr. Ken Chang, who is the Senior VP of Analog Mixed Signal designed at Marvell. He has more than 25 years of experience building series across a wide variety of companies. He's also an IEEE fellow and leads a very capable team at Marvell. I would like to welcome Dr. Ken Chang.

Thank you, Sandeep. Sandeep say, I lead the Analog and [indiscernible] team at Marvell. 1,000 people draw 1,000, amazing. A lot many of them with decades of experience of SerDes and long mix signal. In Marvell to my organization, the in acquisition, we have an awesome optical DSP team led by my peer [indiscernible] 800 people strong. Why am I telling you this? Marvell not only have [indiscernible] talents, we have talents in skill. You heard it's incredibly important IP and this IP needs to be developed early for this incredibly increasing AI market. Again, myself, a 25 years experience in dimension surprising. Actually, no surprise that my PhD was also on [indiscernible] 2 gigabit. Let's see how far we've come.

I'm going to show you Marvell SerDes leadership in a different way, not what we said. Look at IEEE. This chart shows the SerDes over the years and technology. Y-axis is the linear. Each data point is the paper, SerDe's paper published to top tier. IEEE conference, [indiscernible] conference, ISSCC, also top optical conference OFC. This all industry paper, why am I showing this. This show recently moved out to 2x line rate for 2 years. What I keep on talking about everybody else? Where is Marvell. Marvel dominates. These are peer review paper, Marvell was not part of it. It was reviewed by industry worldwide is expertise. I want to emphasize without Marvell, okay, with the Marvell competition, they love it, they recognize this. As my colleague, Rod is going to show you, Matt has will talk about we're the industry first to demonstrate 400-gig optical IO here. I know some of you work on that here. Line rate is not the only metric, Sandeep mentioned already, power, lower the better, a 200-gig long range. We are able to achieve 4 pico joule per bit. What does 4 pico joule per bit? For 100 terabit between XPU and switch, the 400 watts. So every pico jewel saving equate to 100 watts that we can use for compute. Next is reach. We achieved 50 dB in OFC demo. What does that mean? 2-meter cables between XPU switches. The longer the cable, the larger the scale up racks Nick talked about. We can use that for computer game.

Next is bit error rate. We achieved [indiscernible] in these long-reach 30s, clear the IEEE back by 4 order of magnitude. We also need to put a lot of SerDes on the XPU. So we design the SerDes with the lowest area in mind. And we've done that. Show in tail here, most of SerDes engineers like to CEO only. This is the tie. Why am I showing you here. In parallel, we have electrical is optical sets or develop or published. Don't get me wrong, this is not academics, okay? We build this for products, we build this for the XPU. So Sandeep mentioned, it takes years. We designed test chips for XPU.

In 2022, the team developed a 5-nanometer XSR LR in [indiscernible] that is in deployment today. So before I move on, we are going to continue this publication journey to let people know what we have done. So OFC demo, the 400-gig and the 200-gig and the optical SerDes suites, we plan to publish in next year. Sandeep mentioned, the challenges I'm going to go a little bit technical here. So as a transmitter output, we do have equalizer -- we opened the eye for PAM4, 4-level pulse and pet modulation going wide direction, 3 eyes, the larger, the better.

After 2 meters, 50 DB, if you compute that less than 1% [indiscernible] the eye is closed. Eye is closed. We cannot detect the data. So as a receiver without going into detail, we have technique, we have equalization technique, DSP to open them up to 4 distinct level. That's what we do here. So at 200 gig, the team delivered the team delivered. Next is die-to-die. In contrast to SerDes, which is 50 DB, 2 meters, the die-to-die is within the package. The reach is in millimeters, 3 millimeters. What was a big deal here? What was the challenge? As my colleague, Mark Chumley, who is the next speaker will tell you, we need to pass a massive amount of data between the XPU and IO chiplets. And more importantly, a lot more data between XPU and custom HBM, the [indiscernible] die. So there's a different metric [indiscernible] bandwidth density, terabit per millimeter.

Today, in the lab, the team achieved, we achieve 10-plus terabit per millimeter. That's not only that at 0.3 pico joule per bit, Remember, SerDes, 4 pico joule per bit. If we do at here, this thing will be influenced. 0.3 pico joule at 100 terabit is 30 watts. Again, every sub pico joule, I talk to the team, 0.01 pico joule means a lot. We can use our full compute. For tomorrow, [indiscernible] excuse me, Mark, hey, we need more than 30 terabit per millimeter. The team is working on that, and some of you are [indiscernible] here. If you notice, the pico joule even last 0.2 pico joule per pit. There's require innovation since the team is working on that. I cannot say it too much. By the way, we know how to do it. Okay. We know how to do. For the future, the die to die the roadmap is much shorter than the SerDes. SerDes 2x for 2 years. This when I say future, actually in the next few months, okay? We need to deliver 50-plus terabit per millimeter. This is in collaboration with a distinguished and packaging team here. and some of them are here. And my colleague, May Young is going to talk about the 3.5D package. And we can achieve even much less than Popper. So now I show you the die-to-die and SerDes.

I'm going to hand it back to Sandeep to introduce our next speaker.

Thank you, Ken. You could see the insights on how difficult SerDes is and why we are winning and why we are industry leading. The next speaker, I would like to welcome is [ Mark Kumerlay ] who is our VP of Technology of Cloud Custom Cloud Solutions and architecture. Mark has had an industry track record 25-plus years across IBM, GlobalFoundries and Avera, and he is innovating at the cutting edge on custom external memories, internal memories, and you will learn all about it from Mark. Mark, welcome on stage.

As Sandeep mentioned, my name is Mark Kimberley. And I lead the custom cloud solutions architecture team for Marvell. We work closely with our data center customers to develop cutting-edge solutions for next-generation systems. Today, I'll be speaking about 3 different technologies. We'll be talking about embedded SRAM, customized embedded SRAM. We'll talk about customized HBM, and we'll talk about something we haven't spoken about before at Marvell publicly, package integrated voltage regulation.

Starting with SRAM. Most people think SRAM is just something that you get from the foundry or from an IP provider. It's just something that you integrate on a chip, you set it and forget it, you don't have to worry about it. It just is what it is. You bring it in when you need storage, that's what you do. I'm here today to tell you that, that's really not true or that that's not the right way to approach it. If we look at this diagram, custom SRAM can take up 30% or 40% of a next-generation accelerator CPU or switch device. It's an incredible big portion of the overall design.

How can we not optimize that to provide better performance to our customers? It's crazy. I'm here to tell you that at Marvell for the last 25 years, we actually have a crack team that's been innovating in custom SRAM design led by Darren, who's in the room today, somewhere. Darren, thank you. They've been working in every technology node and innovating in every technology node to open up more performance and lower power for our customers. And let's see what they've done recently. This morning, we announced our 2-nanometer custom SRAM that is optimized to deliver maximum bandwidth to the data center.

We deliver an astounding 17x the bandwidth per square millimeter of off-the-shelf SRAM solutions that we can get from the foundry or IP providers. It's an insane amount of bandwidth, but bandwidth is so essential for these applications to feed these hungry compute units. What's even more special about our custom SRAM is that we deliver this amazing amount of bandwidth to our applications at 66% lower standby power. than other competing solutions. Incredible amounts of bandwidth at much lower power than anything else on the market. This enables our customers to use more in their systems, build bigger data centers to have more compute at the same amount of power, phenomenal achievement.

Moving on to another type of memory. And what our custom SRM does for bandwidth, custom HBM really enables for high-capacity memory, delivered efficiently to the compute on the main die itself. At Marvell, we've been investing and developing custom HBM to really make these applications more efficient. And I'll show you how we do that. If you think about a normal accelerator, GPU, XPU device today, Infrastructure, HBM-IO, takes an incredible amount of the main die along with our normal IO. And you can see in the example on the left, that our compute is actually constrained by the amount of infrastructure I/O that's required to basically transfer data between these HBM. A -- it takes up more and more of the die as HBM advances in technology.

With Marvell's custom HBM, you can see that we unlock 1.7x more useful compute area on the main die by removing those huge I/O areas and removing the HBM controllers and using the die-to-die technology that Ken just spoke with you about. We create a huge open expansive area that our customers can fill with compute to make the most competitive device as possible. And this area, there's a reason why we colored at gold is gold to our customers. It's a huge impact for them to be able to scale up the compute 1.7x the amount of area that they had before. In addition to this incredible experience of a new area that we deliver to the customers. Because we develop our custom HBM based die on advanced technology nodes, we open up additional area inside the HBM based by itself, that our customers can use to integrate those high-bandwidth [indiscernible] devices they can integrate I/O, additional die-to-die, unlocking huge opportunities on the accelerator. And even more amazing than all this extra space is what it does for the power consumption of the XPU device itself.

By removing power-inefficient HBM interfaces and replacing them with an incredibly efficient die-to-die interface developed here at Marvell, we're actually enabling 75% lower memory I/O power when we adapt these accelerators to custom HBM. It's quite an achievement, and it lets our customers really scale up the amount of compute in the data center. Now on to a personal note, Anybody who knows me, and there's probably many of you in the room who do know that I often walk around with a pocket full of XPU devices, which I have today. But I'm actually not going to talk about XPU devices because we're talking about memory. So I want to share something that's actually very exciting to me. We talked about custom SRAM, where we have gigabits of capacity. We talked about custom HBM, where we have gigabytes of capacity. And there's some really cool technology at Marvell that's unlocking terabytes of capacity.

So this little device here that we're super proud of as a [ Structera-A ] module, which is a memory pooling device that enables additional compute with the memory access itself to terabytes of memory for our customers. It's a huge innovation for the data center, and I'm proud that our team, many of you are in the room are a part of it. Moving on from memory to another topic. We actually just released a press release about this today. Marvell has created a platform with industry partners to enable package-integrated voltage regulation. What does that look like? Now on this slide, we can see the typical accelerator picture that we grabbed from the last slide on the underside of that accelerator, we'll actually be integrating package-integrated voltage regulation. Why would anybody do this? Well, the thing is these big accelerator devices and even computer switch devices that we customize have incredibly complicated boards, incredibly complex. And just delivering the power to the accelerator through the Board actually takes a fair amount of power loss and inefficiency.

So moving the current through these incredibly big fit complicated boards, actually makes the system far less efficient and consumes power itself. With packaging graded voltage regulation, we can cut down 85% of that IR power loss going through the Board itself because the regulation is directly on the module. This can result in up to 15% of a benefit in total product power, which is quite incredible. Customers don't need to buy so many power plants to feed their data centers using this technology. But maybe even more important and probably a near and dear subject to many people in this room is power supply noise. Some of us have spent many hours trying to mitigate power supply noise because, quite frankly, as an accelerator goes from idle slamming to fully active. There's an incredible draw on the power supply. And you can actually watch the voltage dip as these devices go into operation. what we typically would have to do to account for power supply noise is margin or take sacrifices in the design density, adding that margin can add additional power consumption to the device.

And the amazing thing about this voltage regulation technology and this platform that we're developing with other companies in the industry, is that we can enable much higher speeds to react to this power supply noise and take away 60% of the power noise in the system, lowering the overall power noise by up to 60%. It's quite an achievement. And it means our customers can do more with that extra power and they can use more of that compute more rapidly to deploy on their workloads. We're learning more and more about what this technology can do for us, and it even opens up, we think new opportunities in other advanced integration like copackage optics, which you're going to hear about soon by Rod.

With that, I want to thank you and hand it back to Sandeep.

Thank you, Mark, for really allowing us to understand how you can unlock more compute with innovations and integrated voltage regulators and innovations in memory. Our next speaker is Dr. Radha Nagarajan, who has 30-plus years of experience leading focused innovations in silicon photonics and optics. He is an IEEE fellow. And also the optical engineering society fellow has been inducted into the National Academy of Engineering. You will hear all about the exciting world of silicon photonics from Rada. Welcome Radha.

Thank you, Sandeep. It's great to be here in the presence of all these technical talent in the room. You've heard a lot about co-packaged optics from several speakers -- several speakers at this event. As Sandeep said, I've been doing this for 30 years. And like Ken, I started designing optical interconnects at a gigabit speed 30 years ago. Today, we had terabit speeds. And over this time, optical interconnects have been pretty much done the same way. The last couple of years, co-packaged optics is a sea change in optical interconnects, where you bring interconnects to the custom silicon. And critical for co-packaged optics is silicon photonics.

You've heard a lot about silicon photonics. What silicon photonics enables you to do is to design the entire optical system on a chip. A traditional optical component design is 1 at a time, you design a component, it put it together in a package. But what silicon photonics enables is, you can do all of that on a single platform. Silicon also has very high speed, as we will see, and long reach. And as the name implies, you can use existing CMOS fabs, silicon foundries to build optical components. This is the other sea change, so you don't have to build their own custom foundries. Our fabs. What's most important in this process allows complex electronics and photonics integration.

When I mean complex, you can integrate silicon germanium, CMOS and photonics onto a single common substrate, and we'll talk about it in a moment. Marvell has been working on silicon photonics for over 10 years. This is not as commonly known. Silicon photonics applications for data center, to large application space. One is between data centers, as you saw in one of the slides between campuses, and the other is inside that centers. Marvell chose to attack the harder problem first by introducing products for between data centers, where the reach is several hundred kilometers as opposed to several hundred meters and the data rates are generationally higher. So we started shipping the 100-gig product in 2017. And 400 gig product is shipping in volume, 800 gig is sampling and 1.6 terabit per second. All these are per lambda is in design.

And the brain trust for this silicon photonics progress is in this room. And that's why it's so great to be addressing this group of people. We have multiple generations of field applied silicon photonics. And that's very important. This is Marvels pedigree. Number two, high speed electronics. As multiple speakers have pointed out, things seem to be happening every 2 years, who thought, I mean, one was the darling of the industry not too long ago. We're deploying 224 in volume. Today, inside Marvell, we have 480 gigabits per second single lane electrical. We used the same cell of electronics, which had a lot of extra margin as you can well tell to do a 450 gigabits per second demo at OFC 3 months ago. This progress of 480 is in just the last 3 months.

Again, this 400-gig enables 2 classes of applications. One is the 1.6 ZR between data center application. Single lane and a coherent format. Coherent format, the way [indiscernible] a 4x data density compared to inside data center, and that's the difference. At 400 gig, at 32 lanes of 400 gig, we are well on our way to designing the next generation called packaged optics, 12.8 terabits per second. So how do you put all of these together, another common theme at this event, advanced packaging. This is where it all comes together. We'll discuss 2 levels of advanced packaging. Let's look at it at the die level. This is a cross-section of a light engine. The way to read it is you started the electronics layer on the top -- and then there's a thin sliver of silicon. You may or may not be able to see it, which does the bulk of the work. Process is optical signals and allows for the integration of the electronics above it.

And this electronics could be CMOS, Siggi, in some cases, other material systems as well. And then they are through silicon vias and you go to the substrate. And as Mark showed why stop there, integrate things to the back of the substrate. Decoupling capacitors, electronics. And this is the basis for the 6.40 optical engine, the bottom left, 32 lanes 224. And this is where the beauty of advanced packaging comes in. You take that, which is already a 3D silicon engine and you integrate it onto a substrate to obtain an XPU with 4 of these light engines integrated together for 25.6 optical interconnect to an XPU complex. And to tell you all the details about how you do the next level of integration is my colleague, Mayank.

I'll hand it over to Sandeep. Thank you.

Thank you, Radha. Radha and team make even the most complex technology, simple, but it's not that simple. We will now transition to our next speaker, Mayank Mauk, who is our senior distinguished engineer of advanced packaging. And he will talk about all the innovations and direction of advanced packaging and the innovations will be very entertaining.

Thanks, Sandeep. All right. So like Sandeep said, I'm going to be talking about advanced packaging. And just a few years ago, advanced packaging used to be an afterthought. That's something you did around tapeout, something taken for granted. That's not going to work anymore. It has emerged as a key differentiator. It is as important as silicon, if not more. All I'm telling you is for those we hired recently in packaging, this is a good time to be a packaging engineer. So because of all the drama going on in advanced packaging, we are making investments to expand our road map for current as well as future generations of products for switch as well as for XPUs. How do we do this? We are doing this by partnering with the right partners as well with the right strategy. And we are starting with creating some foundational IPs in design, materials as well as process. Then we forge partnerships with OSATs and foundries and so on. And with their help, we co-create these building blocks. This technology building blocks are fungible across generations which means that we can create a very long range of road map in a very short time.

Now these building blocks can also be mixed and matched which means that we can create real custom solutions in advanced packaging by all these building blocks that can be with different permutations and combination of these building blocks. And also, since these building blocks are fungible, and generational, we are able to pull this up in a much shorter time. So the cadence between generations is short, shorter than ever before. Next, I'm going to show you the evolution of advanced packaging and how Marvell is leading the way. So we started doing a 2D package a while ago, 20, 25 years ago, a 2D package has silicon, which is directly sitting on the substrate and then we have a lid on top.

Next, we transitioned into a 2.5D package in a 2.5D package instead of 1 monolithic silicon. What you have is silicon chiplets, And these chiplets are integrated on to an interposer, the interposer then sits on the substrate, and that's how we create the 2.5D package. Now by doing that, we are able to scale the package from 1x to 4x 4x. The next one is a 3.5D package in which it has everything that a 2.5D package has to offer. But in addition to that, we have a die stack, a 3D stack silicon chiplets. And with that, we are able to double up the capacity or develop the scaling of the package is 8x compared to the 2D package, which just feels like yesterday. The next 2 packaging technologies, the 4D and the 4.5D take advantage of recent innovations that we have done in terms of substrates, for example, in substrates, we are using engineered materials that is helping us to scale from tens of millimeters to tens of inches.

And we are also integrating optics and copper right into the package. And with all of this ensemble, we're able to scale the package to 16x compared to the 2D package. So let's take an example and we're going to build a package ground up. So at the very bottom layer, we have substrates. In this particular case, we have advanced substrates, which, as I told you, has much larger scale. Also, we are embedding active passive and optical components right into the substrate. The next is an interposer. It could be 1 piece or it could be multiple pieces of interposer. And we have both bridge space interposer as well as RDL-based interposer. Beyond that, we have the 3D IC.

In a 3D IC, we have stacking of top and bottom dies using hybrid copper bonding. And with that, we are able to create these high-bandwidth interconnects this high-bandwidth interconnects then help to tie together multiple pieces of silicon, and you can pack in about 2x the amount of compute in the same footprint. And finally, we have integration of optics that gives you high bandwidth, low latency connectivity as well as we have copper integration right into the package. -- that gives you improved signal integrity, that's much lower power and much lower cost. So what I have shown you today is that we can have a combination of different building blocks with this 1 as an example, but you can combine these building blocks in different ways to create unique and custom solutions that is optimized for a particular workload. Now back to Sandeep. Thank you.

Thank you, Mike. I think more well from all the technologies you have seen today is an industry powerhouse, and we have been able to do this with sustained innovations and a track record of extraordinary execution not only for the current generation, but we are on track to do it for the next generation and the generation after. As a result of this I hope you're convinced that we are a handful of fabless companies less than 4 or 5 that can make this happen. And with that, I want to you to the next speaker.

Welcome to Vice President and General Manager, Cutom Cloud Solutions, Will Chu.

All right. Good morning, everyone. I'm super excited to be here to discuss the fantastic progress of my business and the incredible custom opportunity for Marvell. Based on our nonstop pace of activity, it's clear that we continue to see strong customer momentum. Once again, my name is Will Chu. I started my career at Texas Instruments as a design engineer. I then earned my MBA at MIT. Eventually, I joined Maxim, and I've been at Marvell for the past 8 years. My presentation will focus on the unique value Marvell brings to our customers and why we are extremely well positioned to continue growing rapidly. The Custom Cloud Solutions BU was formed through the integration of Cavium, Avera and internal Marvell teams. Cavium brought decades of experience in custom compute networking and security.

Avera brought more than 25 years of experience in custom silicon having delivered more than 2,000 ASICs. I've been involved in the custom business from the very beginning. And I personally drove the integration of these teams to form a custom silicon powerhouse focused on winning in cloud. Now let's discuss the strong design win momentum we have achieved. Matt had discussed the enormous traction we've had in the market. We have design wins with all the top 4 hyperscalers and emerging hyperscalers. We have custom XPU and custom XPU attach design wins, and many of these are multigenerational. Now I'm super pleased with this traction and the increase in the number of customers and sockets.

I wanted to thank the entire Marvell team for making it happen. And as you've seen in the previous sessions, we have a world-class engineering organization at Marvell. So let's dive into how we want so many sockets. Matt discussed the typical system architecture, where there are custom XPU and custom XPU attach opportunities. For the custom XPUs, our customers have specific workloads that they want to optimize with customized silicon. That tight integration between their specific workloads and custom silicon can drive enormous increases in profitability for their cloud infrastructure.

For custom XPU attach opportunities, there are 3 main types: networking memory, networking and coprocessors from memory. Cloud customers today buy billions of dollars worth of memory. And they're looking for specific custom solutions that will improve memory capacity, reuse and utilization that can drive enormous TCO benefits for networking. There's an enormous amount of data movement throughout their systems. And again, the customers are looking for custom solutions that optimize that data movement to deliver increased performance and efficiency. And finally, for coprocessors. The customers are looking for ways to secure and control their infrastructure and custom solutions enable them to do that security and management at scale. Okay. With this as a background, now I'm going to dive into some examples, and I'm going to start with custom XPUs.

Okay. So you see here an XPU and normally, the engagement starts off with an architect like Mark, engaging with the customer, and there's a deep intense collaboration to figure out what to build and identify how to use Marvell's unique value add to make their custom solutions winning. We also showcased all the unique technologies you've seen just presented previously in Nic's and Sandeep's presentations. In that process, we also tend to learn about their customers' next-generation opportunities. Okay. So let's dive in a little bit.

So we go through the process that Nick described, let's talk about system architecture. So I listed here, rack scale enablement and optimization. So this is all outside the XPU before we even get into the silicon. The customers want to scale the XPU as you saw, tens of thousands, maybe up to 1 million XPUs in a single solution. We support that system architecture engagement with the customer. That's called Rack Scale enablement. A perfect example is what Rod described with or Nick described with CPC. So today, we're working many customers right now in those technologies to enable them to scale their solutions for thousands or up to 1 million XPUs.

Rack scale optimization. This is what Mark discussed with our integrated power solutions. He talked about reducing power of 15% or more inside the XPU. If you multiply that by thousands of XPUs. This is a rack scale level optimization that the customers are looking for. Marvell does that uniquely. Next, design IPs. So these are critical IPs. And we talked about those. In a typical XPU, you have SerDes, die-to-die and [indiscernible] Those are all -- we have the team go through that. These are table stakes. Our customers want and need world-class IP, and we have that at Marvell uniquely. And if we were to move to CPUs, of course, we have developed years of experience -- decades of experience doing ARM products, and we can add that in the critical IP list to develop leading CPUs.

Let's go to our silicon services. So what's unique is I list here, design to spec and co-development. In many XPUs, as you saw, there's many different chiplets in the architecture. On the design-to-spec side, in many instances, we at Marvell will own one of the chiplets in the design, right? And we do most of the design work in conjunction with our customer. On the co-development side, typically on the main ASIC, they will own that, and we will support them. But it goes beyond just a traditional physical design relationship. We're typically supporting the customer on the front end of that, doing things like emulation and FPGA work to make sure that their design works really well. Again, this is unique -- this is a unique value that Marvell brings to the table. Packaging. Mayank talked about this.

So Marvell has both custom and we're using the traditional TSMC cooas based packaging. We deliver custom packaging to our customers because, one, it gives them more capacity in their solutions. And number two, it enables a lot more design flexibility as they design their package. And last, manufacturing and logistics, I list here faster time to market. Now all of these things that we're doing already shortens the time for the customer to take their product to high-volume manufacturing. But beyond that, because of Marvell's scale, our expertise, we're able to do many things in the manufacturing side from tape-out to general availability in parallel that helps pull in the time that it takes for the customer to go to production and dramatically improve their TCO. Now this is a really good example of all the things that Marvell brings to bear for our customers. And these are all unique. And these deep engagements enable us to drive multigenerational opportunities.

Okay. Now let's look at an XPU attach example. Okay, in the XPU attach example, it's a slightly different model. So as we mentioned, there is memory, networking and coprocessor type opportunities. We have experts in all 3 fields that have decades of experience in developing these kind of products. and they're working with our customers to showcase our technology to make their silicon dreams come true. What you don't know is that many times, we are writing the specification for those customers because we have this expertise and we have this technology. So we're working side-by-side very closely with the customer to define what the product really is. And that's unique to Marvell.

As we jump into the architecture, I have here production firmware software boards and ecosystem. So again, outside the silicon, we are developing the production firmer and software for our customers today in all of these 3 areas: memory, networking and on the security side. Boards. So in the networking space, we're delivering full production boards for our customers today. In the memory space, we are optimizing the Board with the various DRAMs that go on the board so that the customers have an optimal solution. On the system [indiscernible] side, ecosystem, what does this mean? So we're working with the ecosystem partners to make our solution fantastic for our customers. So on the memory space, we work with the ecosystem of DRAM vendors so that when the Board comes up, it works really well. In the networking space, in a NIC card example, we work with the CPU or the XPU partner on the Board to make sure that they bring up between the NIC, the custom NIC and the CPU work really, really well. In security, we work to deliver compliance certification like FIPS for our customers. Again, these are big unique value adds for our customers.

Next, on the design IP side, critical IPs. So I list a bunch here. And for example, the first couple, compress, decompress and our compute fabric. So in the memory space, what we're able to do with our unique IP that we've developed over decades is we're able to pack more bits into the same amount of memory. Obviously, this is a big benefit for our customers. And we have a compute fabric that can move the data across efficiently across the custom memory accelerator.

Security. So we have leading cryptography IP at Marvell. And with that, we're able to deliver leading solutions for our customers to develop their security products. And finally, on the networking side, SerDes, which Ken talked about, right? So we have leading SerDes, which enables world-class custom networking solutions for our customers. Okay. I'll move on to silicon services, design to spec integrating customer IP. As I mentioned, in many instances, we're helping the customer write the specification for the products. And in this case, design to spec means we're designing the entire chip, but we also integrate our customers' IP. And it sounds simple, at least on the slide, but it's not that simple. So what we're doing is we're taking their IP, we're instantiating it in the chip, and we're making sure it works seamlessly with everything else that we're designed. So we are doing things like emulation and FPGA work, for example, but we are also attaching it to the system level.

So their custom IP needs to work within all the IPs that we have as well as deliver connect to the production firmware and software work well on the board and all the rest of the ecosystem around it. Again, these are all unique value adds from Marvell, and these deep engagements are driving our multigenerational engagements. Now let's see how this unique value supports customers across a single complete program. So here, we have a bunch of different phases. And then the first one is the IP development, which has been discussed. Marvell invest years ahead of time before we get a design award. So we designed the IP and we demonstrated to our customers that they have comps, and this is what they need. In the next phase, in the system architecture, we have this intense collaboration where we decide what to build and how to engage. And then finally, we have full ownership of the end-to-end solution after the design awards been given to us and we take the chip all the way to production.

Now this entire flow is an unmatched combination of technology expertise and scale. This is not something that physical design services does or just the manufacturing only. service can do, right? So this is unique to Marvell. Now by investing upfront in the IP and working so closely in codevelopment with our customers. We also get engaged in the customer's next generation. So let's see how this unique value-added model supports our multigenerational partnerships, especially in light of the rapid pace of innovation in AI. Okay. So as I discussed in a single program, we invest well ahead of the curve. But due to the rapid pace of AI innovation, customers engaged in their projects in a much more compressed time. So the customers are engaging in multiple generations at the same time. And I'll give you an example here.

So when we have a 3-nanometer design win, as I mentioned, we had already invested years ago in that -- but we're also investing at the same time before we even get the award in 3-nanometer [ and ] our 2-nanometer IP. And we're also then investing in our 16 AnshimerA16 IP as well. So the customers are depending on us as Sandeep described, to do all this innovation. And this is what enables concurrent engagement and enables a chip every single year. Now this is just an example on process node. Let's see how examples of this for our IPs that we have announced in just the last 6 months. All right. So here's a list of the press releases of Marvell in the last 6 months, press releases of all our breakthrough IPs. So we have our integrated power solution, which Mark talked about. And our 2-nanometer custom strand, which Mark also talked about. We announced both of these today. We have our advanced packaging platform, which Mayan talked about. And we have our industry-leading 2-nanometer platform, which Ken described. And our breakthrough copackage optics, which Rod discussed and our breakthrough custom HBM, which Mark also discussed.

Every single one of these IPs are early and ultra unique. And each one of these are driving specific customer engagements for next-generation opportunities. So that leaves us the opportunity pipeline. As Matt discussed, we see over 50 opportunities in front of us. about 1/3 are custom XPU and 2/3 are custom XPU attach. We see this many opportunities because traditional and emerging hyperscalers see the unique value that Marvell brings to their programs for this generation and their future generations. Said differently, we have a seat at every table for these opportunities. I am personally involved in driving every single one of these opportunities. And you have seen the outstanding engineering leadership at Marvell driving our innovation and execution.

So to recap. Marvell brings incredibly unique value to our customers. We are engaged in every opportunity and extremely well positioned to win. Thank you.

Thanks, Will. We will now start our Q&A session. Let's give the event team a few minutes to set up the stage. In the meantime, investors and analysts can continue submitting questions through the live video screen window. I will read those questions on to the team. It looks like we are almost ready. So I would like to invite the Marvell team back on the stage. All right. Let's start with the first question. So this one's come in from a few different investors. Can you clarify the difference between XPU and XPU attach with some examples of each.

Sure. And before I start, first of all, just again, everybody on the line, thank you so much for joining today. Thanks everybody in the audience. And I just want to take a moment to thank this outstanding team here -- you guys did an excellent job today. So thank you so much. Very good. Perfect. All right. Well, I'm going to direct traffic here on the questions a little bit. I'm so happy because normally, I'm on an earnings call it's me and William and then the world, but at least I got some teammates up here. So Chris, why don't you take the first question on clarifying the difference for the audience on XPO and XPU attached. We talked about it a lot, but maybe just to clarify for people.

Sure. Great. So yes, so the -- basically, the XPU attach is the portion of the compute TAM that's actually addressed by companionship. And those companion chips are different for every architecture, but it includes things like mix, scale-up fabrics, coprocessors, memory interfaces, Actually, as these architectures continue to customize, we're seeing more and more sockets as those platforms customize. So previously, this XPU attach was included in the compute TAM that we talked about last year, for example, but this year, as it's grown, we've sized it and broken it out separately just to make that portion clear. And the XPU, of course, is still a very important part of the market for us, and that continues to grow very rapidly as well.

The next question is from Ross from Deutsche Bank. Does Martin expect the number of XPUs for hyperscaler to expect to continue to expand to address new different workloads? Or is the expense to right?

Sure. Why don't you take that one also that I can add to it?

Sure. Thanks, Ross. I think in general, yes, that's what we see. We see, as I mentioned in my talk that the workloads are diversifying and that ultimately, having specialized silicon for the different workloads is of benefit as this entire thing scales and the total CapEx dollars go up. So over time, we do expect that to happen.

Yes. And I would just add, if you look at the you compound all the different technologies we talked about today in terms of what's the total benefit to our customers in terms of cost and performance. I think it's very compelling versus what's out there. It actually enables them to create more and to create more solutions for them to optimize their workloads and their technologies. So I think it's actually creating more opportunities the way we're going about this.

The next question is from Atif from Citi. Is the profitability levels, both gross margin and operating margin of XP attach similar to what you see on the XPU side?

Yes. I'll give Willem a shot here for financial question.

Yes. So I think the way you should look at it and will did a good job explaining all the different IP. And clearly, the more IP that's from Marvel that margin profile is sort of on the higher end of the custom scale. And so we're very excited about those opportunities because they're very sticky, and there's a lot of Marvell IP involved. And so -- from a scale standpoint, I would say it's on the higher end of our customer model.

And I'd also add to what contributes to that is there's just a huge magnitude of difference as well, right? The XPUs by nature, just much higher volume, the XPU attaches still actually very large volume, if you compare it to any sockets we ever actually used to go after in Marvell. If you go after the last 6 or 7 years, those alone are just gigantic XPU attached, but because they're also not as higher volume, you tend to get a little bit better on the margin side there.

Right. We have a question from 2 different investors. Same question. So are the XPU attach wins tied to winning also the XPU compute socket at the same customer? Or can they be different?

Where' Will. Will, do you want to take that one?

Yes, they're not attached strictly. So of course, we're trying to win every socket at every hyperscaler. But there's no there's no linkage between them generally. We go after each socket independently. And of course, we're working hard to win them all.

Yes. And I would just add, our customers expect us to participate and be active across everything that they're putting in front of us, right? So we don't cherry-pick. We go all in. We look at where we can participate, where we can bring the value and that ultimately that level of engagement and trust actually brings a variety of opportunities. And usually, that's been the path to actually win some of the larger ones is actually starting small. In some cases, we started actually very small. 5 years ago with some of these hyperscalers. And we built the trust showed the execution and then that leads to bigger and bigger opportunities over time. So that's the model that we're engaging. We're just all in on this market in terms of how we engage with our customers on a range of big to more modest.

Great. The next question is from Tore Svanberg from Stifel. The reports are that custom ASICs may perform below our merchant GPUs. How does this factor into Marvell's view of the market opportunity? And what is needed to close the gap if it does exist?

Yes. Maybe I'll ask Sandeep to talk about that one.

Yes. So custom ASICs are very purpose built to the workloads that each of the customers would have. And when you have to customize for let's say, different kinds of floating point arithmetic, or fixed point arithmetic. Then GPUs, our general purpose and the framework is going to be different. But you can actually fine-tune, just like you saw from the technology presentations and equally capitalize on the advancements that we bring for the TCO, which is very purpose built and you heard Ronnie also talk about it in the partnership. That is what enables an important innovations to really fine-tune the compute performance, the SerDes performance, the memory performance to the actual workloads and where GPUs may not excel from a performance per bot perspective.

Yes. So it's not -- we're not talking about performance against benchmarks here. as Ron talked about, it's software, hardware codesign. And so the performance in the workload is what matters. And clearly, these are superior performance in the right workloads.

Yes. And then the final context I would add is, in light of all this, we're still sizing the percent, right, of the total opportunity of accelerators to be about 25% as customs. So we're not even saying it's addressing the whole market. Now the more competitive those are and the more compelling we can make those then obviously, the higher the percentage contribution those could be, but we're kind of in that 25% number right now, which still spits out just an enormous TAM for us.

Great. The next question is from Ben Reitzes from Melius. When do you expect the 2 new XPU and 4 new XP attach with the emerging hyperscalers to hit your P&L? Can this be -- should we think of this as incremental to what you had discussed last year?

Yes. Thanks, Ben. Chris, do you want to take this one?

Sure. Yes. So one way to think about it is that some of the 18 sockets that we outlined all together, we had 1 last year, and some of those are in production now, that first wave, as Matt said, but many of them are in design execution. And so we would expect them really to start to turn into revenue in '26 and '27 going forward. But some of them, as I said, already are there.

Right. The next question is from Harlan Sur from JPMorgan. In the concurrent engagement model you articulated, we think you're already well into 3-nanometer designs with your lead customers. So is it fair to assume you already went into the design on next-generation programs?

I'll have Will take that one since we talked concurrent or in the middle of all these.

Absolutely. So the customers, as I said, they're looking at multiple generations ahead. As Sandeep mentioned in his presentation, it's maybe 3 to 4 years. to do a full technology cycle. But if you're refreshing the platform every 2 years or less, then by definition, you have to do things concurrently, and this is what almost every customer is looking at because they have to.

Great. One more clarification question. In the 5 XPUs you listed, how many of these are CPUs versus accelerators?

Yes. Chris why don't you do that one.

Sure. So yes, basically, 4 accelerators and 1 is a CPU. But if you look out at that pipeline of opportunities that we talked about, the 5 in the pipeline, it's actually a mix of CPUs and XPUs and then, of course, the XPU attach.

Right. The next question is from Gary Mobley from Loop Capital. For the IVR and custom SRAM solutions you outlined. Are these specifically for a custom AI XP or Xpotach Marvel is currently working on? Or is it available more broadly beyond the custom platform?

Yes. Do you want to do a little bit and then Mark, you add to that as well since you see the whole.

Yes. So we have customers that are -- we're working with closely to design in our 20-nanometer custom SRAM technology for sure. And IVR -- oh, on the IVR side, we have multiple engagements with many customers trying to take that solution to market, like it's very rigorous, let's say, or intense because as you can see, the power benefit that you can derive from something like what we're developing on the IVR side is quite compelling.

I'll just add to that. If you look at the cost of SRAM, for example, the benefit of it isn't really limited just to XPU devices. It's something that could be highly beneficial for not only accelerators for AI, also for processors, also for switching applications also for relatively simple networking applications like a Nick. So the benefit really isn't limited to just the very high-end AI accelerator devices similar with IVR. A lot of products have challenges with power supply noise and power efficiency. And that technology can be brought to bear to help many different applications.

Yes, I mean, as you guys said, this is really a pan Marvell benefit we can get from this deep engagement on the AI market, we can actually leverage that technology across all of our products.

The next question is from Joe Moore from Morgan Stanley. Can you talk about the NRE relationships here? How much of the R&D for these opportunities is funded by our customers? And is there a difference in between compute and compute attach?

Sure. Willem, do you want to take the first part and then Will, you can chime in on the second part.

Yes. So just as a reminder, NRE is nonrecurring engineering. And so when you look at our custom engagement model, the way we operate is that our customers sort of co-invest and we recognize that as a reduction in our operating expense. And so if you look at the operating margin, that these programs drive, that is an improvement in the operating margin. And so when we look across all these programs, you should expect it to be very consistent in that when we look at the compute attach or the XPU attach, there's a significant NRE component, very similar to the XPU programs that we have today.

That's a great answer. Anything to add or.

Great. The next question is from William Curvin from Morningstar. What are situations where a customer may choose a less than full service vendor? And how large do you see that as a piece of the total TAM, especially as you look forward?

Yes. Do you want to leave that for.

Sure, yes. So ultimately, what we see going -- I mean, clearly, those are situations where they might be able to source the IP on the open market, and those tend to be IPs that are generally available. So usually, we see that for sort of slower moving parts of the market, right, that aren't sort of moving at the rates that Sandeep talked about. And that Ken and the team talked about where the IPs are going so fast that it's not something you can generally source available market. And that's where you're able to sort of put together different services from different companies and produce a products. So that does happen. There are plenty of examples of that. It just tends to not be in this sort of accelerated computing going forward.

Yes. Nick, anything to add to that? You had some slides around this topic.

Yes. I think the key point is that as we see continued acceleration in the applications and customization for the workload, the value of having a differentiated technology and doing very, very high levels of integration is just getting higher and higher. If you compare it to 5 years ago, you could build a product that's simple by comparison and be competitive. But now that's just not an option. So that's where the full service becomes a lot more critical.

Yes. And we see opportunities kind of up and down when we engage, but typically, when it gets real, and we actually have to go deliver and there's a tight time frame and you want to just underwrite your execution. And that's where our success has been. It really kind of floats to the top of that model in terms of needing to get the full benefit of what Marvell can bring to the table.

Question from Adam Wackers at Wells Fargo. Can you talk a little bit more on how we should think about the custom HBM for logic die timing and how this could be leveraged as a differentiator in your custom XP or attach design wins?

Got you. Will, do you want to talk a little bit about just rough time frame on some of these things? And Sandeep, you can add if you've got a perspective.

Yes. So on the HBM 4s, I mean, this is public information. They're coming out to market like later this year, all the samples. And so we are engaged with customers to figure out how to incorporate our custom HBM technology to support that. I think most customers are targeting HBM 4E, which is the 4 echo, as they would say, which is the kind of higher performance solution that comes after custom HBM 4. And I think there's a bigger kind of interest there just because of the performance that's needed and the benefit you can drive with that higher performance HBM coupled with custom HPM technology.

Right. The next question is from Craig Ellis at B. Riley. Given the large number of wins 18 you outlined and the pipeline of 50 plus, can you please discuss capacity planning across the ecosystems to ensure sufficient supply to meet customer demand. Are there any upside limits or concerns? And if so, where?

Sure. I'll have Chris take this one. He leads our operations as well. So we've been deep in that planning for years, actually, you and I in terms of getting ready for the ramp we've got now and then in the future?

Yes, that's really the answer. I mean, starting back in 2020 when the original supply crunch had, we started doing long-range planning for all of our supply chain, including packaging, substrates, foundry, et cetera. In fact, we do a 5-year forecast for our suppliers, and we give it to them every year. We've done long-range contracts in some cases where we need to. We've been planning for these ramps for a long time, and we're very confident in the capacity that we have lined up.

Yes. I think I'd just add, I mean, we've come so far on that front in terms of shifting from kind of reacting to demand to thinking that, hey, we got to plan multiple years in advance. And we actually made investments right in the supply chain. We built partnerships. We put in contracts to make sure we had access to the best technology. We staffed the team. So from that standpoint, I think we've got a very, very robust supply chain set of partners they're deeply committed to us and our success. And clearly, with the magnitude, which is a great question, we need to plan even more aggressively in advance as we chart our path on data center, right, from a couple of billion dollars to $4-plus billion last year and this ramp we've got through '28 and beyond. But I think we're in great shape as a company from a supply chain perspective and the partners we have.

The next question is from an investor. Could you talk about the various flavors of SerDes that are used in your XPU portfolio specifically? For example, extra short read and others and specific attributes where your portfolio is better versus merchant solutions?

Yes, Ken, do you want to lead off with that? And maybe Mark, you guys can team up on that.

And you the different flavors. Different flavor SerDes. We primarily focus on long reach SerDes. We focus on electrical and optical SerDes, and we're able to kept coherent SerDes. On electrical SerDes, we primarily focus on long reach, although we support show reach as part of that. We do have XSR capability so that can turn on any single time when there's a customer demand.

Just to add to that. I think from a broad portfolio perspective, all the data rates that Ken added 112, 224 on all different process nodes that you see 5, 3, 2. So it's a deep portfolio of multiple data rates across electrical, optical and coherent modulation that makes it a fully blown portfolio.

Yes. I think the concurrent planning, in particular on this part of the engineering of the company is extremely complex because you're trying to shoot basically 3 years out in advance with multiple nodes, multiple reaches, multiple line rates. And so it's a nontrivial job to do the architecture planning. But the team that has been assembled under can is just really world-class in terms of their ability to think through what's needed and then actually plan the engineering and then get on the shuttles and -- but this is a machine now we've built. And I actually feel it's a machine that's driving leadership actually in terms of product release and IP availability. And that's what's really winning the customers, right, as we're able to show real working silicon in the lab -- and that's giving customers confidence that we can continue to provide this IP, which ultimately is a key determinant of do they keep working with us or not. So it's been a great job, can you and the team on this.

Maybe just add 1 more. So we not only focus on Tapi. We also focus on, as metal say, how you can support on the customer support. That's actually to [indiscernible]

We have a question from an investor, a 2-part question. First is can you clarify how you trade NRE in your OpEx Second, can you also speak to how your OpEx will scale as you move from supporting a handful of wins last year to basically 18 wins you talked about today, and you're looking at 50 plus in your pipeline. So how should we think about OpEx in the future?

I'll let Willem handle that one. I might add at the end.

Sure. Yes. So in our nonrecurring engineering, so we recognize that as a reduction in our operating expense. And so customers effectively invest with us on the products that we're codeveloping with them. And so what that results in is that we get additional leverage on the R&D investment, where the actual investment is actually quite a bit larger than is shown churn on our P&L. If you look at the work that we've done in terms of funding all these investments, when you go back over the last year, the data center has become more than 3/4 of our business. And so when we looked at the remainder of our portfolio, we had heavily invested in bringing those to the leading sort of level of technology. And so the required investments on the rest of our portfolio is frankly just -- we assess that as being lower today. And so what we did is we very actively moved resources and redirected our investments to the data center. So not only do we have the benefit of the NRE, but then we've really optimized our portfolio. And so as we look forward, you should expect us to continue to drive very significant operating leverage as we grow that top line.

Yes. Maybe just to add, I think on top of all that, the NRE is critical. But I think also the company, we've been very thoughtful in our capital allocation over the years. And I think back to when I was interviewing you and we're talking about making the jump to light speed on nanometers and like how are we going to pay for this, and that was always a big worry. And we actually did not make that jump investors problems, right? We drove a lot of operating leverage through the cycle we've been through. We've gotten now on the leading-edge train. We do get this benefit where we have co-investment from our customers to go do this. But we also have a lot of revenue scale we're driving. And so -- and even recently, you keep seeing this trend where we're growing our revenues, but we're growing our operating income at a faster rate.

So we're getting some leverage in the model, but we're clearly prepared to invest with our customers. This is a monster opportunity. And we're going to continue to grow R&D spending in the company. And despite the ups and downs and semis is a cyclical business and the cycles, we all feel, especially the employees of the company when we're going through it. But if you actually look back over the last 9 years, I think we looked at it. I think we've actually grown R&D spending in Marvell every single year since 2016, consistently. Now we've done a lot of reallocating and we've sort of put our bets where the where the future is, and I think that's got us to where we are. So I'm very confident with the customer investment, the leverage we're driving and the way we run our capital allocation in Marvell. I think we can continue to have a compelling financial model, but really be very, very in position to win all the designs that are in front of us and compete at the highest levels on technology and readiness.

All right. The last couple of questions. First one from Srini Pujari in Raymond James. Given you're 1 of full-service custom providers plus all the design wins you articulate today. I guess I'm a little surprised by your 20% share target, why not higher.

Well, who am I going to give that one too. I mean, look, we don't look back and say 20% is so low, but maybe I'll just comment on it since it's such a big question. I think I think it's a journey we're on. And I think just looking at the progress where -- and this customer, as an example, I mean, we're less than like 5% market share, right, just a couple of years ago. When you take it to the data center Inc. level. And even there, we were like 10% market share just a couple of years back before ChatGPT or with the ChatGPT and sort of Gen AI rise, now we're at 13% share. So it's like anything, right? It's a journey that we're on, and I think that's a great benchmark along the way. But we're here for the long game here at Marvell, right? And so it's not a, hey, we get to 20% and then we spike the ball and then we all retire and we're done.

And if we can get there earlier, by the way, because some of these designs we've got in the [ 18 ], I mean, you don't know how big they can be, quite frankly, you don't know on the [ 50 ]. You don't know what some of these new emerging customers what they can do or where the traction can come from. So this is our best estimate. We tend to be -- people could call us conservative. You could call us also just sort of thoughtful and judging the business the right way and giving ourselves some room. But I think it's going to be an absolute home run for employees and investors. If we can achieve our goals there in the data center and drive that kind of a revenue level. And the thing is we've just got to keep executing guys, increasing our market share year in and year out, let the market evolve, and that's going to have its ups and downs, too along the way. But if you look through cycle, it's going to be a big, big market.

And then on top of that, just for a moment to zoom out to the Marvell Inc. level, you've got the core business roaring back as well, right, which is our carrier business, enterprise business, industrial business, those kind of businesses are coming back. So when you add those up and you look at what the opportunity in this company is, it's quite substantial. So I just view the 20% is a good goal, and it's somewhere in that time frame. But I think it only gets bigger if you just look at the momentum that we have.

Great. Last question from Chris Caso at Wolfe. Can you just help investors just understand better what you're counting when you show sockets. For example, are the 5 XPUs you showed? Are any of these follow-on projects or existing customers? Or these are all discrete projects.

Yes. Great question and happy to keep clarifying. So Chris, do you want to take that one?

Yes. So the 18 sockets that we have today, those are sockets, multigenerational in nature, all independent of one another. Now of the 50 opportunities that we're chasing, some of those would be follow-on to the 18. So any -- whatever generation we have at 1 might be part of the 50. But -- and then, of course, obviously, there's a lot of incremental opportunities in there that will be brand new sockets as well.

Perfect. I think that was the last question.

Okay. Yes. That was actually rapid fire. I thought that was great. I got so much help Like, we can do when we all go parallel, right? Yes, exactly. I think next earnings call, I might have a few new friends my conference room. But anyway, just to wrap it up, again, thanks, everybody, for joining us today. Everybody in the room, all the Marvell employees listening and all the investors on the call. We appreciate the interest. I think there was great questions from the analyst community and our investors we're clearly going to have a lot to talk about, and we're happy to go do that. I mean there's a lot of just new things we talked about today, new reveals, bigger opportunity. But again, I just appreciate everybody here and the outstanding job you guys did.

A few final points. I mean if you sort of wrap it up, first, custom silicon, it's a major, major growth engine, not just for Marvell, but if you look in the industry and just pure TAM, I mean, I think you put it in perspective, like the data center silicon TAM out in a few years is going to be as big as the entire semiconductor TAM of everything all in like a year or 2 ago, right? So it's just -- so that's exciting. -- custom silicon is driving a lot of that. And we're in the sweet spot there given how long we've been investing and how long we've been at this, and I think it's a credit to the team that the really smart people in this company and in the room and across Marvell saw it coming. And we were able to start preparing. We didn't know how big it was going to be, but we certainly had an instinct that this is where the puck was going.

The second is we really have established these deep relationships with these very significant important customers. We value those immensely. Like I said, we're all in with our customers. We are going to invest with them to make sure that they're as successful as humanly possible and that they go make it happen. Our portfolio is very broad within custom. If you just look at the capabilities we outlined today. Big moat in terms of what we can bring in and showcase to our customers and leverage those technologies to get them their best solution, but also the fact that we can come in with our interconnect portfolio other product lines to sell to them. And of course, this new emerging XPU attach area, right, which really adds value to their architecture and how they design their systems.

So I think we're just extremely well positioned to capitalize on this $95 billion-ish opportunity for Marvell. And I think there's a huge, huge growth potential in front of us. So anyway, I just want to thank everybody for attending today our senior technical leadership conference. It's a great way to kick it off, isn't it guys? All right. Perfect. Thank you. And I thank everybody for joining and your interest and thanks, everyone, for joining and your interest in Marvell. We'll be in touch. Thank you.

And I'm Vivek Arya from BofA semiconductor team and really delighted and honored to have the management team from Marvell join us this morning, Matt Murphy, CEO; and Willem Meintjes, CFO. I'll go through typical fireside questions, but if you have anything you'd like to bring up, please feel free to raise your hand. But with that, a very warm welcome, Matt and Willem, really appreciate you joining our conference.

Maybe, Matt, let's just give us a state of the union as you see it, a lot of kind of macro cross currents. So how are you seeing the demand environment right now versus what you thought at the start of the year, and then we'll can get through some of the more detailed questions.

Yes. Great. Thanks, Vivek, and it's great to be here. I love coming to the conference. It's good to see everybody here and also whoever is listening in. Yes, a couple of things, maybe big picture to start and then we can go dive as deep as you want. I think the first is, yes, I'm actually -- I'm actually very happy to see where we're at right now, given kind of to your point, all the churn that's gone on in the macro and the global and tariffs and all these things that I think investors and management teams have been concerned about it. I think those concerns are still sort of hanging out there, but the fact of the matter is, we just guided the strongest quarter in the history of the company, and we guided $2 billion, which I was thinking back and Four years ago, in Q2, we guided -- we achieved actually like [ 1076 ] or something, which was the first quarter with Inphi.

And that was like when we crossed the $1 billion mark. And we were cracking the champagne -- per quarter, I'm saying. And it was like how do we get this company up to $4 billion. This is fantastic. And now basically, 4 years later, it's doubled, right, in terms of the growth. And I think from a performance perspective, I couldn't be more pleased. I mean our data center business has been growing now annually at a 70% kind of plus clip. And that was off of a strong year ago compare, right? This wasn't like we had some bottom a year ago. And so those numbers are sort of very, very competitive given our scale and I mean just think about the overall data center business for Marvell is bigger than the entire company was basically a year ago, OldCo.

So from that perspective, I think very, very excited. I think the other thing that's really progressed has been the ramp both of the -- continued ramp both of our connectivity and electro-optics business. I think the asset we purchased from Inphi plus all the good Marvell stuff we put into it has performed very well. And then there was a lot of concern in 2024 about custom and could it ramp or not? And are these programs real? And is the football going to get pulled and can you make it work to -- remember all these concerns is even going to happen, right? And now you look and we started shipping high volume at the end of last year and now through the second half.

And there's multiple custom silicon programs now that are in production, future ones coming to production, and we'll talk about that more at our custom silicon AI investor event in a couple of weeks. But from an overall perspective, I'd say super solid. And then the final one, which is very encouraging is we had this very sort of strong post-pandemic cyclical decline in some of our core business, right, which is carrier infrastructure and enterprise networking, and those bottomed at about $900 million a quarter -- a year kind of run rate where really we think end demand is about $2 billion, something like that. And so now you fast forward and based on our guidance, we're like $1.3 billion to $1.4 billion, so that's on the way up. And it didn't come back.

I think the stuff never happens linearly, right, never. There's no linearity in this sort of cyclical stuff in semis, but it took a little bit longer than we thought, I think, to rebound. But now the last couple of quarters, we've seen kind of 10% plus, even mid-teens sequential growth on that business. So you can see that fall through starting to happen now. Operating margins have expanded. So net-net, we stand here, given all the challenges in the world and all the dynamics and I'm thrilled to have and the team and very proud of the Marvell team for have gotten the company to a revenue scale of $1 billion annualized run rate with a lot of gas in the tank on the go forward.

Absolutely. So Matt, let's just walk through the different segments starting from the custom, right? Custom compute side. On this last earnings call, I think you really gave us a good flavor for how the program is ramping at your largest customer that you have this multi-generational thing. And I noticed that the confidence was a lot more than -- or maybe at least externally, right, you were willing to share a lot more. So what changed, right, in the last 3-plus months to give you that confidence to give us a better forward view of how that program is shaping up.

Yes. I think there's 2 things maybe to decouple. And this I think is the case with any kind of material piece of business or end market that we've gone after. You sort of have your bogey you set or your target. And then -- and are you more or less confident on that? And then there's also, as you -- as time progresses, and you see progress, you can let out a little bit more information, right? It kind of makes sense. So from a confidence standpoint or kind of where we think things are landing, actually, nothing's changed. I mean we had an AI investor event April of last year, and we were getting questions about the multigenerational nature of some of these engagements.

And what I said back then is that that's what we believed to be happening pretty much across the board because that is fundamentally the nature of these types of programs. There's a lot of benefit in doing that. So from that standpoint, actually, our expectation is really nothing has changed. I think as time progressed, there's certainly been an investor appetite to get more updates here. And then I think there's kind of been a unprecedented level of noise and kind of something I've never seen this level of sort of kind of public discussion and news cycle on something that's actually highly proprietary.

So we have to balance that because ultimately, all that really matters here for us to grow the company in the long term is to take care of our customers, right, and make sure that we're really protecting their confidentiality. But at the same time, I understand the need for investors to get as much transparency as possible so you guys can make decisions. So that's the balance I try to strike every quarter. And so there was not a confidence change per se, but I think since things have elapsed another 3 months, and we have more line of sight, we thought it would be helpful to investors to kind of reassure people that we still see strong growth in the coming year and years beyond on a number of these programs where there's been doubt cast about whether they can continue or not, but we still strongly believe and we feel really good.

And we'll talk about at the AI silicon event. I mean, the pipeline of opportunities that we've got across the board, actually, as the whole cloud network ultimately becomes more and more optimized and more and more customized, that's really playing to our favor. And it's not just a couple of sockets and that sort of it. I mean it's actually, I think the breadth and the scale of some of these things will be interesting to talk about. But yes, so far, so good. It's all tracking according to what we found.

I know you -- I assume you will talk a lot more about this at the upcoming AI event. But give us a flavor Matt, how should people think about Marvell versus some of your Asian counterparts, right? What helps you win business? And how is it different than the business model is, right? Because people tend to equate, I think it would be helpful to understand what makes you guys different.

Yes. Yes, happy to do that. And I think we're also planning as part of this event to kind of break it down a little bit more simply for people because, ultimately, if you haven't been doing this for a long time or you're a generalist or you're just trying to figure out what does all this mean? It might not be intuitive. But historically, the way it's worked and the way it's played out so far is that the -- that companies when they want to do their own custom chip. And this could be for AI, it could be for networking, it could be for a consumer. I mean, really anything you want, a lot of the companies that do this in-house on an ASIC, they typically have what's called a front-end design team. So that's the architecture team.

That's the team that's going to actually define the product, do the front-end, design and then those companies don't necessarily want to spend all the money on what's called the back end, which would be like the physical design, place and route, maybe package design. Some of those things that doesn't really make sense to get in-house because you can use I think, very low-cost suppliers typically from Asia, right, Taiwan and China to some extent. And so that whole design process gets augmented, right, which is typically in physical designers called PD.

So that's where that expertise is typically laid and then maybe the license IP from other people or they'll just use that as part of a bigger chip project, but yet they'll work with a big full turnkey supplier like us, right, or one of our competitors. And by full turnkey, I mean, we can do that. We can certainly do the back end. We can do the package design, we can do the manufacturing the test, but we can also contribute all the core IP or at least a lot of it. We can help with architecture. We can assign our front-end design resources. We have a lot of expertise there, and we get hired to do that quite often. And so -- and then we can deliver literally the full chip and underwrite it, underwrite the quality, underwrite the yields, failure analysis it, send FAEs out if it's not working. I mean the whole 9 yards, right?

And so it's kind of 2 models, right? One is roll your own and use a partner to help you with some part of the design that isn't as critical and then find somebody else to do it or go the full turnkey approach. So we don't have a back end only business per se. I mean, where we really add value is when you want to do a very complex, very large die size, complex packaging, all the different challenges you're going to need in terms of signal integrity, power management, package design, cooling, how to yield it, how to do best-in-class sort of test, how to get it to production, like without any failures and no re-spins? And you kind of hear about this now like these big chips.

If you need a respin, I mean it's like 6 to 9 months away at 3-nanometer or 5-nanometer because of the number of the manufacturing cycle time. I mean we've been able to bring up virtually all of our 5-nanometer fleet so far with effectively A 0 or first pass success on massive chips. So that's a time-to-market advantage, and that track record has really helped us with our customers that are like, oh my God, you guys are actually able to execute, which was not something Marvell as a hallmark of historically.

If you remember in the past, right? I mean I joined this company, we were spinning ships 8, 9, 10 times. I'm not joking. I'm not joking for like simple small die sizes. So we've really built a machine to go and execute these programs and do it where it actually works. And final thing I'll say is it's really about risk. I mean you can do anything on your own. How long is it going to take? And is it going to work? And I'd encourage everybody to look back over time and how many times did something pop out of the rumor mill supply chain over the last 5 years about somebody's got some socket with some other -- and it just sort of where did they go? What happened to that one, didn't work.

So we continue to believe and what we see in the market is that there is a place for that model. It's certainly there's going to be revenue generated. There's no question. It's a giant TAM and we size this custom thing at like $40-something billion a year ago and that's only floated up, right, since we -- so this is like an unprecedented TAM. And -- but I -- the way we see it still is that the vast majority of the volume that's going to ship in the future is still going to be from companies that can provide the full total solution.

And I think until such time as this market slows down or something and you have just more time than maybe you'll try to do it a little cheaper on your own. But it's really just a cost issue. It's not a -- but if you really want to execute and use the latest nanometers, latest package and get to market, you're way better off picking Marvell to go do that for you.

One other thing, Matt mentioned during the call was that some customers, given that they have these really joint CapEx, right, and a lot of projects that they could choose multiple paths right? So do you see that dual sourcing on the ASIC side become a trend? Because we have heard your other U.S. peer also talk about their number of projects has gone from 3 to 7, there is potentially more competition in the project, right, that they are in. So do you think this becomes a norm where these large hyperscalers just have multiple basic projects running, and we just have to get used to it.

Yes. I think it's still early to make a call on that or not. I do think if you look at the industry perspective and you filter the signal from the noise on a lot of the information that's out there, it seems like that would be the case. And I think it would be logical to think that, that would be part of the market because to your point, like what we -- I mean I remember when we bought Avera like 5 years ago, right? We -- a home run like ASIC project for any ASIC business, actually, even us or our big peer was like you get like $100 million a year plus -- I mean, this thing was like you're ringing the bell, right? It's a gold mine. They just never -- and now you're talking about these programs that get to like $1 billion a year or more.

And so when you get to that scale and the criticality of these, it just -- it wouldn't surprise me that again, from an industry perspective, there is some bifurcation and there is some multipath that occurs. But I'm not ready to make that call exactly. I mean, we'll see how that all plays out. But just given the scale, I would think that it probably opens up opportunities for multiparty to do things. And I don't think that this is going to end up. And it's really not -- and when we look forward, is we're not seeing this. It's not going to be a one-trick pony type of business. There's not going to be 2 sockets in the world over time that drive all the volume.

And if you have one, you've got and if you don't, you're 0 and I think that's some of the discussion we'll have at the AI event is we see just a really strong breadth of opportunities across the stack. Some of them are bigger than others, but they also carry more risk, right? I mean if you -- whenever you chase the big, big shiny object, usually, there's a lot more risk with it. So I think that's -- there's some interesting trends we're seeing, and we'll talk about that more. But it certainly wouldn't be out of the question that ends up becoming somewhat of a trend, just given the pure scale and the revenue.

But that's -- but does that impact your ability to capture revenue, so you continue to grow even if there is a different player in that same...

No, remember, we're the up and comer, right? I mean, sometimes I appreciate some of the credit we've gotten. But we just started shipping high-volume custom in this area like last year, like it's sort of like this is all still in front of us and I think there's obviously a lot of pessimistic view that sort of our best days are behind us, but we just got started. These are designs we won in the first generation a few years ago. We've delivered them. We've won subsequent generations. Those are in front of us. So from our standpoint, especially given our size and our scale, and you just think about a TAM that was $40-something billion a year ago, that's going to now be a lot higher, and we were like I think last year, well, we overachieved it. But at one point at the AI day last year, I think we said like $500 million of custom revenues like last year, and then we shot over it.

I mean that's a very small percentage of the total TAM. And so if you think that the -- if you see that there's a diversity of opportunities and the direction is heading that way, absolutely. There's no -- these trends are going to happen the way they are, but it doesn't affect our line of sight in terms of the revenue capture we can go get.

Got it. And then I think at last year's the April 24 event, you had also mentioned a third customer, right, that could start to ramp up and during this earnings call, you also mentioned that you now have these wins. So is that something that helps in '26? Is that something that helps in '27? How do you kind of conceptually look at the ramp and contribution from that third customer?

Yes. We -- I think from the beginning, and again, you're trying to call the ball fairly early on this, so we always try to -- so in the case of last year, we talked about the fact that our customer wanted to go into production in '26. And the way I signaled it even back then a year ago was just kind of park it is like late '26, and we'll see how it goes. And we'll see. I mean I think that's still our hope is to do that and certainly getting some revenue in '26. '27 always, we said was probably the bigger year and then it kind of grows from there. But as I said in the call, projects are tracking well, and we have no reason to believe that, that can't happen. And we're all working hard together to make it happen as fast as possible. So -- but that all is -- our view on that is very consistent still from a year back, and now we also have a lot more track record in terms of the time that's lapsed and the design and everything.

Got it. Makes sense. On the electro-optics side, Matt, so size of clusters growing, right, need for connectivity growing. But at the same time, we have also been hearing about incremental competition as people, right, whether it's the move towards CPO, right, whether it is some competitors coming up with DSPs at the 1.6 terabit generation. And Marvell has been such a dominant part of that electro-optics market. So how do you see the competitive landscape shaping up in that industry?

Yes, I'd say a couple of things. Well, first, I mean, we're in great shape. We're in great shape from a competitive standpoint, never been stronger. I recall, and I was following this from the outside looking in, and you can go back in time to, take your time machine back. If you remember when the PAM transition happened, and there was like how many companies were hanging around the hoop, right, that all had some PAM DSP revenue plan that was going to go. I'll add them all up, they're way over 100% of the market. And ultimately, the team at Inphi just really executed well. It's just a top-notch team that now we supercharged with the Marvell engineering teams as well. So the first wave of PAM did great. We've continued that trend, 400-gig, 800-gig, 1.6T.

We've been the clear leader at OFC the last 3 years in terms of technology leadership, whether that's demonstrations or announcements of new products and technologies, most recently now the transition to 400-gig. It's really funny our people in our company, including some of our engineers, they do listen to the earnings call, right? So they got -- they picked up on the question about hey, is Marvell SerDes durable? And is there any problem with it. I can't tell them how, I rate e-mails I got. I mean they're ready to unload. I mean they're just like, where is this coming from? So our technology base optical and electrical SerDes and the leadership we've demonstrated is clearly going to keep us in that top position. Now markets expanded dramatically. I mean the whole transition from NRZ to PAM, that kind of opened up a new TAM because all said you had DSPs in the mix. And now all the ports have happened.

So there has been increased competition. And certainly, there's people today that are still saying there's going to be solutions at 1.6T, just like 800 and 400. So we have to deal with that. But when you're first and you have the incumbency and you've got -- you've got the best products and technology, we're confident in our ability to keep our share and keep growing, and we're very excited about the coming years here. A lot of exciting programs we are involved in.

Got it. Now this is a little bit of a nitpicky kind of short-term question, but I get it, so I'm going to ask you. When people look at Marvell's earnings revisions, right, they seem to be a little more modest than what we -- so year-on-year growth rates are obviously exceptionally strong. But when we look at sort of these quarterly run rates, they seem to be a little more modest than what we see from some of the peers. How do you address that issue that Marvell's business is strong, yet predictable. I know predictable can be a good thing or not a good thing. But how do you address this issue of kind of more modest earnings revisions through the year?

Yes, it's a great one. I think -- I guess I never thought I'd be punished for being consistent, but apparently, that's the -- that is what's happened. But that's how we historically have run the company. I mean, guys, I've been CEO for 9 years, right? And I'm open to change, okay? But just if you look at kind of -- this isn't the last like 3 quarter thing. I mean in general, we've been very focused. And over time, I think, have demonstrated that just consistently managing the business and trying to be very thoughtful about the expectations we set in doing what we say has kind of been the mantra. Certainly, we've been banged in cases where peers at different times have had beaten raises, right? And we sort of deliver in line and then there's a little expectation.

We're in a cycle now where the stock is retreated. People are like, hey, why aren't you beating and raising. But there was a time when we were doing the same thing and outperforming. And I remember during the pandemic, getting a similar type of criticism at times, which is, hey, everybody is like blowing their numbers out. Hey, there's mobile phone companies blowing their numbers out or there's these auto things being their numbers out and then analog guys blowing their numbers out, well look what happened, look what happened.

I mean if you look at -- go track it, go back to prepandemic and draw a revenue line for the analog bucket, what's it done? Fully round-tripped. Some companies have outperformed, some haven't, but look at all the mobile phone stuff, round tripped. Look at the 86 package, right? One did well, one didn't, round tripped. Now the data center focused companies and the ones that executed have just outperformed, like I gave you our numbers, right? We doubled in the last 4 years. I don't think -- so I think along the way, we'll find quarters where people are disappointed because we didn't have the same. But I'm not being defensive, I just think that that's how we run it. And if we get a bluebird, we get a bluebird.

But certainly, I think in the market we're in right now, being raises or -- big raises are sort of appreciated. But long term, I think consistency is what matters. And I try to do that for you guys in terms of being predictable. And also even in long term, new markets we get into, hey, here's automotive, here's what we think it can do. Here's custom silicon, here's what we think it can do. Here's 5G, here's what we think it can do. And we've had, I think, a very good track record of kind of sizing opportunities going after them and meeting or exceeding them over time. And so we'll continue to do that. But I got the feedback. You've been telling me for a long time, bigger beats.

That would be nice. In terms of the recent announcement with NVIDIA on this NVLink Fusion, can you give us a little more background what drove that -- what is driving that partnership? And more importantly, when does it start to actually result in something tangible for Marvell?

Yes. Actually, this discussion is not a new one with them. As a leading provider of custom silicon, I mean I probably have a discussion with them, I don't know, 1 year or 2 ago. I mean, hey, is there a way to kind of co-market our IP together, right? Because when you look at it, all the different pieces of custom compute ultimately are complementary and they're in the same system and network as NVIDIA. And so is there some advantage to actually having interoperability through the scale up by including NVLink with our chips as an example. So the manifestation of that was the announcement that they did of which we were one of the partners, but it's not like -- it's not a new concept, but it's been formalized now, which is good because I think you needed something more concrete.

And we have a lot of customer interest so far. It's very early. This sort of was just released broadly at COMPUTEX. But we're -- we just view it as another addition to our custom platform. It's an offering we have. We're going to do the same thing with other industry standards like UALink run that consortium. We have partners over there. We're doing stuff there. So it's really just about having the broadest kind of most open platform to allow our customers that choice. And if they want to go down that path, we'll fully support it. And we have a long-standing, very good working relationship with NVIDIA across a number of different vectors. And so I think that ability to go partner and really deliver, I'm very confident we would be the best-in-class partner to go do that for sure.

Got it. One other TAM expansion, Matt, yesterday, one of your peers on the active electrical cable side, right, had very strong results and outlook. When does AEC start to become tangible for Marvell?

Yes, it's in the -- yes, so I think they did a great job, and I think it's been further validation of that category, which we've been investing in. And we are participating now. We -- there's revenue in the system today, and there's more opportunities, especially as that business is now is transitioning to a PAM-based approach. Remember, the initial wins in that area and the initial sort of instantiation of AEC was kind of a full solution, which is NRZ plus kind of with one vendor in a cable. So now I think there's just going to be -- there's just demand from the market, right, as more and more of these links get accelerated optically and electrically that you're going to want diversity of choice. And so we've had -- we have a different approach. We've partnered with all the key cable manufacturers to kind of broaden that ecosystem and then we're bringing our best-in-class PAM technology to that field.

So long story short, it's in the revenue today, and we expect more meaningful contributions throughout the year and next year. But I view the performance of the Credo team is they did a great job, and they proved that this category is real. And I think that more broadly, kind of the demand for data center connectivity across the board is strong and that trend is going to continue. And we're going to be -- continue to be with our mission being really the one-stop shop broadest, deepest supplier. So customers have choice on what they want to do. We're not a point solution where we have the breadth and scale to go tackle the most complex tasks.

Got it. Then one thing you mentioned last year at the AI event you had given, I think, a total of like mid-70s total market and then in that, I think 40-plus was on the custom and compute side, since that time, we have only seen the AI market size grow much, much bigger. What do you think is that ASIC versus GPU mix on that. I guess, since you guys are smaller, it doesn't even matter to some extent because you have that headroom for growth. But do you think that ASIC to GPU mix has evolved since you first thought about it.

Maybe you want to take this one, you were driving the Investor Day by yourself.

Yes, I think you're going to have to wait for Gen...

No preview.

Yes, I think this is one of the key messaging. I think just at a higher level, 2 things, right? One is you clearly see a proliferation beyond the top 4, right? And so we'll give you some flavor of that. But also, we've been saying very consistently that customers not just on the compute but it's the entire system. And so we'll unpack that a little bit more. Yes. But I think we have some really exciting things to you.

And we'll date the TAMs to you at that point. Because yes, we did get criticized for low balling I guess, last year, and we were calling out like the $200 billion data center, 70 TAMs and $70 billion this and $40 billion, it was just like, yes, you guys were sandbagging. It's like, okay, no problem. But that's fine. I mean I think it's given us some headroom to watch the market flowed up over the last year. And so I think there's a strong message there about the opportunity only getting bigger. And then to Willem's point, maybe a lot more diversity and optionality for growth from a number of different vectors. We'll update those numbers for you.

On that optimistic note, thank you so much. Really appreciate. Thanks everybody.

Yes, thanks, Vivek. Appreciate it.