Astera Labs Aktie

Hello, everyone. Thank you, everyone. Welcome to day 2 of Citi Tech Conference. My name is Papa Sylla. I'm the U.S. Optical and Internetworking analyst here at Citi. And I'm very pleased to welcome Jitendra Mohan, Co-Founder and CEO of Astera Labs; Desmond Lynch, CFO of Astera. Welcome both.

Thank you.

Thank you.

We'll start with my questions. And towards the end, I will open it up, if the audience has some questions as well. I guess to get started, Jitendra, because Astera is so involved in so many of these server designs, I thought it would be helpful to start by giving perhaps the state of AI from your perspective, but also for Astera Labs, it's been now 2.5 years, I believe, since you IPO-ed. It's been quite a fast and furious journey. So if you can just talk us through the journey and what's ahead?

Yes, definitely fast and furious. You put it well. Thank you, Papa, first of all, and good to see all of you, some familiar faces, some new ones. So let's start with the AI question first. And I believe that we are still early in the multiyear build-out of AI that we are seeing. And I think more than my word, you should look at the CapEx expenditure from the hyperscalers, which is now expected to cross $1 trillion next year. So that's a very good sign that not only are our hyperscalers investing a lot of dollars, but actually the pace of that investment is even increasing. So in many ways, you can see that the tide is rising, and of course, all boats will rise with the tide. But at the same time, I'm also very excited really about the pace of innovation and change that is there in the AI industry. Every 3 months, every 6 months, there is something new that comes up.

Talking about the IPO, when we went IPO, the unit of compute was basically 8 GPUs in HGX board, the 8 Hopper GPUs, that's how ChatGPT movement happened. Now the unit of compute has become a rack. And the reason for that is these models just keep getting bigger and bigger. IPO time, we were talking barely just getting into 1 trillion parameter models. Now we have crossed 10 trillion, and there is no sign of stopping. And you look at the latest Astra model that came out was trained on more than 100,000 GPUs. Inference is driving a lot of new workloads, including longer context windows and so on. And even for inference, now you need racks of compute. So we know the rack has now become a unit of compute.

The other trend that has come up more recently is supply. The demand for compute is so much that it is really outstripping the available supply. So all of our customers are also looking to figure out how to get more out of the hardware that they have available. Fortunately for us, Astera was really founded back in late 2017, we started operations in 2018 on this premise to solve the challenges with AI connectivity, remove the bottlenecks for data, for networking and for the memory bottlenecks. And this is really what we have been doing. Our vision is to provide connectivity infrastructure at the rack level, which includes all of the different components, whether it is switching, signal conditioning components, running over copper, running over optical hardware as well as software. And so we are well on our way to do that. And that has been our journey since IPO.

At the time of IPO, I recall that our revenues are about $60-something million, give or take. And last quarter, we were $392 million. We guided for $550 million at the midpoint for Q3. Similarly, our earnings per share were $0.10 and then last quarter was $0.80. So we have come along quite a distance in there. In fact, I brought something to show you guys just to put this in perspective. Apologies for those on the webcast. But this is where we started. This was our Aries retimer chip. And then this is the Scorpio X that we have just kind of recently introduced. This is the 320-lane Scorpio X. That kind of gives you a little bit of perspective about the journey.

And going back to the point I made about, A, the tide is rising with AI, with the CapEx spend, but we are growing a lot faster simply because we have managed to introduce new product families and capture a lot more content on a per XPU basis. In fact, when we started our content on a per XPU basis was sub-$100, largely driven by the AeresZimer products that are attached to the Hopper GPU. Now we are already at close to $1,000 of content. If you look at the Aries family, the Taurus family, the LEO CXL family and then, of course, the most recently introduced Scorpio family. And every generation of XPUs, we increase our content, especially with the Scorpio X, we are on track to get $1,000 or more content on a per XPU basis from the Scorpio X family alone.

So we are very pleased with where we have come over the years. We will continue to work on new product families that gets us into multiple thousands of dollars of content on a per XPU basis. So really proud of the growth that we've seen, really proud of the team to have come up with so many different products and the strong trust that our customers are placing in us and inviting us to the table years in advance of an actual deployment and helping us define these products and bring them to market.

Yes. No, that's absolutely fascinating. And if I knew you would be showing Scorpio, we'll do a video webcast instead so everybody could see.

It's fine.

But you mentioned kind of Scorpio and perhaps that's where I want to start. It seems like moving forward, this will be probably your most important kind of business. I mean you mentioned this will be a product that will drive kind of content per accelerator beyond $1,000. I'm curious if you can parse through the different types. You have the P version, the X on kind of content, but also the growth outlook of both as we go into '27?

Yes, yes, certainly. So for those of you not familiar, we have 2 series of Scorpio devices. Both are purpose-built for AI applications. So before Scorpio, the PCIe switches were used for general purpose compute applications to connect storage and sometimes networking devices, et cetera. So when we decided to enter this space, we wanted to build something that was purposely done for AI, that understands the AI workloads and so on. So we started with the Scorpio P-Series, which is meant for scale-out connectivity, connecting GPUs to NICs and to CPUs to connect these different clusters together typically over Ethernet. And that was a product that we introduced in October of 2024 at OCP. And then it very quickly became our fastest-growing product line just in 3 quarters last year, it was 15% of our total revenue. So definitely one of our fastest-growing product lines.

And that growth has continued. We've added another hyperscaler customer. We are into multiple customers that are now deploying Scorpio P into their AI infrastructure. Just because of some of the unique things that we built into Scorpio. Scorpio X, especially the 320-lane device we introduced just in last May, and it has just got a tremendous amount of traction from our customers. So just to explain what Scorpio X is, it sits right in the center of our scale-up network, right? As I said earlier, it's connecting all the GPUs together so that it looks like 1 GPU. So from a software perspective, it's very easy to just address, let's say, the 72 GPUs as just 1 big GPU. So a lot of that intelligence goes into this chip.

And on top of that, we've added more features like in-network compute and Hypercast, that really understand the traffic patterns that go for both training as well as inference workloads and essentially make the GPU work better. So a GPU that's connected to a traditional PCIe switch will produce a certain amount of tokens at the end of the day, whereas the same GPU connected to a Scorpio X will produce more tokens per second. Just think about that. That allows our customers to have less infrastructure for a given workload or for the same amount of infrastructure, service a lot more workload. They need at the end of the day, fewer GPUs, fewer racks, fewer nuts and fewer bolts. So it's a phenomenally good device, and we are so thankful to our customers for helping us define and bring this to the market.

We are already shipping this to our lead customer. This device was introduced in Q2 and already entering volume production here in Q3. Is also behind the strong guide that we provided for Q3. So we're very thankful for that. It is shipping to an additional customers. A lot of customers are actively designing this, and we expect to convert many of these to full design wins to ramp in 2027.

Even so, just in Q3, Scorpio will become our largest product family by revenue. That's just incredible. And within the Scorpio family, Scorpio X will overtake Scorpio P. And recall that I just said Scorpio P was our fastest-growing product line. So we are very excited by this milestone. There is a lot more to do, a lot many customer applications to solve. And people are just finding different use cases for Scorpio X now that this is available to them as a toolbox.

Yes. No, no, absolutely. And it seems like at least from last earnings that kind of the expectation that Scorpio will become the main product is almost pulled forward by a quarter or so. I guess I'm curious on what's the driver of that? Is it kind of a faster conversion of qualification or design win into revenue? Is it kind of market share gains? Or is it just the TAM overall of the intra-networking ramping faster?

No, we are very happy with the progress that we had in the last 18 months or so with the Scorpio family. One of the reasons that this is ramping so fast is actually not something that happened now, but started many months, even years ago with this extreme amount of co-design that we did with the lead customer on both defining the product, simulating it, emulating it, just a team doing intense amount of planning even before the chip taped out and certainly before the chip showed up. And then after that, the teams on both sides, the customer as well as at Astera just did an incredible job of bringing up this complex chip and then a complex system in really record time. So I do want to give a shout out to the team for just the immense lift that they provided in getting this complex system up and running. And now it is getting deployed, and we are starting to see the fruits of their labor.

In general, we do try to be conservative when it comes to the amount of time it might take to qualify a system in production. We control -- certainly our chip, we control a few things, but we don't control everything. So there is certainly some amount of, call it, just old-fashioned luck that everything landed in the right way, and we were able to ramp this as quickly as we have been able to.

Right. And then quickly, one for you, Jitendra, and I have one for Desmond as well. For both P and X, I'm curious if you can frame the competition there. I think in P, you are more of a kind of coming into a market, you found a player there, but it seems like you're doing really well, while X seems to be kind of its own domain, although there are other kind of proprietary scale up. I'm just curious if you can discuss a little bit the competitive dynamics of both Scorpio products.

Yes, yes, certainly, certainly happy to do that. So before we got on the scene, the dominant switch supplier was actually Broadcom with the PCIe Gen 5 switch products. What we did when we entered the space is we decided to focus on AI. We did not have any baggage. We did not have other customers to support other applications to support. So we purpose-built the Scorpio family, both P as well as X for AI. Now folks have realized since then that this is a very big TAM. Both for Scorpio P was already always there, but also for Scorpio X. So it is understandable that the competition will come. And we welcome that. There is a lot of demand. It's not a zero-sum game.

Having said that, the fact that these chips were purpose-built for AI have all these different feature set that is built into the chips, there is really this help us stand apart. If you look at Scorpio P first, we are shipping this in the highest volume for PCI Express Gen 6. So we are in a very good position. I feel very confident where we are with our Scorpio P series. And then Scorpio X is just getting started. And at the root of all of this is the trust that our customers have placed in us in helping us understand what the applications are, both for Scorpio P and for Scorpio X. And therefore, giving us the direction on what features to build, how to build it and so on.

So first of all, very thankful to our customers for that. Secondly, as I've mentioned now 3 times, these are just custom built for AI. And just think about it one more time that a GPU attached to a Scorpio X just does better tokens per second than a GPU attached to a non-Scorpio X device. So that's just a phenomenal. I mean in today's supply-constrained world, that is gold. If you can get more tokens per second from your infrastructure. So I think that's fantastic. We are already shipping this to multiple customers, just a strong amount of interest that we have from our customers to design this in. So really looking forward to deploying this at scale, not only this year, but then ramping into next year. We are just getting started with a multi-quarter deployment for Scorpio X, multiple applications within the lead customer and then many more customers.

Right. No, that makes sense. And Des, I'm curious on when we think about gross margin, and you gave out the longer-term kind of guide of eventually around 70%, but the company is hovering around low 70s, low to mid-70s. I guess when we think of Scorpio X coming in, it seems to be a very kind of unique product, but also I assume part of it will probably early on as the ramp starts, the gross margin for the product will continue to ramp as well. But just what would be the impact of kind of these products coming in, in late '26, '27? Is the risk to the 70% gross margin on the upside or perhaps on the downside?

Yes. We've been very pleased with our performance on the gross margin side, where we've been operating around the mid-70% sort of range. What we have indicated to investors is that over time, we do see the margins trending down towards 70%, which is in line with our long-term target.

I would say a couple of factors here would be on the product mix side. Firstly would be the mix of silicon versus module content. And secondly, it relates to the Scorpio X sort of series contributions there. And if we drill into the sort of Scorpio X, what we're really seeing is a wide range of sort of margins. We have a broad sort of portfolio today. And the margin profile will really be dictated by use case as well as lane sort of configuration. But overall, I would say that our Scorpio margins are in line with our corporate averages. And in any given quarter, depending upon what's shipping, the margins can sort of move around from there.

But it is important to remember that we are going after large opportunities. And what you do see is a significant fall-through to gross margin dollars as well as operating sort of leverage from there. But I would say, overall, I would say that our margins will trend down towards the 70% range, which is very healthy for a product semiconductor business, and it's in line with our overall expectations.

Got it. No, that's very helpful. And coming back to you, maybe, Jitendra, on slash, maybe moving to a different topic, kind of UALink switch perhaps one of the key kind of server that you would be potentially kind of getting more content, the Helios server as well. So to my understanding, at least the first Helios server would be kind of UALink on top of Ethernet. So the fabric is really Ethernet, but is your understanding as we move kind of generation to generation, it will end up being fully UALink? Or do you see kind of 2 paths kind of UALink on one side and some type of a kind of Ethernet version of it?

So first of all, I think we should not like talk on behalf of AMD. So I'll kind of reiterate what they have stated publicly. So first and foremost, you're correct that for the MI450 series, they are going to go with UALink running over Ethernet. And that has really got everything to do with the fact that UALink switches are not yet available. So in order to deploy MI450 series, they have to use UALink that is tunneled over Ethernet.

As we go into future generations, what AMD has stated publicly is they believe that UALink is the highest performance protocol on the planet. And the reason UALink is the highest performance is because it is a memory semantics-based protocol, which gives you the low latency and the ease of access of multiple GPUs, but also gives you the fast speeds of Ethernet. Now they've also stated that they will continue to offer both options, running UALink over Ethernet as well as running UALink natively depending upon where the customers are. So I think that's probably a fair way to say where AMD is. So we do think that both of these will continue to coexist for a while.

No, that makes sense. And in terms of the content you have with that particular customer, I'm sure it doesn't go from 0 to a lot of content when we get to UALink. You already have some content prior to Helios at this early version of Helios and eventually with UALink. Can you perhaps just kind of contrast for us how content can potentially grow from prior to Helios to ultimately a UALink version of it?

Yes. So again, I can't comment specifically on the Helios platform. That's AMD specific. But at the high level, what you can look at is what is our attach rate when the fabric is something that we support, whether it's PCI Express or UALink versus its fabric where we do not support a switching content. So in the case where we don't support a switching content, our content is typically focused on retimers, whether they are PCI Express retimers on the scale-out side, Scorpio P series, again, on the scale-out side as well as Ethernet Taurus retimers or active electrical cables on the scale-out side. So we are more limited to scale-out, and that's the content or the TAM that is available to us.

Now when you have a scale-up protocol that we do support, whether it's PCI Express or UALink or something in between, then our content really does explode a lot because scale-up is a very rich area for connectivity solutions. So as and when we have versions of Helios or other systems that deploy UALink natively or, let's say, a PCI Express-based protocol natively where we can have a Scorpio X solution, then our content grows very meaningfully.

And then on UALink, I think you were kind of one of the companies or the company that led the discussion there. I'm just curious on engagement there as well. But recently, we heard many companies also aiming to do some type of UALink switch or kind of jumping into the UALink bandwagon. So I'm just curious on, for investors and sell-side analysts, what should we look out for to try to determine who will lead this race moving forward?

Yes. I think maybe I'll answer 2 ways. The first is to look at UALink itself. So UALink, we believe, is going to be an evolution of other memory semantic-based protocols like PCI Express. So while there is a lot of focus on UALink as there should be, there is a lot more to it than the protocol itself. The things we talked about earlier, in-network compute, Hypercast, these are all of the feature sets that are really important for GPUs to deliver high utilization. And we will carry forward all of this knowledge from our current generation of Scorpio X to the one that we do for UALink. Diagnostics, telemetry, all of these things are extremely important, and we offer all of that through COSMOS. To somebody who's already using our Scorpio X and therefore, using our COSMOS software, it just becomes very easy for them to move from Scorpio X to a UALink-based Scorpio X.

However, if I kind of level this conversation up a little bit, if you look at the overall scale-up TAM, we have proprietary protocols like NVIDIA's NVLink. We have memory semantic-based protocols like PCI Express and UALink and then we have Ethernet, which is a messaging protocol. Our vision, our goal over time is to address the full scale-up TAM. We are already now active with the NVLink Fusion. So I think that is going very well. We are active with PCI Express. You're seeing the results of that now. We are working on UALink as well. And over time, as our customers demand it, we will address the rest of the TAM as well because, again, protocol is one thing, but understanding how a scale-up network works, deploying this at scale, taking those learnings and applying them into new generations will really determine who gets what share of this overall big market.

Got it. And maybe switching to a different aspect of your business that I think a lot of investors are very excited about is the optical side of the aiXscale acquisition. To my understanding, the early kind of sales or the early products are more kind of the fiber kind of type of products, but you will also have or you have the optical engine as well. I'm curious on how you see that optical business in terms of both growth and how large you see it being. But also what's your strategic approach to that business? Is it mostly kind of to support, let's say, your Scorpio portfolio? Or do you aim for that business to be a stand-alone business on its own?

Yes. Great question, Papa. So the optical TAM is actually very large, tens of billions of dollars in TAM that we are able to go after. The strategy that we have laid out for our optical is really straightforward. We will start with our optical journey. We've actually already started the optical journey 2 years ago when we put a team together to address this. But from a revenue perspective, that will happen in 2027, starting with component level sales. And these component level sales will come in the form of the connector, which we get from the acquisition of a company called aiXscale that we acquired in December of last year. So they build this connector that is really very good for CPU applications in particular. And some of our customers are trying to qualify that component for use in their own CPU applications for scale-out applications. So that might very well end up becoming the first revenue that we see in 2027.

That will then lead into NPO. So at the 200-gig data rate, we believe NPO is a very good solution that does not stress the supply chain, but also gives you the distance and the density that you need for scale-up networks. So we are fully ready to participate in the NPO side of the house with optical for the 200 gig per lane data rates. And then as we go to 400 gig per lane data rates and even higher densities, that's where CPO comes in. And this is where we will have an optical engine that uses an electrical IC that is done by Astera Labs, a photonics IC or PIC that is also done by Astera Labs as well as the packaging technology to attach the fiber to the silicon photonics also offered by Astera Labs. So we'll have a full optical engine, which will reside not only on the Scorpio, it becomes an optically enabled Scorpio SKU, if you will, but also on the XPU side.

So on a link where previously there was no such component, you get 2 components. So that's a very, very big opportunity and will drive our content on a per XPU basis into multiple thousands of dollars. And the good news is we're already well on our way in order towards delivering this full solution, starting with component level sales, then going into NPO and then eventually to CPO.

The last thing that I'll mention here is that while we talk about optics usually sort of in isolation as a media choice for carrying this data, having a switching platform is really key to deciding for customers to decide which optical solution or in general, which signal conditioning solution they will use. So for us, having this switching platform with Scorpio X and then expanding into multiple different protocols is going to be really key to also influence a full end-to-end link, both for copper as well as for optical.

Got it, that's very helpful. I have a couple of more questions, but I want to pause for a second and maybe open it up for, if the public has some questions. Maybe just give the mic.

Can I quickly ask about customer concentration? As you grow your Scorpio line very quickly and obviously, some customers, they will ramp up faster than others. So what's the sort of medium-term plan in terms of your breakout into different customers? Or is your goal more like concentrated than others?

We continue to be pleased with our customer engagements. What I would say is that across all of our sort of portfolio today, we are shipping to all of the U.S.-based hyperscalers. But given the end markets that we're in of data center and AI, customer concentration is to be expected from here. But what we see is a continued diversification sort of efforts across the portfolio. Jitendra touched upon Scorpio P engagements. Again, that's led by our sort of lead customer. But what we're seeing is hyperscaler customers coming into the mix as well as AI infrastructure providers there. So we're seeing the diversification efforts take place there, and that will continue into 2027 and beyond.

As we look at Scorpio X-Series as well, what we've talked about initially was having 10 customer engagements. That number now today is up into the high teens, and we expect to convert several of these opportunities into design wins in -- by the end of the year and going into '27, we'll see that continued revenue from there.

Also on Taurus and our LEO product families, we see diversification as well. So we're really pleased with how the customer diversification is taking place and the engagements that we have with customers, and we'll see that contributing into the model going into 2027. But it is important to note that our lead customer will also continue to grow as well. So overall, we're very pleased across the entire portfolio of the diversification efforts that we continue to see.

A quick question. In terms of scale-up domain going past one rack+ to multiple, how do you view that in terms of growing opportunity on a per XPU basis or just as a total?

Yes, that's a great question. So it is exactly what I was talking about the optical side earlier. We do expect that optical and copper will coexist. And in particular, what that means is at the rack level, within the rack, the connections will predominantly remain copper. They are copper already today at 200 gigabit per second, but even at 400, we believe that there is a way to keep those connections copper.

And then when you go from one rack to another rack, as cluster sizes increase and your people are no longer satisfied with 72 GPUs in a rack and you want to have 144, 288, even 576 GPUs in a cluster, you have to go to multiple racks. And when you go to multiple racks, as physics gets in the way and copper will not cut it at these increased data rates. So then we go into optical solutions, starting with NPO. And our content goes up very significantly for those links because you still have the switching content that we had before. That does not change. But now we introduced 2 NPO devices, one on either end of the link. So this is actually a very exciting development, and we will start to see initial revenues in the back half of 2027 from our NPO efforts and then full year ahead in 2028.

Yes, yes, of course, switching goes up as well. Switching goes up as well because now the cluster size has gone up and so the radix of the switch goes up. There is more data rates are going up. The number of GPUs connected to each switch is going up. The advanced features that we talked about, they are there. So roughly speaking, that translates into higher ASP on a per lane basis as the radix of the switch increases. Contrary to popular belief that price per lane goes down, it actually goes up with radix.

Maybe I will kind of pursue with some additional question. I think you mentioned also the custom solutions part of your business, let's say, like NVLink Fusion. I'm curious on that part, how large do you see that opportunity? And also, I guess, as you get perhaps more NVLink Fusion coming in, does that take away, I guess, from other parts of your business, let's say, the Scorpio X or UALink?

Yes. No, not at all. So first of all, custom is a very big market opportunity. It is tremendous. Just look at the amount of revenues that Broadcom has and Marvell has and others. We are not going after all of it. We are very selective in the parts of custom that we are going after, but we are very excited with the opportunities that we have. We've talked about 2 of those. One is a custom -- the most recent one is actually a custom opportunity for LEO CXL to get deployed in a KV Cache offload application in AI. So very excited about that. That is in qualification now and will start to ramp in 2027. And that hopefully will open many more doors for similar applications at other customers as well.

And then the second one is NVLink Fusion. With NVLink Fusion, we are able to help our customers take their XPUs and attach them to NVIDIA's NVL72 or MGX platform. So that's an incremental TAM for us. There are some cases where the alternate, for example, in case of, let's say, AWS, as they have said, they will have a native solution with a UAL switch as well as an NVLink Fusion solution. So with our NVLink Fusion solution, the total revenue opportunity on a per rack basis is about the same, given the differences in attach rate and ASP. So maybe that's -- it's a wash. But if you look at other opportunities that we are finding customers wanting to use the NVIDIA infrastructure to deploy their XPU, that becomes net positive incremental TAM that is available to us.

But more importantly, it is a great sign of NVIDIA trusting us to deploy the NVLink infrastructure with other customers. So I think when you have that level of trust with the biggest customers, certainly good things will happen. And like I said earlier, our goal is to play in all of these different ecosystems, NVIDIA ecosystem, PCIe, UALink ecosystem and others over time.

Got it. No, that's very clear. And I'm curious, maybe this is a high-level question. In terms of Agentic, when we think of Agentic AI and when we think of kind of CPU racks and kind of combining CPU and GPU racks, I guess, earlier on, there was kind of questions around the retimer content ratio. But now with this Agentic AI, it seems like that ratio could even go up in those type of developments. So I'm just curious on Agentic AI, how can you tie Agentic AI, I guess, to your opportunities in the retimer and other products?

Yes, absolutely, absolutely, not just the retimer, by the way. Of course, it's an opportunity for retimer, but actually it's bigger than that. If you look at an Agentic system where people are trying to have, our customers are trying to have different ratios of CPUs to GPUs and disaggregate GPU and CPU or create a composable system, PCI Express just naturally emerges as a protocol of choice. Because all of these components, CPUs, GPUs, storage, NICs, they all support PCIe natively, and it is designed for exactly this purpose. And we have the largest portfolio of PCI Express devices, starting with switches and signal conditioning like retimers, active electrical cables, active optical cables, we can run PCI Express over optics, gearboxes. So we are a one-stop shop.

LEO, I forgot about the CXL. So when you want to attach more memory, you attach it over LEO. So we have a full portfolio for our customers to optimize for Agentic. And it creates a large opportunity for all of our devices, which we are happy about, in particular, of course, retimers because as these systems expand multiple racks, the reach requirement goes up, data rates are going up, and that gives us additional opportunity for retimers as well. But it's an overall broad portfolio story.

Got it. And I know we are getting close to time, but just 2 more questions, one for Des and one to finish up with you, Jitendra. So I guess with the balance sheet kind of, it seems like now the cash lockup expiration, we are kind of getting past that period. So just curious on how you think about cash allocations kind of moving forward? Is it mostly kind of reinvesting in the business, R&D? Or perhaps could we think of investor kind of returning cash to investors there?

Yes. We're very fortunate to have a robust balance sheet with continued strong cash generation and our approach to capital allocation really centers around organic investments as well as inorganic investments. Organically, we've done a really nice job of expanding the portfolio, as Jitendra has outlined today, growing from retimer sort of business all the way through to the switching sort of portfolio, and that's increased our overall sort of market opportunity there. We'll continue to invest at the right rate in the business to go after these long-term opportunities, and we're making investments in optical and custom solutions, which will continue to grow from there.

Inorganically, we've made some smaller acquisitions, and that's something we'll continue to look at going forward. aiXscale was a great example of that, where we're able to bring in some of the glass coupler technology in-house, and that strategically is going to be important for our optical strategy from there. And we've also done some smaller acquisitions or acqui-hires, which has brought skilled design teams in-house to the company. And that's something we'll continue looking at sort of going forward. But I would say our same playbook would sort of play out. We do expect to continue to invest both organically and inorganically in the business, and that will be our focus from a capital allocation perspective.

Got it. Very helpful. And Jitendra, maybe to finish it off, any part of the Astera story that you believe is still underappreciated you wanted to address? Or any question I didn't ask that you want to touch on?

Yes. Just briefly, I understand that we are over time. But just briefly, I think the part that's not very well understood is this is a multigenerational opportunity that we have to not only benefit from the rising tide of AI, but to just gather more and more content on a per XPU basis as we come up with these different product lines. Already, our content is now up to $1,000 per XPU. With the new Scorpio X devices, we get to $1,000 per XPU on Scorpio X alone. And then to add to that, the optical multiple tens of billions of dollars of TAM with optical. So we are looking at a significant growth opportunity that we have with our products. That's one.

And the second part is just to reiterate what I said earlier, once you have a switching platform, that gives you a significant influence over what happens in the rest of the system. So once you have a switching platform, really, it's yours to lose for when it comes to signal conditioning, whether it is running over copper or over optical. Over time, we are very pleased to be providing solutions that scale from chips to hardware to software to allow our customers to really get the best out of their AI infrastructure.

Got it. Thank you so much, Jitendra.

Thank you.

Thank you, Des. Thank you for joining us.

Thank you, Papa.

Good afternoon. My name is Holly, and I will be your conference operator today. At this time, I would like to welcome everyone to the Astera Labs Q2 2026 Earnings Conference Call. [Operator Instructions] I will now turn the call over to Leslie Green, Investor Relations for Astera Labs. Leslie, you may begin.

Thank you, Holly, and good afternoon, everyone, and welcome to the Astera Labs Second Quarter 2026 Earnings Conference Call. Joining us on the call today is Jitendra Mohan, Chief Executive Officer and Co-Founder; Sanjay Gajendra, President and Chief Operating Officer and Co-Founder; and Desmond Lynch, Chief Financial Officer. .

Before we get started, I would like to remind everyone that certain comments made in this call today may include forward-looking statements regarding, among other things, expected future financial results, strategies and plans future operations and the markets in which we operate.

These forward-looking statements reflect management's current beliefs, expectations and assumptions about future events, which are inherently subject to risks and uncertainties and that are discussed in detail in today's earnings release and the periodic reports and filings we file from time to time with the SEC, including the risks set forth in our most recent annual report on Form 10-K.

It is not possible for the company's management to predict all risks and uncertainties that could have an impact on these forward-looking statements or the extent to which any factor or combination of factors may cause actual results to differ materially from those contained in any forward-looking statements.

In light of these risks and uncertainties and assumptions, the results, events or circumstances reflected in the forward-looking statements discussed during this call may not occur, and actual results could differ materially from those anticipated or implied.

All of our statements are based on information available to management as of today, and the company undertakes no obligation to update such statements after the call, except as required by law. Also during this call, we will refer to certain non-GAAP financial measures, which we consider to be important measures of the company's performance.

For example, the overview of our Q2 financial results and Q3 financial guidance refer to various non-GAAP financial measures. These non-GAAP financial measures are provided in addition to and not as a substitute for financial results prepared in accordance with U.S. GAAP.

A discussion of why we use non-GAAP financial measures and reconciliations between our GAAP and non-GAAP financial measures and outlook are available in the earnings release we issued today, which can be accessed through the Investor Relations portion of our website.

With that, I would like to turn the call over to Jitendra Mohan, CEO of tester Labs. Jitendra?

Thank you, Leslie. Good afternoon, everyone, and thanks for joining our second quarter conference call for fiscal year 2020. Today, I'll provide an update on AI infrastructure market trends, our Q2 results and recent announcements. I'll then turn the call over to Sanjay to discuss Astera Lab's long-term growth profile. Des will then cover our Q2 financial results and Q3 guidance.

Since our last earnings call, investment in AI infrastructure has continued to accelerate. Multiple industry forecasts now point to hyperscaler AI and cloud infrastructure spending in 2027 reaching into the $1 trillion range with sovereign AI initiatives, inference workloads and enterprise adoption expanding globally.

These secular trends are increasing the need for intelligent connectivity and expanding our long-term growth opportunity. Astera Lab's delivered outstanding results in Q2 with record revenue of $392.4 million up 27% sequentially and up 104% year-over-year. This strong performance was driven by broad-based strength across our entire product portfolio, reflecting the diversification of our business as we continue to win new designs across multiple customers and product categories.

[indiscernible] momentum further accelerated in Q2, representing more than 50% of our total revenue driven by both Scorpio AI Fabric switches and [ Ares retimers. ] This milestone demonstrates the maturity and scale of our Gen 6 portfolio and clear market leadership.

On the AI Fabric front, Scorpio delivered significant growth in Q2. High-radix Corp X series has entered volume production and will continue to scale materially in the second half of this year. We expect Scorpio to become our largest product family in Q3, which is one quarter ahead of our prior expectations, marking an important strategic milestone.

We are now shipping multiple configurations of Scorpio-X Series into scale-up applications to our initial customers, and we remain on track to ship Scorpio-X series to additional customers by year-end. Scorpio-P series also continued to expand across multiple customers with new shipments to hyperscalers and AI infrastructure providers.

These programs are currently in preproduction and are expected to ramp more meaningfully in 2027. We're also closely engaged with our customers to support their future system architectures with purpose-built AI fabric, including next-generation PCI Express and UALink protocols.

Upcoming AI systems will employ more complex switching topologies to enable larger XPU cluster sizes and higher bandwidth. These designs are expected to drive a material increase in our silicon dollar content per XPU. To summarize, our Scorpio fabric switch family is well positioned for a multiyear growth cycle driven by the falling factors.

First, higher dollar content per XPU due to better attach rate and increased capabilities. Second, expanding customer base with exceptional traction in designing pipeline; third, a greenfield TAM growing rapidly with broadening use cases across training and inferencing.

Turning to our signal conditioning business. ADS delivered record quarterly revenue in Q2, driven by strong adoption of retimers and gearboxes across scale-up and scale-out topologies in AI and general-purpose compute platforms. AC's growth is being fueled by increased AI system deployments, expanding AI accelerator cluster sizes, greater unit attach rates and higher ASPs associated with the transition to PCIA-6.

Taurus also provided strong revenue growth in Q2 with increased units shipped across a variety of AI and general purpose platforms. During the quarter, we also shipped preproduction volume of our 100 gig per lane Taurus smart cable modules for 800-gig ACs for scale-out applications.

We expect Taurus to continue its growth trajectory in the second half of the year as 800 gig deployments expand. Driving improved attach rates and higher ASPs compared to 400 gig solutions. We also expanded our Taurus portfolio with a new family of 3.2 smart retimers and smart re-drivers supporting 200 gig per link for next-generation Ethernet and ULN connectivity.

These are the industry's first solutions to enable smart swap, giving platform designers the flexibility to choose between low-power smart re-drivers and long-reach smart retimers without requiring a board redesign. This architecture of flexibility is critical as AI systems evolve rapidly and designers need to address increasing signal integrity challenges at higher data rates while optimizing system power consumption.

We expect our new high-speed 200 gig per lane retimers and re-drivers to double our Talus portfolio market opportunity to over $4 billion by 2030. In CHL, we are seeing renewed momentum as industry is recognizing its potential to unlock performance and utilization benefits in agent AI applications.

For a growing set of memory-intensive inference in agentic AI workloads, memory capacity and utilization are increasingly becoming system constraints and CSL attached memory can offer attractive cost performance relative to local HBM.

As our hyperscaler customers look to address current memory supply and pricing dynamics, we are also seeing additional opportunities for CXL in mainstream general-purpose server platforms. During Q2, we closed a new design win with our standard Leo memory controller at a U.S. hyperscaler.

Looking into 2027, we expect to ship both standard and custom OCSL memory controllers in volume across general purpose compute and AI inferencing applications to 2 U.S. hyperscalers. Activity and engagement within the CXL ecosystem continues to be robust, and we expect to convert additional designs in the coming quarters.

In closing, I'm out of the results we have delivered and what excites me most is the trajectory ahead for Astera Labs. Score Series is in volume production, and we expect our Scotia family to become our largest product category by revenue in Q3, a significant milestone in our evolution as a company.

At the same time, new and existing designs across Ares, Taras and LEO are contributing to our broad-based growth. Intelligent connectivity is now fundamental to your infrastructure. and we are winning with the customers and platforms that define the market.

With 2-1 execution, deep customer partnerships and growing design win visibility, we are confident in our ability to continue outgrowing the market. With that, let me turn the call over to our President and CEO, Sanjay Gajendra to provide more details on our product momentum and strategic initiatives.

Thanks, Jitendra, and good afternoon. I'll walk through our product institution and the expanding market dynamics that are positioning Astera Labs to capture a significantly larger share of AI infrastructure spend over the coming years.

Our laser focus on execution and innovation has delivered a purpose-built intelligent connectivity platform to our hyperscaler customers and is solving sophisticated technology challenges ignite aggressive time lines. This results-oriented mindset has helped us strengthen the trust that our customers and partners place in us.

As a result, we continue to unlock many new opportunities in next-generation designs across multiple customers to rapidly expand our top line growth. Let me deep dive into some of the key product categories, starting with AI fabrics.

Our broad portfolio of Scorpio switching solutions is perfectly suited to address the increasing complexity and bandwidth requirements in AI architectures. With leading hyperscalers planning large-scale training and inference deployments with ever-growing accelerator cluster sizes, our Scorpio Series high ratings fabric switch family provides intelligent AI fabric with industry-leading features to maximize total economics.

These capabilities are driving exceptional demand for our Scorpio smart fibrin switches to address the $20 billion merchant scale-up switching TAM. Scorpio exceeded differentiation comes from its intelligence layer. Hardware accelerated hyper cast and in-network compute capabilities delivered up to 2x improvement in collective performance for large-scale training and inference workloads.

These capabilities are implemented through our Cosmos software platform, which now extends beyond basic fleet management functionality to enable dynamic traffic shaping and advanced real-time performance management. And we are not stopping here, drawing on our multigenerational partnerships with customers and the lessons learned from deployments at scale we are building an ambitious road map for our scale-up smart fabric switches.

Our Scorpio road map is closely aligned with our customers' next-generation XPU architectures enabling more complex topologies, larger cluster sizes and higher bandwidth. Looking ahead, we expect the content opportunity for Scorpio-X series solutions alone to grow well beyond $1,000 per XPU in future generations of AI platforms.

Scorpio-P series momentum also remains very strong. with both early stage and production deployments expanding across multiple AI platforms. We are shipping Scorpio to several customers to enable high-speed connectivity across AI networking applications.

Over the long term, we aim to proliferate Scorpio T-Series across a broader set of applications spanning modular AI servers, disaggregated KV cash appliances, storage platforms and enterprise level service. On single conditioning, Aries delivered record quarterly performance in Q2 and is well positioned for sustained growth over the long term.

PCI, re-timers and gearboxes have become critical infrastructure for AI servers where single integrity directly determines whether systems achieve rated performance. Aries established itself as a gold standard across major XPU and CPU platforms for PCIe 5, and this position has continued for PCIe 6, we continue to invest in our customers' technology road maps, and we expect to have PTIE7 solutions ready to meet their time lines.

Taurus is also expected to deliver strong long-term growth propelled by 800 gig deployments for AI scale-out networks in the near term and 1.6T and 3.2T-based systems over the long term. We are pleased to report that Taurus has entered preproduction ramp for 400-gig per lane in Q2, which will support platforms utilizing 800-gig links.

Our Taurus portfolio expansion supporting 200-gig per lane for 1.60 and beyond, positions our road map through the next Ethernet generation and the forthcoming ramp of UALink driving continued content gains and above industry growth for the tollers product family.

Moving to UALink. The consortium's 2.0 specification validates our architectural approach. UALink 2.0 codifies in-network compute, advanced RAS features and 200-gig service as next-generation scale-up fabric requirements. These directly align with our investment areas.

We remain engaged with lead customers and ecosystem partners to enable UALink deployments supported by our upcoming fabric and single conditioning solutions. We are on track to deliver UALink enabled Scorpio-X Series sub switches in 2027. Optical connectivity represents a substantial expansion vector with the potential to unlock tens of billions of dollars of additional market opportunity beyond our current corporate connectivity TAM.

Over the past 2 years, we have assembled capabilities spanning analog and mixed-signal design, DSP, electronic and photonic IC integration and optical packaging through organic investment and the [ xScale ] acquisition.

This enabled us to deliver complete optical engines in addition to optical components. Our optical strategy unfolds in 3 phases: first, high-density fiber attach solutions, leveraging our ex-scale core technology enables up to 50-meter fiber connectivity with ultra-low latency.

Our initial opportunity with the Tier 1 AI platform provider is expected to enter volume production in 2027. Second, near packaged optics or NPO chipsets enabling multi-rack scale-up closes via optical links are also targeted for 2027 production.

Third, fully integrated Scorpio-X Series fabric switches with CPO optical engine module represents our longer-term road map for 2028 and beyond supporting larger fabric domains with terabytes per second optical I/O per switch. Beyond standard products, we continue to make solid progress with our custom solutions, the increasingly diverse XPU landscape and workloads are driving demand for custom connectivity solutions.

We are executing across multiple vectors with custom solutions for hybrid drag architecture, including envying fusion-based designs and separately custom LEO memory controllers in specialized interface applications. We expect initial shipments for these custom design wins to commence in 2027. We're also engaged with multiple customers on additional high-value custom connectivity opportunity to deliver end-to-end purposeful silicon solutions.

Together, these design wins and active opportunities validate our capability to deliver differentiated custom silicon and represents a multibillion dollar incremental opportunity over time. Overall, our platform philosophy, which we call AI or way reflects our strategy rather than prescribing a single architecture approach we provide the broadest portfolio of connectivity solutions that let customers deploy AI infrastructure optimized for their specific workloads, whether that's standard spaced or custom or copper or optical interconnects and scale up or scale out connectivity across a plethora of standards.

This architectural flexibility is becoming a competitive advantage as the market diversifies. To summarize, Astera Labs has hit another important inflection point with Scorpio high-raiI fabrics reaching volume production. This marks our evolution to a complete AI fabric infrastructure provider extending beyond signal conditioning and memory connectivity solutions.

Our expansion into optical interconnects and custom solutions further broadens our addressable market in 2027 and beyond. With that, let me turn over the call to Des to walk through our Q2 financial results and Q3 guidance in more detail.

Thank you, Sanjay, and good afternoon, everyone. Today, I will review our Q2 financial results and then discuss our Q3 guidance, both presented on a non-GAAP basis. We delivered outstanding second quarter results with record revenue of $392.4 million, up 27% sequentially and up 104% year-over-year.

Growth was broad-based across our AI fabric and signal conditioning portfolios led by PCIE6 products and the initial production ramp of Scorpio. Our Scorpio product family delivered a strong Q2. Scorpio-P Series momentum continued during the quarter, driven by the scaling of existing customer deployments and the early ramp of new programs into production.

Scorpio-X Series also grew strongly as we began shipping initial production volumes across a variety of radix configurations. Aries revenue reached a new quarterly record, driven by strong PCIe 6 retimer adoption across both scale-up and scale-out applications. Overall, revenues from PCIe 6 across our AI fabric and signal conditioning portfolios represented more than 50% of total company revenue in Q2, up from 1/3 in Q1. [indiscernible] revenue also grew strongly in Q2, driven by increased unit shipments across both AI and general purpose compute platforms.

Non-GAAP gross margin for the second quarter was 73.7%, above our guidance of 73%. Non-GAAP operating expenses for the second quarter were $135.8 million, up 10% sequentially, reflecting continued investment in our road map.

Non-GAAP operating margin for the second quarter was 39.1%, up 290 basis points from Q1, driven primarily by strong revenue growth and the resulting operating leverage. For the second quarter, interest and other income was $12.1 million.

Our non-GAAP tax rate was 12% and non-GAAP net income was $145.8 million resulting in non-GAAP diluted earnings per share of $0.80, an increase of more than 30% from Q1. We ended the quarter with cash, cash equivalents and marketable securities totaling $1.25 billion, up $68.5 million from Q1 and driven by cash from operations of $87.7 million.

Now turning to our outlook for the third quarter. We expect revenue to be in the range of $540 million to $560 million at the midpoint, this represents 40% sequential growth driven by the transition of Scorpio-X Series to volume production, continued strength in Aries PCIe sector timers and clean production shipments of Taurus solutions supporting 100 gigabits per line for 800 gigabyte Ethernet.

As Jitendra mentioned, we expect Scorpio to become our largest product line by revenue in the third quarter, marking an important milestone in our journey from smart signal conditioning products to mission-critical AI fabric solutions.

This milestone reflects the progress we have made over the past several quarters as Scorpio exteriors transitions to volume production. We expect third quarter non-GAAP gross margin to be approximately 72%. We expect third quarter non-GAAP operating expenses to be between $156 million and $160 million.

Our non-GAAP operating margin is expected to be approximately 43% and up 400 basis points sequentially, which reflects significant operating leverage as our revenue expands. Interest and other income is expected to be approximately $12 million and we expect our non-GAAP tax rate to be approximately 12%.

We expect our Q3 diluted share count to be approximately 185 million outstanding. Overall, we expect non-GAAP fully diluted earnings per share to be between $1.16 and $1.21. In closing, Q2 was an excellent quarter with strong execution, driving record revenue and expanding operating leverage.

Our strong Q3 guidance reflects continued momentum and confidence in sustained growth across both our AI fabric and signal conditioning portfolios. We will continue to invest strategically to capture the substantial opportunity ahead while maintaining strong profitability and driving long-term value to our stockholders.

This concludes our prepared remarks. I will now turn the call back to our operator to begin Q&A. Operator?

[Operator Instructions] Your first question comes from the line of Harlan Sur with JP Morgan.

Great execution by the team. It's great to see the volume production ramp of your Scorpio-X scale of fabric switching solutions here in the third quarter? In the press release, you said that Scorpio-X ramp would be led by your newer 320 main products I think that's with one hyperscaler customer, right?

You have another hyperscale cloud customer that is targeted to ramp your lower rate Scorpio product family as well. Is that more of a Q4 high-volume rent? And then with the strong ramp in Q3, is Scorpio-X revenues going to exceed PT revenues this quarter? Or is that still looking to be more like the Q4 time frame?

Yes. So thanks, Harlan. So a couple of points. First, the corpus series has been gaining tremendous amount of interest in the market. So we have started shipping the parts for preproduction for multiple customers starting this quarter.

We do expect Q3 at least for a lead opportunity and lead customer to ramp to high-volume production in Q3 and essentially taking over the -- or surpassing the revenue from Scorpio-P series, so which should give you an idea of the value that we bring in scale-up use cases. We do expect to ship to revenue for additional customers on Scorpio-X by end of the year.

So in general, what I would say is that there's been an exceptional traction that we're seeing, both for Scorpio-X and extends to be more tied to scale up and various different configurations. Speed goes to a multitude of opportunities. But overall, it's been great to see how the customer base is seeing the value that we deliver with our Scorpio family.

And like we noted in the call, Q3 onwards, we expect Scorpio as a family to be the largest product line within the company.

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

Great results. I actually -- I know there's going to be a lot on Scorpio, but I want to ask on the Aries product. You said record revenue I think when you gave initial outlook, it was directionally growth. I think people had pretty low expectations.

Just kind of curious if you could talk about the drivers within Aries. And then as you look to kind of the next generations of AI servers, are you seeing -- what kind of design traction are you seeing and retime or attach in those systems?

Yes. So 2 parts to that. One is, again, Scorpio 5, which is our Gen 5 portfolio that's been around, I want to say, definitely continues to expand especially as we're seeing inference-based applications start showing up.

So that's driving a new wave of growth is how I would put it. 86, which is for Gen 6 supporting the higher data rate, the 64 gigabit per second. Obviously, that is -- comes in with higher attach rate and going into several new platforms.

And also, there are other use cases around the cabled application and so on. So overall, like we've always said, the retimers ease a socket that somewhat keeps on giving in the sense that you need it for a variety of use cases. If anything, the surprise is how long Gen 5 has continued and given that there is a new set of use cases opening up for inference, all of those are contributing to help us achieve a new record on the Aries portfolio.

And by the way, we expect this trend to continue as new Gen 6 platforms and GPUs are being developed. So this is not a trend that will stop. And then, of course, then we move on to Gen 7, and you can expect some of the cycle to continue for a long term.

Your next question comes from the line of Joe Moore with Morgan Stanley.

You mentioned UALink in your road map there. Can you talk a little bit about that road map? And do you see, I guess, a transition from PCI to UALink? And how broad do you expect the customer adoption to be there?

Joe, thank you for the question. Yes, we continue to see a lot of interest in Link for scale of networking. As we mentioned during the prepared remarks, the ecosystem is evolving. We have new IP vendors coming online, new tools coming online -- some of the other vendors have also introduced their own products or announced their own products, that kind of all points to a nice vibrant ecosystem.

From a customer standpoint, we continue to stay engaged with our customers. There is continuing interest in Link A lot of it is actually driven by not only the additional throughput that UALink brings, but also because of the low latency that UALink will deliver compared to some of these other protocols.

So we find a lot of traction for that in inferencing applications and the throughput and the latency and the open ecosystem is definitely continuing to drive the traction from our customers. We are completely plugged in, in terms of the development cycle for UALink.

We plan to intercept the XPUs that are on track to be released in 2027 with our own UALink solutions. And as we have discussed, we want to come in here with a full portfolio of devices, which would be anchored around our switch platform. and then also providing signal conditioning solutions, both for copper and optical.

In fact, we recently introduced our tours family, which support you in at 200 gig per lane line rates. So it's all looking very good. We continue to stay completely engaged. And as things move from the PCI Express generation to UALink, we do expect the dollar content per XPU to increase, and that's going to be a favorable trend for us.

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

Congratulations on the strong results. You talked about new software for the Scorpio-X deployment. I think you even talked about really improving efficiency, [ toconomics ] and so on and so forth. Can you maybe lab a little bit on that? And I'm especially curious, is that new software is also lending you more new customers? Because I assume that's something that's going to be very, very beneficial as they launch their switch fabrics.

Yes. Sanjay mentioned, Tore, there is exceptional traction from our new customers for the new [indiscernible] that we introduced. Part of it comes just because of the fact that we have higher ratings and higher lane count and therefore, we can connect more XPUs to the switch.

But a lot of it is also stemming from the advanced capabilities that we built in. And I think this is what you refer to when you talk about software, which is the in-network compute as well as Hypercast. Both of these are very important technologies that really set us apart and enable very low latency inferencing as well as very high efficiency for net per compute.

Almost 2x improvement in the collective operations that are required for training. And all of these are hardware-based features, but they are enabled through software and this software resides in our Cosmos platform. And so while previously, Cosmos was mainly responsible for optimization, telemetry diagnostics.

Now we are becoming part of the orchestration layer and the XPUs are talking directly to this -- our customers platform to run the workload -- and this drives a lot more stickiness into our solution. Many customers that has adopted our platform in this generation will continue to use this for the next generation, the easiest thing for them to do.

And we continue to support these type of advanced capabilities in our current devices, and we'll carry them forward for the next devices -- next-generation devices as well.

Your next question comes from the line of Natalia Winkler with UBS.

I wanted to follow up on the CXL opportunity. So you guys mentioned the CXL controller and kind of 2 new hyperscalers that you will be ramping in 2027. But curious how that CXL opportunity plays out for the switching content as well? And how do you see that market kind of progressing from here?

Yes, I think it was about time isn't it. I think it's finally, great to see some traction back on CXL. As you know, we have been investing in the technology for a while now. And it's great in some ways because we understand exactly what is needed, both from a hardware and software standpoint.

So now we see actually 3 main use cases. One is largely driven around sort of the memory market dynamics. People wanting to find alternates to the availability or the high cost of memory today. And they're trying to do that through CXL and try to leverage the fact that you can stick different kinds of memory behind the CXL bus.

The second one, which is very exciting for us is the GPU attach or accelerator attach where combined with our KB cash acceleration type of function, we're able to do unique things that fundamentally are improving the latency and performance for influence applications.

The third one is something that we've been talking for a while, which is general compute or HPC application, where for use cases like SAP, where you need a high memory or memory intensity workloads. The CXL provides a way of adding more memory to the CPU. So those are the 3 main applications or dynamics that we see. And for us, what we are tracking is, of course, multiple design wins and opportunities at this point with most of them in qualification or early stage this year. and getting into high-volume production in 2027.

So we do see this trend to continue, by the way, and we are investing heavily in coming up with technologies and capabilities to sold the memory wall problem and the memory availability problem in some ways given where the market is on that front. So overall, we are positive and we are continuing to invest.

Your next question comes from the line of Sean O'Loughlin with TD Cohen. Please go ahead.

Congrats on another great set of results. And thanks for letting me hop on. I wanted to ask maybe a bigger picture question on the product portfolio, especially as we start to think about optics really starting to be something that sees at scale deployment within most investors' time horizons. I guess the question is, as I think about UALink NPO or CPO switch, you have an optics-based Ethernet PHY, Is there a fundamental reason why like Taurus over time, couldn't also transition that physical media dependent layer to one that's based on fiber rather than copper?

And does that end in effect, extend the assumed lifetime revenue of what maybe most of us are thinking of as an inherently copper-based solution today. Apologies if I'm leading the witness here a bit, but just wanted to get your thoughts on that.

No, no, that's a question. So let me comment on that. So first of all, we do fully expect copper and optical to coexist. As you know, most of the scale-up connectivity today is done in copper, and our customers are asking us to continue to push that forward and we will continue to do that.

And as these data rates go up and the signage challenges increase, there is definitely room for additional products and additional product categories, just like we announced our Taras 200-gig both retimers as drivers in support of these applications.

Now we do expect that for this generation and the next generation to scale up within Iraq will continue to be copper. But then as these cluster sizes grow and the data rates increase, optical will definitely play an increasingly important role as you start connecting these racks to each other and our trustee expands from maybe 1 or 2 racks to 4 fiber or even a larger number of racks.

And we are very well positioned to play our part in this developing market segment. We are very closely engaged with our customers. We understand when they want to deploy these increased cluster sizes. And we expect that the first deployment will happen with NPO solution.

As we go from copper in addition to copper to NPA solution, our dollar content per XPU or per link actually goes up. So we definitely welcome this change. Overall, if you look at the optical market, it's several tens of billions of dollars in TAM that gets unlocked.

Once NPO gas deployed and database continues to increase, we will start to see CPO getting deployed. So we look at it as 2027 being the year where NPAs deployed and then 2028 and beyond is when CPO gets deployed. From a product development standpoint, we are very comfortable. We started investing in this a couple of years back. We have a team of engineers working on the EIC solution, which is the electrical component.

We have a team of engineers that is working on the silicon photonics. And of course, late last year, we acquired xScale Photonics which gives us the ability to develop a connector. And put all of these things together, we have now the ability or we will have the ability to have full optical engines which will eventually lead to not only optical links and taught us like components, as you pointed out, but actually a fully optically enables copies.

So we are really looking forward to that supplying both that optically enables Scorpio switch. As well as the individual components that make up the full optical link. And that's a tremendous opportunity for us in addition to what we have at copper.

Your next question comes from the line of Papa Sylla with Citigroup. Please go ahead.

Congrats on the impressive results. So I wanted to double click on the content per accelerator. I think when you initially kind of IPO, it was more around $50 to $100 were accelerated over the past quarters also we went to $1,000-plus and Sunday, you mentioned today's pop alone could be well over $1,000 per accelerated.

So I'm wondering if that $1,000 plus or core, is it a pre-line metric? Or perhaps if you can provide color on the path to that $1,000 plus per accelerator for corpulent, finally, maybe tying everything together, if you can speak more generally on your content per accelerated currently? And where do you see that growing over the next 2, 3 years?

Yes, good question. I think this is a fundamental question that you asked, and it's a very important focus for us, which is how do we keep ensuring that our long-term business continues to grow and we are able to sustain the kind of growth that we're able to deliver.

And it really, like you highlighted, comes down to how can we keep increasing the dollar content that we can get per XPU attach. Like you correctly noted, we started with small numbers in IPO time. Now we see our business contributing multiple thousands of dollars per XPU with the scale or scale up switch itself offering over $1,000 per XPU.

And this is already starting to happen, of course, with the high-grade switch that we are shipping right now. But at the same time, this is just the beginning of it because when you start adding several other features that are important in terms of optical, for example, or adding more single conditioning because of the higher speed and the higher attach rate.

We do believe that this number will continue to scale up, and that is what we are focusing our investment right now, which is how do we ensure, a, we keep increasing the dollar content we get per XPU targeting multiple thousands of dollars; and b, how do we continue to keep diversifying our business, adding more hyperscalers around the anchor socket that we have, which is a scale-up switch and continue to grow significantly above where the market is.

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

This is [ Trip Smith ] on for Tom. I was just curious about how we should think about the Scorpio gross margins now that it's becoming a bigger part of the business? And what are the puts and takes there for gross margins looking out? And I believe you still have the Amazon warrant impact. So I just wanted to double check on that and how we should be modeling that over the next few quarters?

Thanks for the question. It's Des here. As we mentioned in the prepared remarks, we do expect our Q3 gross margin to be approximately 72%, and really as our portfolio continues to sort of diversify. We continue to see a wider range of gross margins with the main impact being silicon versus module revenue mix in the different use cases associated with our switching portfolio.

On Scorpio across both P and X series, we do have a wide range of margins, really driven by use case and man count utilization. But I would say, on average, Scorpio across all use cases has an average margin profile broadly in range with our sort of quarter gross margins. In any given quarter, this mix can sort of change.

But what I would say on the switching side is that we are going after large opportunities. And as the top line continues to scale, we'll see the gross margin dollars flow through and contribute to the bottom line. We clearly saw that in our Q2 results and also our Q3 guidance from here.

But really, given the broadening of the portfolio, I would expect our gross margins continue to trend towards a long-term target of 78% from here.

Your next question comes from the line of Ananda Baruah with Loop Capital Markets.

Thanks, operator for the question what's a useful way to think about the contribution that a hyperscale to make across the portfolio as you continue to introduce the new products and move up from generation.

And it's a good observation. In China, as you can imagine, the open protocols are quite popular, which includes PC Express, Ethernet and increasingly, UALink. So we definitely see that as a good market, not nearly as big as what we have here. But certainly, there is a lot of interest, and we are engaging with the customers there with our PCI solutions as well as our [indiscernible] Ethernet solutions.

So in aggregate, that can be a very promising opportunity, especially as PC Express, is the nervous system of servers and probably the best way to connect many add-in card format GPUs together to deliver good inferencing capabilities.

I appreciate -- is it a way that can be noticeable in the P&L over time?

I think it will be in an absolute scale, yes, our revenues from China will continue to increase. However, we do expect the rest of the world to increase even faster. So as a percentage of total revenue, they will continue to be in the single digit.

Your next question comes from Suji Desilva with Roth Capital Partners. Please go ahead.

[indiscernible] the results here. I'm curious, you talked about the software supporting telemetry and now orchestration gaining more kind of functionality there. Is there kind of a soft line where you might cross over to helping the execute itself with some of the functionalities that co-process or would you draw sort of a straight line there is connectivity versus that dimensional trend in memory compute or things like that?

Yes, very good point. No, actually, that is what I was trying to say when I answered the question earlier that with the advancement of the in-network compute and Hypercast features in our latest Scorpio [indiscernible] device, we are indeed crossing over that line where the XPU is directly talking to the switch to orchestrate the workloads to reduce the latency and increase the throughput for both inferencing as well as training workloads.

So that's a very important development for our Cosmos platform, where not only are we doing the traditional optimization, customization diagnostics and telemetry, but we are really helping to improve the workload and directly improve the utilization of the GPUs that are connected to us with -- so it's a very important development.

And as I mentioned, it makes our solution -- software solution more powerful and a lot more sticky as customers designing our current generation of switches and then look to upgrade them in the future.

Your next question comes from the line of Simon Leopold with Raymond James.

[ Jeff Kochi ] in for Simon. Maybe you could just talk qualitatively about your end market concentration. How is the trend in the quarter? How do you expect it to trend over the next I guess, through 2027, especially with the new launches

Yes, it's a really good question. What I was really pleased to see in our sort of first half is actually has been the broad base strength across our sort portfolio. And really looking at our sort of drivers, I would say ARS delivered a record quarterly revenue in Q2 and remained our largest product line, really supported by the ongoing transition to KCI across both scale out and scale up connectivity.

And I do expect Aries to continue to grow in Q3 and beyond. As we mentioned in the prepared remarks, Scorpio continues to show strong momentum in Q2, really driven by X and P series. And as we mentioned, we do expect to see Scorpio becoming our largest product line in terms of revenue as a result of the production ramp to high rating solutions in Q3.

Taurus also continues to show nice growth, really supported by AI and general purpose and certain platforms. And I do expect to see further growth in the third quarter driven by the strong base of the 400 gig demand as well as the early contribution to the 800 gig as well.

And I think, as Sanjay has sort of mentioned, we continue to see sort of strong progress on the Leo CXL products, and we continue to see some revenue here in the back half of the year, which will ramp more materially into -- as we look into 2017, we'll continue to see further opportunities on the optical and custom side.

But we're really pleased with the overall performance of the portfolio and the continued diversification.

Our next question comes from Karl Ackerman with BNP Paribas.

Jitendra, Scorpio is ramping significantly in the second half of this year, largely driven by the product transition at your main customer. But can you discuss the breadth of customer wins including a growing mix of Neo clouds and whether your Scorpio-X opportunity extends across both liquid cooled and air cold servers?

Yes, I'll take this question. This is Sanjay here. So like you said, the demand for Scorpio-X and Scorpio has truly been exceptional. We have over 10 customers engaged on the Scorpio-X itself. Many of these opportunities are either getting into preproduction or getting deeper into the qualification cycle. So to that standpoint, we are already shipping Scorpio-X to multiple customers right now as they get ready for their own internal qualification cycle, and we do expect that we'll have additional customers ramping to production towards the end of this year, early next year based on the progress that they can make.

So this trend is something that, of course, given the features and capabilities that we have, especially with the 320 lane, we're able to unlock several things when it comes to inference type of applications. So there are several unique things that we are able to achieve.

And all of these are contributing to diversifying our customer base on Scorpio-X and like I noted on Scorpio-P, it's truly broad-based because the PC is going to go into standard PCI switching application. So from storage or networking and some of the scale-out use cases. We have several design wins and design ins that are ramping to production.

So in general, I want to say Scorpio family has taken off in a very positive way for us, and we intend to add more customers and more revenue growth in the second half through 2027.

If I may just add to that. The chip itself supports both liquid cooling and air cooling. In fact, we do have customers that are designing it for both of them. When your point about the new cloud is also very valid because when they start to deploy these solutions for especially inferencing applications, sometimes they can do that without giving a full rack level solution.

And so we do see this increasing use case of enterprise customers and new clouds looking at solutions that are based on Scorpio as a scale up and using Advent card form factor GPUs to deliver good inferencing TCO.

And also maybe to add one point that is somewhat lost is even with the lead hyperscaler we have, there are multiple platforms where we are designed in. It's not like there is just one, right? So I think that's a nuance that needs to be considered as you look at our business.

There are no further questions at this time. I will now turn the call back over to Leslie Green for closing remarks.

Thank you, everyone, for your participation and questions. And please refer to our Investor Relations website for information regarding upcoming financial conferences and events. Thanks so much.

This concludes today's conference call. You may now disconnect.

I think we can go -- all right. All right. Welcome to the Astera Labs' fireside chat. My name is Mark Lipacis, I'm the senior semiconductor analyst at Evercore ISI. And so very excited to have Astera Labs. So we have Desmond Lynch, who is the CFO; and Nick Aberle, who is Senior Vice President of Finance and Head of Investor Relations. So gentlemen, Welcome, and thanks for joining.

Thank you for having us, Mark.

Yes. So Des, you're the new guy right? So maybe as a way to get started and kick us off, can you tell us like why did you decide to join Astera Labs? Like what did you find attractive? And how have the first months been?

Yes. Thanks, Mark. Joining us at the labs, it's been a really good sort of first couple of months here. I've been really excited to sort of join the company. And if I go back to my initial decision to join the company, I would say there's a couple of things that stood out. Number 1 was the overall size of the market opportunity that the company is going after. And the company is very well positioned within these end markets. Secondly, I would say the broadening of the portfolio that the company has been able to achieve over a short period of time, the company has made the right investments to go after this market opportunity, which is really exciting from there. And then thirdly, I would say that I've been really the superior business and financial model that the company is driving towards.

Having been within the company for the past couple of months, I would say that all of these factors are sort of playing out but being within the sort of 4 walls of the company, what's really stood out to me has been the engineering focus of the sort of company, the focus on execution and delivering to customer needs. I saw that up close hand with the Scorpio X release that we've just done earlier in the quarter, and that was a phenomenal effort by the team. And the other thing that really stands out to me is really the strong relationships that we with the customers, and we continue to solve complex engineering sort of problems. And as I look further ahead, I continue to see the sort of growth opportunities within the company. So really excited to be in the company. And the first 90 days have played out really well for me.

Excellent. All right. Well, I think you talk about the broadening of the portfolio. We have a lot -- I want to talk about lot of these products. So we're going to jump right into it. So starting with Scorpio, the smart fabric switches, can you provide like a brief overview of the product segments, how the business has grown how do we think about growth over the next couple of years?

Yes. We've been very pleased on the Scorpio switching family. This was a product line we announced in 2024. It does have 2 different sort of product lines under it. We call it Scorpio P, which is associated with scale-out technologies and then Scorpio X, which is for scale-up applications. Our growth rate on the sort of business has really been phenomenal sort of last year. This quickly became our fastest-growing product line within the company and accounted for about 15% of total company revenue last year. Much of the growth was really driven by the P-Series sort of products and scale-out applications. And as we look ahead to the sort of second half of the year, we just announced our 320-lane high radix solution for scale up.

That will contribute in volume production in the second half of the year. So we're very excited about the opportunities there. We continue to have strong engagements with our sort of customers. Much of the growth really has been driven by our lead customer. But if you look at our diversification efforts on the P-Series and what we will see is 2 hyperscalers contributing to revenue here in the second half of the year. And then on the X Series, we continue to have strong engagements with our customers, and we expect to see some design wins towards the end of the year, which will contribute into the sort of '27 and beyond sort of time frame from there. But we are going after a large market here, Mark. We're talking about $4 billion on the scale-out applications as well as $10 billion on the scale up, which is really a greenfield sort of opportunity for us. And we feel we're very well positioned to continue to grow. And by the end of the year, I do expect that Scorpio will become our largest product line within the company. So very excited about the growth opportunities ahead of us here.

How is the -- high growth. It's a great market. How do you think about the competitive dynamics? What is -- how do you think about share? What kind of share can you achieve in the market?

Yes. So I would say that as D pointed out, I mean, we're talking about a pretty new greenfield opportunity that's very open-ended from a market opportunity and growth standpoint. So very excited to be one of the first guys to really be shipping in material volume. We've done over $100 million of P-Series business last year, first to be shipping Gen 6 solutions for PCI Express for both scale out and scale up. and we're just really getting started. So I think what you'll see over the balance of this year is us continuing to grow that very aggressively and positioning ourselves to continue to grow in '27 versus the competition, which is still trying to get kind of standing up with a solution and get it into customers' hands. I think our approach is just a bit fundamentally different than other folks when you talk about addressing the market.

What we've seen historically is general purpose switching solutions trying to be jammed into AI types of applications, where from the beginning, we specifically designed and developed our solution for AI-based applications. So you would see that with the P-Series in terms of how we staged the lane counts and the radix and how it would support all the different elements within the head node within these AI racks. And then now you're seeing it with the X-Series in terms of what we can support both for smaller Radix solutions and the larger Radix solution that Doug just alluded to on the 320 lane 80 port that's really going to allow you to do single hop to an entire kind of 72 XPU rack. I think the other piece, and you've seen this in our signal conditioning business that we're taking and kind of applying to our approach on fabrics is this eye towards providing information, diagnostics, telemetry metrics not just raw horsepower and low latency and low power, right?

Being able to tell our customers and fleet managers exactly what's happening in their systems. And when you talk about Cosmos, our telemetry software working on a single conditioning platform, we can tell you about temperatures and link health. But when you're talking about on the AI fabric side, you have much more information to digest and feedback and serve as a feedback loop to the customer. So being able to have that baseline of software and that visibility into the platform across multiple products and now higher value, higher protein products like the fabric is a huge differentiator for us from a competitive standpoint because we're basically providing higher utilization, higher productivity back to our customers and our parts are effectively more than paying for ourselves in those situations.

So -- and when we when we think about the original product you came out with the retimer, you guys made the argument that the Cosmos was a real differentiator, right? Is this -- is what you're doing on the Scorpio product, is this an extension of this? Is this like -- is this an idea where your customer is used to Cosmos and then this makes it easier for them to Scorpio assumption...

That's exactly right. So I would view it as Cosmos being the baseline infrastructure that's sitting down and integrated within hyperscaler operating stacks at a fundamental layer. The blood, the sweat, the tears, all that that's been poured in over the last couple of years to help support signal conditioning can now be reused and leveraged with the AI fabric portfolio as well. So a lot of the work has been done, and now you have more tools from Astera to be able to monitor and observe exactly what's happening in your systems, not just from a link-to-length standpoint, but actually what's happening with this traffic. And like I said earlier, I mean, it just yields higher productivity, better utilization, and that's really what these guys are striving to deliver. They're spending tens of billions of dollars on infrastructure. So if we can squeeze in an extra couple of percent, it starts to really.

Yes. I think it's forecasted to be over $1 trillion right now for -- at least in our model for '27. So you started off with in Scorpio 64 lanes and then you -- now you have a product line 32 to 320. So what is the motivation to go down and have the broader product line? Could you just tell that? Normally you think, oh, you want more lanes.

Yes, yes. So I mean, I think there's -- if you look at the lay of the land, there's going to be like lots of applications across AI topologies that are going to demand different shapes and sizes of solutions. It was very strategic from our standpoint to get into the market with the P-Series initially with that 64-lane solution because there was customers that were demanding that. And it's been very widely deployed across NVIDIA NVL-based platforms that are customized by hyperscalers. But we've always known that this was just the initial foray into this market. There was going to be a portfolio of solutions, and we've been developing that obviously, for the last couple of years.

So this kind of formal announcement we did kind of in conjunction with earnings was really to highlight that we have multiple solutions across both scale-out and scale-up types of applications to service all these different sockets. And it's not kind of trying to jam our 64 lane into places where it might not fit or it might not be able to service a wider radix kind of demand profile. So super excited about expanding the portfolio. It's going to generate a lot more shots on goal for the company. And now we can kind of go in with more strategic products and attack these different sockets and try to pick up share as a function of a breadth -- a broad portfolio rather than just kind of 2 solutions.

We talk -- we spend a lot of time doing field work, talking to all the different hyperscalers. And the one thing that comes through is the hyperscaler as a group is not a monolith, like everybody has a different approach. And I guess I'm wondering, as you talk about this broad portfolio, is this a reflection of your -- the different customers on the hyperscalers and Neo cloud? Are they -- because they have different approaches? Or is this a given hyperscaler will be using a 32 and 64 and a 320 lane solution? .

Yes. it's a one-two punch from us, and we've been saying this since IPO that no 2 clouds are the same. What these guys are trying to accomplish with their customer bases are completely different. Their workloads are different. So you want to be able to the best of your ability to service all these different clouds with all the different products and applications that they're trying to deliver to their end customers, right? You're never going to have the perfect choice for each socket. So I think broadening the portfolio is definitely one big step towards that direction.

But I think the other piece, and we've been talking about this for a long time as well, is fundamentally building the solution with the software-first mindset such that we can go iterate and tweak on the fly and be able to turn off and turn on or turn off or enhance certain features or optimize certain capabilities in order to specifically support that customer and their cloud. So I think it's a great one-two punch for us to be able to provide a larger buffet of solutions from a hardware standpoint, but then also being able to leverage the core software-defined architecture that we've had embedded within our products since day 1 in order to give them as much flexibility and optionality as possible.

And then on Scorpio X, can you just talk about what -- why your customers are adopting that for scale up when you compare to other alternative solutions that they have.

Yes. I mean, so we're still in the early stages, like we mentioned, of clustering XPUs and GPUs in large arrays. So you had NVIDIA really kind of was the first guy to go down that path with NBL 72, and now you're starting to see the hyperscalers do very similar things, right? So they all skin the cat in different ways, like we just mentioned and the choice of protocol or approach on the back end for scale-up is going to be a customer-by-customer decision. And at the end of the day, you have -- if you kind of add all those up, if you exclude NVLink and you just look at the merchant-based switching or fabric solutions that will ultimately be available in the market. We see that as being a $20 billion-plus opportunity. And there will be opportunities for PCI Express.

There will be opportunities for PC Express evolving to UALink. There will be opportunities for Ethernet. and each of these guys will pick their different paths. So where you've seen us gain traction initially is in this PCI Express domain for Gen 6 specifically where we're first to market. We have engagements with over 10 additional unique customers that will use PCI Express in some form today or tomorrow and then eventually evolve to UALink down the road. But we've been very kind of strategic in terms of supporting open protocols and open standards to allow for these customers to kind of build their systems around an open approach. So that's why we've seen a lot of the initial solid traction, especially on the PCI Express side and then ultimately expect to see that on the UAL side over the medium and longer term.

Got you. And then like how do we think about like pricing here? Is there a do we think of a rule of thumb in dollars per PCI Express lane? Or as you add features like hypercast or in-network computes this do you get more.

Yes. So we try not to talk specifically about ASPs or even maybe even dollars per PCIe lane. I think what we've talked about in general is since the beginnings of the company, targeting with ARS $50 to maybe $100 worth of content for ARS. But as we sit here today, being able to service a customer with over $1,000 of content per accelerator as we bolted on additional products like Taurus, P-Series Scorpio, X-Series Scorpio. So the broadening of the portfolio has been certainly a nice content driver for us. If you look at it on a per product basis, I mean, I think there's -- it's very obvious that as we move from generation to generation, you're going to be having faster speeds.

You're going to have more features and functionality like you mentioned. So the capabilities that we're able to provide continue to be higher value add for the customers, and they're certainly willing to kind of pay out for that. So I would say as you look from generation to generation on a like-for-like basis, you will continue to see ASPs go higher in general, driven by faster speeds and greater functionalities. And it's incumbent upon us to continue to deliver these kind of new technologies so that they can grab that productivity and utilization gains like we have in prior generations.

Well, I mean, from my standpoint, that's exciting to hear because normally in semis, you think you're giving the same ASP for improved functionality over time, if I think about historically the CPU. So when ASPs goes up, that's a very good thing. So you mentioned UALink. Amazon and AMD have publicly stated plans to deploy UALink. I think Jitendra has also mentioned additional customers who are planning to deploy. Can you just talk about the UA Link ecosystem? Where is it in its development cycle, the maturity, how -- talk about your position in this market?

Yes, yes. It's been very dynamic and fast kind of evolving situation. So UA Link consortium was formed almost 2 years ago now. They put the first spec out in April of last year. And as we sit here today, they just released the second version of the spec month, 1.5 months ago. So we've gone through a couple of rounds now. The ecosystem continues to mature across kind of the supplier base, the technology and the IP availability so that folks can go use UA Link IPs and embed it within their solutions. The testing and the qualification machines and equipment are starting to come available as well. So you see a very well-rounded ecosystem starting to form.

And similar to what we talked about our evolution as a company and our ability to be successful in the market being driven by building something specifically for AI, that's exactly what UAL set out to do from the beginning, right? So if you think of PCI Express, PCI Express has been around for decades. I don't think they ever assumed that scale-up clustering applications were going to be even a thing when you go back into the '70s and the '80s. So we're really kind of leveraging kind of older standards and protocols to drive up some of these initial scale-up systems today. So it works and it's effective, but it's not as efficient as you could be. So UAL is really taking that next step and to say, hey, we're going to build something from the ground up that's going to be specifically tailored to driving scale up within AI infrastructure. We're going to take -- we're going to pick and choose the best of all the worlds. We're going to use low latency memory semantic approaches from PCI Express.

We're going to use the high horsepower front-end SerDes from Ethernet, and we can combine it together to provide like a best-in-class solution for the masses of folks who want to leverage an open ecosystem and open architecture to drive the platform. So I think that's the big piece that's driving people towards checking out the technology and ultimately adopting it. Like I mentioned, we have over 10 customers that are planning on using PCI Express to scale up over the course of the next year, 1.5 years. And almost all of them are like, hey, we're going to use UALink as the next extension of this, whether it's next generation or the one after -- so it's a very natural evolution. It's optimized and specifically tailored for these applications and the progress that we've seen has been certainly compelling.

And is the long pole in the tent, it's just the maturity of the ecosystem? Is that?

Yes, I would say, yes. I mean, the ecosystem will continue to mature. Folks have to build products. We're building products. So we'll have fabric solutions supporting UALink. We'll have signal conditioning solutions supporting UALink. So those will start to kind of become prevalent and in customers' hands, I would say, towards the later part to early -- later part of this year, early part of next year. Customers, as you mentioned, a couple, AMD and AWS have endorsed it as kind of the go-to scale-up protocol. I think those guys in general have pointed to late part of next year is initial platform deployment on UAL. So I would say revenue kind of 28%, but in between now and then, obviously, a lot of work to be done on our customer standpoint, from our standpoint, getting these products out, getting them up, getting them qualified. But yes, I'm really enthusiastic about the progress that's been made to date.

Great. Let's shift gear to the optical road map. You acquired AI xScale Photonics. How do we think about your rack opportunity -- content opportunity? How does it relate to CPO scale up? Is this a scale-up technology, scale-out technology? -- give us kind of the lay of the land on this product.

Yes. I mean, so optical has been leveraged for scale-out for some period of time now. There's a very large market there, as you well know. Optics within a scale-up domain is still very nascent. And when you think about it, the characteristics that make copper so great in terms of kind of performance, low power, low cost, reliability, you're going to continue to use copper everywhere you can. And if you listen to some of the keynotes at CpyTax, it's like copper, if you can, only optical, if you must. But I think in terms of looking at the next generation to 2 generations out, certainly, you will see an expansion of cluster sizes, and we will start to move to these multi-rack scale-up topologies where you're not only scaling up within 1 rack or maybe 2 racks, but maybe 3, 4 or more than that, more racks than that going forward.

So then you start to get into the problem of distance to scale up and you start to expand beyond the domain of copper. And in these situations, you want to have an optical solution in order to support those larger cluster size and scale-up opportunities. So we see it as a big market enhancer for us. We plan on obviously shipping a ton of copper-based product into the market for the foreseeable future. But to the extent that we can grab these longer distance kind of bigger dollar opportunities on the optical side as you move to multi-raack scale up, that's something that we definitely want to participate in as well. I think one of the key points to remember for us, as we attack the market with an anchor socket like Scorpio X, we become a very key piece of that connectivity backbone for scale-up topologies. And we want to be able to support electrical connections with that. We also want to be able to support optical connections with that.

So I would say the long-term goal, you mentioned CPO is to ultimately optically enable Scorpio X, and you will have a CPO-based switching solution with Scorpio X at the heart of that and optical engines basically bolted on and around it. In the intermediate term, there certainly will be opportunities to ship the fiber coupler that we acquired through the xScale acquisition. And then there will be an intermediate step towards NPO as well, where we plan to service that market with some analog mixed signal solutions as well. So very -- so we see the signposts. We're working closely with customers to intercept these different market opportunities as they evolve. But having that critical position within the backbone of connectivity with Scorpio X puts us in a really nice place over the long term to benefit from the transition to optical.

What IP have you developed on the copper side that you leverage into the optical side?

Yes. So I mean the switch itself is going to be a huge piece of the equation. And I think that we've also talked about, we haven't unveiled the full plans yet, and we'll continue to provide signposts going forward. But a lot of the telemetry and observability data that we drive and support with Cosmos today will be very critical within the optical domain. We're working on building out custom Pies that will be very prevalent and material within the optical domain as well. So there's different pieces of the puzzle that we continue to gather. The fiber coupler is a differentiated technology that I think that will provide us some very nice advantages in terms of the connector capabilities of the solution. So all in all, software, hardware, IP, I think we're in a really good spot. But yes, I mean, we've got to continue to execute. There's a lot of work to be done still, but pretty excited about the opportunity.

So let's come back to the original product, the retimers -- by our estimate, it grew 65% in 2025, what's continuing to drive the strong growth here?

Yes. So we've been very pleased with our performance on the retimer business. As you mentioned, it really has been the backbone of the company, and we saw some really nice sort of growth sort of last year. And we expect that growth to continue in '26 and beyond. It's really speeds and complexities that continue to sort of grow. The attach rates for retimers continues to remain sort of strong. And the other factor I would point out is we're in the middle of the sort of transition from Gen 5 to Gen 6. Just in the recent earnings call, we talked about 1/3 of our sort of revenue be attracted towards the Gen 6 solutions. And the important part here is as you move from generation to generation, you usually see a 20% to 25% uplift in the ASP. So that ASP, coupled with the strong attach rates gives us sort of confidence in the sort of growth rates of the business going forward, and we're very pleased with our performance here, Mark.

Got you. When you guys IPO-ed, I don't -- I'll admit I did not know what a retimer was. So -- and now it seems lots of companies are coming out with retimers also. And so what is -- what helps you keep share in this market?

Yes. So I would say that incumbency is a very strong factor and especially within an environment where you're sitting on a link and conditioning signals back and forth across an extremely eclectic backdrop of different types of CPUs, different types of memory, different types of networking devices, different types of storage, different types of CPUs. You have a whole variety of different things that you need to support in different combinations. So interop and understanding all the nuances within all these different endpoints is a highly critical kind of valuable piece of the equation here.

And being able to -- I mean, we've shipped millions and millions and millions of these devices into the field. There's very few AI systems that have been deployed globally that don't have ARIES embedded in them today. So the amount of learnings and tribal knowledge that has been gathered over the course of the last couple of years through this very wide deployment, basically battle tested through the gauntlet is a huge differentiating factor for us. As we move to Gen 6, kind of same story, kind of first to market. We've talked about 1/3 of our total business being PCx+ Gen 6 in just the most recent quarter, I don't think anybody else is really shipping any type of material volume at this point.

And it's a very easy transition for customers that know that we've been proven, we've been battle tested with Gen 5, and we're well positioned to kind of help and service and support on Gen 6. and you can continue to use and leverage the Cosmos infrastructure that you've already kind of spent money and time and resources on to embed into your operating stack. So certainly, it's going to become more competitive. This is a very large and fast-growing market. So it's not like going to be handed to us, but I feel like we have a very good strong position, and we view ourselves as being the leader in this space for Gen 6 and the generations to come.

Got you. All right. Let's shift to the LEO CXL controllers. When I think about this market, I always think whenever there's this new protocol, it's hard to get the industry to embrace it and like it always seems to take longer to play out than once it starts playing out, it plays out faster than you expect. And it seems that, that may be the case with CXL. Can you talk about what -- where we are in the cycle, what's driving increased interest?

Yes. I would say some of the supply chain dynamics within the memory sort of space is really leading to renewed interest from sort of customers. Historically, our sort of play here has really been associated with general purpose compute. And just at the end of sort of last year, we did announce our sort of deployment with the Microsoft Azure platforms, which will come into the revenue profile in the back half of this year and continue to ramp 27 mark. What I would say is we're also having interesting sort of discussions on AI inferencing sort of opportunities. Just as part of the earnings call, we did announce our KV Cash option, and we did announce the fact that we did have a design win with a hyperscaler, which will ramp into sort of 2027 time frame. I agree with your sort of comment. We've seen the market been sort of stop start and also the deployment of these sort of new technologies does sort of take time. And with that, we will be cautious with the overall sort of outlook. But we do see this as being a sort of large market opportunity for us. We've sized it over sort of $4 billion. But what we'll continue doing is providing signposts for investors as we sort of go along to indicate the sort of progress. But again, we've invested very early into the sort of CXL space, and we're excited now we're seeing this come back around from opportunity perspective.

Got you. Is there a synergy on the LEO CXL and Scorpio when we talk about like KV cash offload use cases?

Yes. So synergy from the perspective of, as I've been talking about Cosmos running underneath all of this, right? So Cosmos is going to be leveraged across Ares, Taurus, LEO, Scorpio. So that's all a common framework that's going to be utilized by our customers. But I think that is an important piece. We're getting a lot more questions around CXL. And certainly, we've seen activity and engagement kind of start to percolate around this for obvious reasons. So it's a situation where we just need to kind of block and tackle along our way here. Everybody is trying to skin the cat a little bit differently. But we have the ability to ultimately service these customers and these appliances and these applications with not just the CXL controllers, but these oftentimes you're going to need switching products that could need signal conditioning products. So again, kind of speaking to the breadth of the portfolio, our goal is always going to be to maximize the amount of content that we see on a per system or per platform basis. And we see CXL just being additive to that and being able to kind of pull in some of our other pieces.

Got you. Okay. We didn't talk about Taurus and the AECs. Where is Taurus in kind of the deployment cycle? What kind of traction are you seeing on...

Yes. We continue to see sort of strong traction and sort of good growth on the sort of Taurus product family. Last year, it was probably around 15% of total company revenue. Most of this growth has really been driven by the 200 and 400-gig solutions. We do expect 400 gig to continue to be strong throughout this year. And really importantly, our growth has really been driven by our sort of lead customer. But as we move to the 800 gig transition, which we expect to take place in the second half of this year, we will see diversification beyond our sort of lead customer, which will be nice for us to see here. And then looking sort of further beyond that, we'll continue to see the 1.6T deployments, and we have strong sort of engagements with customers. So we continue to see strong sort of growth rates in terms of the Taurus and it remains an important part of the portfolio here.

Okay. I want to ask about your content per accelerator. I think when you IPO-ed, you were in the hundreds of dollar range. I think now you're in the $1,000-plus range. So how do we think about the next 3 to 5 years? As a sell-side analyst, I would pull out my log scale, semi-log scale chart and draw a line from $100 to $1,000 and then say, well, in 3 to 5 years, it should be $10,000. What's wrong with that now?

Yes. I mean that's certainly the goal. And like I mean, we have -- we're in a great position today with higher value, more critical pieces of the overall portfolio such as Corpio X that we're sitting at the table with our customers talking about the next generation, the next generation after that. And the connectivity challenges that you see coming down the road are going to be only more complex and more critical than they are today. So there's a very kind of long-term visibility here in terms of what we need to do in order to service those next pieces. So you bring up a very good point. When we started cutting our teeth in the market, it was kind of with these IO-based products, signal connectivity solutions, which are great, and they've built a very substantial piece of business for us. but the ability to layer on and build out the portfolio to service these next-generation connectivity challenges is really what's driving that content growth.

So you have the big secular trend that you mentioned earlier, $1 trillion worth of infrastructure going in the space will certainly benefit from that trend. But our goal is to grow faster than the market. And the way that we do that is with more capable solutions as we go generation to generation and then the additive nature of new solutions coming to market. So when you start to think about us getting into the custom solution space, we start to get into the optical connectivity space. These are all going to be incremental opportunities above and beyond what we service today and drive that $1,000 of content with. So I don't want to quote an exact number at some point, maybe you draw a line, you do get up to that level. But from what we can see today in the next several generations, we very confidently saying that our content opportunity per accelerator will continue to grow at a very robust clip.

And I think a lot of people in this market, they are concerned about customer diversification. Can you just -- just at the highest level, at the IPO, GPU versus ASIC attach? And then today, where it could go 2 to 3 years from now?

Yes. So I mean -- and this is another amazing point of just how much the market has changed from when we went IPO just to today, is not a long time ago, a couple of years plus. And what we called an AI system back then was an hopper-based platform, HGX with 8 hopper GPUs. And as we sit here today, it's a rack of 72 GPUs altogether in a very dense situation. So yes, I mean, a lot of our initial traction in the market was, by definition, GPU-based when we started to come into the market. NVIDIA was the frontrunner and deployed widely across all the big hyperscalers, all the Neo clouds, AI factories. But as we've evolved over the last couple of years, you've started to see the rise of the XPU and these guys designing and developing and delivering solutions that are optimized for their own workloads and their own platforms to give them optionality to use AMD and NVIDIA and then their own solutions.

So yes, as a function of that, we have seen kind of a faster-growing business on the XPU side. And I think -- so just in terms of those volumes growing, number one. But number two, NVIDIA does a great job of servicing their customers with the full shop, right? Up and down, left to right, you have switches, you have GPUs, you have the GPUs, now you have CPUs as well. In the XPU domain, we can service and support more because they don't do their own scale-up switches. So we participate in all the same places we would on a GPU-based platform, and we can help support the scale-up platform, which is good for our signal conditioning business, tons of connections with Ares that we can service for scale up. And then obviously, with the big Scorpio X switch to scale up as well, that's another big opportunity. So both sides growing very fast, XPU, just more dollar content per platform and then those volumes have started to kick up as well. So it's been a great spot.

Got you. We're running close to the end here. I want to ask, what do you think, as you talk to investors, is the biggest either misperception or the biggest disconnect between your view of your market position and the fundamentals and where the investment community is?

Yes. I would say I can start. I think folks tend to get hung up on exactly what's happening in the next month or 2 months or a quarter or 2 quarters and what's happening with the next generation of this or that, what are the attach rates? But hopefully, what we try to paint a picture of today is we're really looking at multiple generations, and we have high confidence that as these generation -- next-generation solutions come to market, the content opportunities start to get a lot more meaningful for guys like us that are supporting the connectivity backbone, which is becoming a much more critical piece of the overall platform to be able to service and support the scale-up of very dense cluster sizes. So I would say that we're really excited not just about what's happening in the next quarter or 2 with certain design wins, but what we can accomplish over the next 2, 3, 4 years down the road. And as our position as an AI infrastructure connectivity provider continues to become even more critical and drive that content higher. So I don't know if that's misunderstood, but I feel like you're almost kind of come up out of the weeds a little bit to realize that there's a longer-term path there, a lot of opportunity, and we're building the company and really scaling our operations to be able to support that.

I've only been here a couple of months, but I would say the couple of things that really stand out is investors are really looking at how big the switching scale-up opportunities could be. I think we've touched upon that today. We see this as being a greenfield opportunity and really a sizable sort of market by the 2030 time frame. And the other thing that I would say that gets a lot of attention is the optical sort of road map. I think Nick gave you some sort of good color on our sort of progress. We've been investing here the past couple of years into the optical road map. We see some nice plays on the NPO and then ultimately leading to the CPO sort of road map, and we'll be a good participant in this market in that sort of time frame. But again, as Nick sort of said, it's usually dealing with sort of what's going on today, tomorrow, but we are much more focused on the long term. We have that very robust sort of road map here, Mark.

Continue.

I thought -- Nick, I thought you were going to lead with the death of copper for...

That was so That was 3 minutes...

Right. Is that -- do you think investors are aligned with you? I think you probably alluded to this in your -- when you talk about your road map. Are investors fixated on this still? Or there's a fixation on the...

Yes, I was like the metaphor is when everybody freaked out about deep seq and inferencing was bad for demand for compute, right? I think we view optical as not something that's going to take away from copper, but something that's going to be an incremental opportunity for -- so -- and we'll provide more color and information and market data around this over the next couple of quarters, but it's going to add meaningful amounts of incremental market opportunity for Stera Labs above and beyond what we support today. And we want to be a diverse supplier across all these connectivity mediums. So yes, we see it as a positive. And yes, I think as we've talked more about it and started to give some signposts around it, I think it's giving people a little bit more confidence. But we have a lot of work to do still. So I totally understand that there will be a lot more questions and discussions around it.

And I think you talked about CPO and your package optics and the photonics product, that growth vector, kind of a plight way to suggest that maybe people are not completely tuning...

I would say it's been -- again, we've been dropping sort of breadcumbs along the way here, Mark. The acquisition of the xScale was really important for us as it helps to provide that sort of piece. But as we continue, we see -- we're developing our sort of NPO solution in-house just now that will come out to the market and we will be in time for that sort of ramp. And then further out, we'll have the CPO sort of solution. So excited about the continued investments we're making within the optical space.

Well, the red light blinking means we ran over as I thought we would, Des, Nick, thanks so much for joining. Thanks for sharing the insights about Astera Labs. Looking forward to...

Yes. I appreciate it.

Success. Yes. Thank you. Thank you.

Thanks.

All right. Well, ladies and gentlemen, thanks for joining us at the 54th Annual TD Cowen TMT Conference.

Before going any further, I think I'm contractually obligated to remind you that Extel voting is open. And so if you appreciate the work that either of us up here on stage do, please lend your support. So with that out of the way, we have Nick Aberle from Astera Labs, VP of Finance and IR, and thanks for joining us.

Appreciate it, Sean. Thank you very much, and thanks to TD Cowen for having us. I would vote for you guys. You guys did some really good work. So we appreciate your support.

Appreciate that. So let's just dive in. Starting off near-term environment. We're about a month removed from you guys reporting a pretty solid set of results.

Maybe just start by highlighting the key points that investors should be focused on there, whether it's mix or strong outlook, what should we be keying in on? And maybe if you want to take it more broadly, how does that fit in the context of who is Astera and how do you deliver for your customers?

Yes. Got it. So I would say in terms of the near term, yes, I mean, we've obviously been on a nice run the last couple of quarters. We added good results, guided for a bit of an acceleration to the growth rate in Q2 and pretty optimistic about continued inflection of growth going into the back half of the year.

So nothing really new to report relative to a month ago. I mean, I think as everybody knows, demand continues to be very robust from a secular standpoint. But I think most importantly, for Astera and what we're trying to accomplish, right, is really building out this connectivity backbone for the industry.

And in order to be able to do that, we need to be able to broaden and diversify and grow our portfolio to be able to service and support customers in a variety of different ways.

So what you've seen from us over the last several years is a diversification and an expansion of the product portfolio beyond kind of our core kind of I/O connectivity products like Aries and Taurus into more complex solutions like our Leo CXL memory controller and then most recently, the Scorpio Smart Fabric switch portfolio, which addresses both scale-out and scale-up market opportunities.

So kind of taking a step back, I mean, the name of the game really for us is there's tens of millions of accelerators that will ultimately ship annually over the course of the next couple of years. And we want to be able to staple as many dollars to every one of those accelerators that goes out the door as we can.

So if you go back to around the time of IPO, Aries attached maybe $50 to $100 worth of content per accelerator. When you started to layer in the scale-up piece of Aries, you kind of expanded that kind of by 2x because of the higher attach rates.

Taurus started to layer into the model as well for scale-out kind of driving that number up overall to several hundred dollars of content. And then with Scorpio, P-Series adding several hundred dollars of more content and then X-Series bumping that whole thing up to about $1,000 in total.

So in a very relatively short period of time, we've gone from sub-$100 of content per XPU or GPU to upwards of $1,000. And as we think about growing the business and expanding the business going forward, it's going to be layering on additional parts of the portfolio from a connectivity standpoint to continue to grow that $1,000 up much higher than that.

I think that's a perfect summary and a good setup. And why don't we just go through the star signs. And starting with Scorpio, maybe specifically on Scorpio X, I think it would be fair to say that, that's likely to drive a significant portion of the ramp here in the back half of the year. And maybe just throw to you and say, well, first off, is that accurate?

Second is, what are the points of risk on that? Is this -- and to the upside or downside, are there technical stepping stones still left? Or is this just we're going to hit the knee in the curve and you're going to ship from inventory and from your partners?

Yes. So I would say at a high level, yes, so Scorpio X, of course, is going to be a great catalyst for the top line for the remainder of this year and even into '27 as we diversify and broaden across incremental customers as well.

But all businesses are growing for us. So if we look at the growth rate sequentially into Q2, we're seeing growth across every single one of our product lines into Q3, Q4 and into next year, we expect every single product line to continue to grow as well.

And this is a function of kind of secular trends, more accelerators going out the door, more platforms, more racks being shipped, but also higher attach rates, next-generation solutions driving higher ASPs and higher content on a per product basis. So every single one of these pieces is driving and contributing to growth.

With that being said, as you brought up, Scorpio X is a big piece of that puzzle. You have a market opportunity just for Scorpio X of kind of $10 billion plus which is almost half of our total addressable market.

So clearly, there, the content, the ASPs are extremely rich, so they can drive and move the needle a lot more than some of the other pieces of the portfolio.

And Scorpio X has been in the works for a long period of time. So it's not something that just has come on to the scene recently. We've been shipping preproduction volumes of Scorpio X back in the second half of 2025.

We moved into kind of early volume production in Q1. We've expanded the portfolio now to include both smaller Radix and larger Radix solutions, of which those are shipping in Q2 as well.

So of course, there's always going to be milestones. The volume deployment is not an easy task when you're talking about extremely systems and platforms with thousands of components and many accelerators all needing to scale-up and work perfectly together in unison.

So tons of work still always being done, but very confident about where we stand today and Scorpio X will be a nice driver for us in the back half.

Yes. So I think in a similar way, you've talked about Taurus likely to be a contributor to growth this year in maybe a more material way than it has been or I guess, maybe was last year as your main customer ramps 800 gig.

I feel like taking both a near and a longer-term view on the AEC market for Ethernet, we have a lot of debate amongst us investors about whether or not AECs are dead at pick your line rate, 1.6, 3.2T. And yet that's in the context of we're not even now ramping 800 gig.

Maybe just talk about where we are in the AEC life cycle. And is that a durable market for you that you see as a longer term? Or is that sort of a stop gap is a terrible word to use here, but before we move to a more optical-focused ecosystem?

Yes. I mean I think at the highest level, our view is that the optical opportunity within scale-out or scale-up is incremental and additive to our copper opportunity.

And you've heard many in our industry, just as recently as the last couple of months talk about copper having a very long life cycle and based upon the characteristics that it brings in terms of power, reliability, cost that it's going to continue to be a preferred use case for wherever you can manage to get the length that you need or the reach that you need out of it.

When you talk about Ethernet-based AECs, for example, or in particular, you've seen the growth of that market really be driven by physics such that speeds continue to go up, distances kind of continue to stretch and you've needed active solutions versus passive solutions in order to kind of bridge the gap on these connections.

So you've seen the market opportunity for AECs kind of grow as a function of these secular technological trends -- and we believe that's still in the early stages as well.

And to your point, as we're just on the cusp of moving to 800 gig and ultimately 1.6T behind that, there's still a tremendous amount of passive copper opportunity to kind of take share from in the marketplace. So we would expect just the market overall to continue to grow over the next several generations.

But overall, I mean, we're excited about not just that, but also kind of optical. I'm sure we're going to talk about that in a little bit as well. But yes, I mean, for us, Taurus is going to continue to be a good driver of growth. We see the market growing overall. And then we see ourselves expanding into additional customers as well to kind of broaden our reach there on the customer side.

Great. On the other active copper side, on the PCIe side, Aries was, as you mentioned, sort of the flagship for Astera when you guys came public. And by my estimates, and maybe this is too high, continues to drive maybe 60% of revenue even in the early stages of this year.

But we've -- we're now 10 minutes in, and we're only talking about it for the first time. So how should we be thinking about Aries both this year and longer term as it almost mechanically has to shrink as an overall share of your revenue, but maybe is still growing internally quite nicely.

Yes. So yes, I mean, you bring up a pretty good point. So in 2025, Aries portfolio grew by roughly 70% year-over-year, which is not too shabby of a growth rate, but came down dramatically as a percentage of revenue. I mean it used to be 90% plus of revenue.

And yes, I think your kind of 2/3-ish number is in the ballpark of where we stand today. So we've seen tremendous growth, but everything else is growing faster, and it kind of goes back to that kind of core theme that I was talking about earlier.

And our ability to grow has really been driven by development, expansion of the portfolio, bringing new products to market, not just new products, but products that carry kind of higher ASP and higher content opportunities.

But Aries has got a great story ahead of it. We're just in the early stages of transitioning to PCI Express Gen 6. We talked about that being a little bit of over 1/3 of our total revenue in the Q1 time frame, not aware of anybody else shipping any PCI Express Gen 6 product in volume today.

So we continue to feel very confident in our market position. And for Gen 6, on the PCI Express side, both for scale-out and scale-up, you'll see a higher ASP on a like-for-like basis on or about 20% generation over generation. We also expect to see higher attach rates in general, again, kind of going back to that physics problem where speeds are going up, speeds have doubled again, and you're going to need retimers in places you didn't need them before.

So we expect to see very healthy growth rates for Aries in 2026 and then follow that with the 2027 as well. So we continue to be very bullish about PCI Express.

This is a curve ball I didn't prep nick for it. But is the dynamic with Aries maybe that it moves from the -- as the scale-up domains move to optics, the physics on the scale-out or front end dictate that Aries maybe becomes much more prevalent and the use of retimers for PCIe where you -- as you just mentioned, you previously don't need a retimer to get from the CPU to the BMC or whatever.

Correct.

Is that where we should think about Aries starting to move into, call it, 27 plus?

Yes, I would say, yes. So the attach rates on that scale-out or kind of head node connectivity domain where you're connecting between CPUs, GPUs, storage, memory, networking, all of that is going to continue to be wired by PCI Express.

PCI Express is going to continue to evolve and provide faster speeds. And even though those traces aren't getting a ton longer, it's still going to be trickier to get from each endpoint to endpoint. So we do expect to see in that domain, a higher attach rate of PCI Express retimers.

For scale-up, I mean, I think that market is still going to be a very nice opportunity for us as well. I think -- and we'll talk about optical in a second. But as you start to get the multi-rack and you need to go over 7 meters and you need to go hopping 2, 3 racks over, then you start to need an optical solution. You can't use copper or PCI Express.

So that will be an additive market opportunity that we'll try to prosecute.

But yes, certainly a lot of opportunity still for PCI Express on the scale-up side as well. I think the last thing that I would point out is we continue to talk about 10-plus customers for scale-up for Scorpio X, of which we talked about the lead customer and how that's going to ramp, and that's great.

But there's a whole slew of incremental PCI Express opportunities for X-Series that will pull through PCI Express retimer business as well, both for scale-up as well as scale-out. So we're excited about those opportunities as well.

Yes. That makes sense. Let's talk about optical. So you guys have talked about sort of the 2028 time frame where that scale-up will start to flip, appreciating that, especially as we've seen with NVIDIA and Google, how they are sort of deploying a hybrid copper optical scale-up domain. We'd assume that will look similar.

You guys made it a public investment, aiXscale, I think I said that right. I think what many, including myself, have failed to appreciate is what you guys have been doing behind the scenes on sort of getting your EIC and PIC portfolios ready and really heavily investing not in an M&A style, but in an R&D style to be ready for this optical transition. Maybe you could just fab on that?

Yes, yes, definitely. Yes. So we've been at it for at least a couple of years now on the optical side, putting together the team, gathering the resources and trying to build out the road map of how to attack this market opportunity.

And of course, as with everything that we do, a lot of this is done in very close kind of in concert with our customers. So we understand their road maps, their timing, what challenges that we need to be solving for in the next 2 to 3 years, whether it's copper or optical. And we do see opportunities to kind of intercept on the optical side with the products that are in development for us currently.

So I think that you'll see over the course of the next several quarters, us continue to lay out very kind of finite signposts about our progress and how we're heading down this track. We've talked about generating optical-based revenue in 2027, both from the fiber coupler on a stand-alone basis and from NPO applications as well. So kind of stay tuned for that, but that's on the horizon for growth in 2027.

But yes, I mean, it's -- there's a fair amount of complexity with what we're trying to do from a road map standpoint. The aiXscale acquisition for us was very meaningful. As we've been doing our research and homework and discussing with customers over the last couple of years, this coupler technology, it's a little bit more of an art than a science, something that's outside of our typical kind of capability and flow.

So it made sense to go acquire it. And we were attracted to that because it was a solution that is probably not talked about as much as the EIC in the PIC, but something that's very meaningful and influential to the overall scalability of the overall optical engine. So that's an important piece. That will be sold as a stand-alone component.

We talked about a leading AI infrastructure provider, building that into their CPO solution that will be ramped next year. So we're excited about shipping just that stand-alone piece into volume.

And then there will be another step as we move towards the back half of next year where NPO becomes an application that we can begin to address. Jitendra and Sanjay have been talking about building out our analog mixed signal capabilities, and so we'll have new products that will be able to address that part of the market.

Those products will be integrated into an EIC piece that will ultimately drive our own PIC as well. All of that development is being done in-house organically currently. And then there is a packaging and a test and a manufacturing flow that is very complex as well that kind of brings all this together and allows you to deploy an optical engine at scale as well, and that's all being worked on as well.

So tremendous amount of work being done on that front. The expectation is that the end of the road, we'll have an optical engine that we can bolt around Scorpio X and create a CPO, our own CPO solution to provide to customers. And that optical engine can ultimately be used on the other side of the link as well and be integrated around customers' AI accelerator as well.

So we want to be in a position where we can provide full link accountability across the entire domain. So those are some of the pieces that we're driving towards. So of course, we'll announce more formal products and get a little bit deeper on exactly what we're doing, how we're differentiated, what the market size is. But yes, I would just say a tremendous amount of work being done, and it's a big market opportunity.

Yes. And I think one of the things that Jitendra has talked about as well is that you're looking to be photonics agnostic. Maybe you could talk about what that exactly means because I don't know that just those 2 words makes a lot of sense to me without going deeper. But maybe you could just talk about what -- how you're fitting into your customers' plans and road maps.

Yes. So I think that what we've discovered in conversations with customers and in our engagements is that certain customers have a certain religion around photonic ICs, what modulation to use, different features and functionalities that are going to be priorities to their systems and their workloads that they're trying to operate.

So from our perspective, again, I mean, we're -- I think as you've seen probably across our entire portfolio and how we kind of manage this is we want to provide customers optionality. We want to support flexibility.

So to the extent that a hyperscaler customer would want to use their own specific photonics IC for whatever reason that they might want to do that, we're building the rest of the puzzle in an agnostic manner to be able to incorporate and leverage that PIC.

We'll have our own organic PIC technology as well. So that will be offered and kind of hopefully utilized broadly as well. But we wanted to provide a little bit of flexibility there just in case folks wanted to go down a different path and focus on their own piece that was going to be their own special sauce to that equation.

I think another example of that philosophy and your approach to how you deal with customers is the NVLink Fusion concept. I will say that I think it's one of the more debated topics for your company these days.

Can you just walk us through what that looks like from a -- either from a product level or how you actually expect customers to deploy an NVLink Fusion configuration versus maybe your own native UALink silicon?

Yes. I mean -- so to your point, yes, I mean, it's another kind of example of us trying to provide flexibility and optionality for customers. I think this is a pretty exciting one. Any time you get to work with guys like NVIDIA and AWS on a project like this is something pretty meaningful and kind of shows you there has been a track record of execution and trust kind of built up between our companies over the course of the last several years. So very excited to be part of that program.

And yes, I mean, I'm not sure I want to get into the actual specifics around what they're specifically solving for trying to accomplish. But I would say that in general, there could be applications where a customer, a hyperscaler would want their accelerator to be able to interface with NVIDIA's back-end scale-up topology.

But it's not native for that accelerator to be able to communicate with an NVLink protocol that has been proprietary to date. So where we step in as we provide kind of the bridge of that translation layer with a pretty I would say, almost fabric level solution that will sit between the accelerator and NV switch on the other side on a one-to-one basis to be able to provide that translation.

So that requires a pretty hefty amount of kind of collaboration and coordination and development with both of those companies to make that happen. So very excited about that project. I mean it's kind of opened some new doors as well. NVIDIA can be a good matchmaker. So we're seeing additional kind of doors opening as a function of that. I haven't talked about customers or any specifics there, but I would say that there's the one kind of -- or the one collaboration that we've talked about openly, but there's others percolating as well.

I mean -- and it sounds very, I don't know, bespoke and hand-in-hand. How should we think about the competitive environment in those kind of projects?

Yes. I would say, yes, I mean, it's a custom solution. So it's not something that's going to be handed out and multi-sourced or something like that. So yes, I mean, it's going to be coming down to relationships, track record of execution, trust. And there will be a whole slew of different opportunities and sockets to go after. I'm sure that we will not win every single one of them. There's other folks in the space as well that will certainly be looking to attack and kind of grow along with this market opportunity as well.

But nice to have a lead customer to kind of hang our hat on. I'm sure there'll be a tremendous amount of learnings gathered through this process that will give us a good leg up and leverage as we look to kind of support and service additional customers. But yes, it will be competitive. But yes, it won't be a multisource situation. It will be you'll win on your own merits for that specific socket and go execute against that.

Yes. It would be fair to characterize it as sort of a project-based and you would get significant amounts of that, if not all of it.

Exactly. Yes.

Okay. Let's spend a few minutes just on supply chain. The investor base is -- it seems investors are searching for the next bottleneck around every corner.

I guess, open-ended question, where are you seeing constraints? How are you thinking about dealing with it? We saw AMD invest $10 billion in their Taiwanese supply chain relatively recently. How are you going about managing that side of the thing?

Yes. I mean we've been building out our kind of back-end operations for a long period of time, and it's always been kind of built with an eye towards scale volume and big numbers ramping, right, even from when we were a smaller company.

And that is largely driven by the customers that we were trying to obtain and service and support. And we've been successful at that. It's kind of paid dividends. So I would say from a supply standpoint, I mean, we have a very good understanding what our allocation profile looks like on both the front end and the back end looking out through '27. -- and feel very comfortable about where we stand from a supply standpoint.

And of course, we've buffered that with kind of upside as well to the extent that customers come back and need more. So does it solve for infinite upside? No, of course, it doesn't. But in terms of what we've built in terms of our model and how we're looking at things and what the customers are conveying to us, we feel very comfortable with our supply situation with room for upside.

To your point, that's also not a foolproof plan because there's other pieces of the puzzle that are within these racks. So to the extent that they're short on something else, could that delay a program rollout? Could it delay the timing of something?

Sure, it could. We typically try to provide a decent amount of buffer around our expectations for timing and ramps anyway to kind of hopefully kind of protect and stay conservative against those types of scenarios. So nothing I would call out specifically. I know everybody is working very hard to kind of not let that happen. But for us, in terms of what we control, we feel very comfortable at this moment.

Yes, hard to hold you responsible for someone else's part if it becomes a bottleneck.

Yes.

Just on that same topic, but thinking about the future road map, one of the places, I think, is the most bottlenecked if that's a word, is the optical space. And I would -- I think it's fair to say that, that's a relatively new area of supply for you guys is that work would presumably need to start getting done today. Is that fair to say? Or any comments there?

Yes. I mean, yes. So those -- again, this has been something that's been kind of going on for the last couple of years. So we've been working and partnering with -- across the ecosystem for both the electrical side of the house as well as the photonics side of the house.

We're building out manufacturing capabilities around the fiber coupler as well that kind of came with the aiXscale acquisition. So we're expanding that profile kind of proactively to be able to service kind of good volume in '27 and ramping well beyond that.

So yes, I mean, I think we feel -- I mean, again, it's a very tall task in addition to developing these products, kind of we've been kind of -- the guys have been mandated with building out that ecosystem and the supply chain in order to give customers confidence that we can support and supply. But yes, I mean, it's in concert with the customers as well. I mean that's 2 punch to go out into the market and kind of get what you need, so.

Yes. I think I'm required to use the word Agentic at some point during the fireside. So I think it was surprising to hear maybe some discussion of Leo and CXL. I mean we wrote a deep dive in 2024 and CXL was this whole chapter and then nobody talked about CXL for another 2 years. How does Agentic and agentic workloads sort of fit into the Leo portfolio?

Yes. So as you're kind of alluding to, like the dynamics within the memory space have become quite challenged over the course of the last 6 to 9 months with rising prices and supply constraints. So I would say that where we were previously very focused and engaged with customers on the more general purpose compute applications, your more standard servers that we're now seeing a lot of increased interest and engagement activity around leveraging CXL within AI inferencing applications.

And we talked a little bit about this on the conference call that basically hyperscalers are looking very closely on different ways to optimize and get every penny they can out of every -- get every bit of performance they can out of every penny that they spend on the memory side.

So that activity and engagement is all very good. That's very broad, and it could be very meaningful. I think the one kind of most tangible piece of that is we did talk about a new design win with a hyperscaler customer for a customized version of Leo that will be used for its CXL kind of connectivity purposes within kind of AI inferencing appliance.

And so that will be kind of hopefully the first kind of shot across the bow in terms of some real revenue and some real momentum within the CXL space. I agree, it's been a long time coming. We've been talking about it since IPO, and it's unfortunately kind of lagged behind because of the kind of AI kind of took the day and CapEx started to go that direction. But yes, kind of funny how things come back around, and we're certainly seeing a lot of interest there and excited about it finally kind of kicking off.

I think your P&L has done okay. Big picture to close, we're at time, but what's the biggest difference between the conversations that you're having with us and the conversations you're having with customers and ecosystem partners? Like what are we too focused on? What are we not focused on enough?

I mean I think the biggest thing that we hear from customers that you guys probably don't have as much visibility toward is really kind of that 3- to 5-year lens where they know exactly what they're trying to build, where they're trying to go with their accelerator capabilities, with the rack capabilities and then how many of these things they want to ultimately cluster up and scale-up.

And what that does is create a tremendous amount of challenge and complexity for the connectivity backbone of all of this. So they need to be very tight with folks like us and some of our peers to be able to solve for some of these challenges in order to meet their deadlines and their time lines and service the workloads that they're trying to prosecute.

So I would say that the conversations that we're having aren't about what's going to drive revenue in '26 or '27, even it's how we're going to solve these next-generation connectivity challenges, '28, '29, '30 layered across a broader portfolio for Astera Labs across both optical and copper.

So there's a tremendous amount of work and collaboration and co-development that's being done in order to make that all happen. So that's probably a bit of stuff that you guys are not seeing the full kind of all the way down to the lowest levels.

Great. Well, perfect place to wrap. Nick, thanks so much for your time.

Yes. Appreciate it, Sean. Thanks.

Okay. Great. morning. Thank you. Welcome to the second day of JPMorgan's 54th Annual Technology Media Communications Conference. My name is Harlan Sur. I'm the semiconductor analyst here at the firm. Very pleased to have the team from Astera Labs. We've got Jitendra Mohan, Chief Executive Officer and Co-Founder; and Desmond Lynch, Chief Financial Officer, here with us today.

Company, obviously, as most of you know, leader in accelerated compute, connectivity, networking, memory and storage controller solutions. Their silicon optical software solutions are integrated into 90% of the world's AI compute servers and clusters. It's been a busy earnings season. So I've asked the team to start us off with a brief overview of the March quarter results, June quarter outlook, and then we'll go ahead and kick off the Q&A.

So gentlemen, thank you for joining us today. Jitendra, let me turn it over to you.

Yes. Thanks, Harlan. That was some high-speed communication there. So yes, we've been at it now for two years now actually. So we went IPO about two years ago, March of 2024 and at that time, our revenues were $65 million and change, and our EPS was about $0.10. And 8 quarters later, we closed our last quarter at $308 million in revenue and $0.61 in EPS. So about 5x increase in revenue and more than that for EPS. So very happy with where we are.

Last quarter was $308 million. We grew sequentially about 14% and about 93% year-on-year. And the growth there was broad-based all of our products grew but certainly, the highlight was our Scorpio X family, which started ramping into volume production in Q1 of last year. Overall, very pleased with where our PCI Gen 6 portfolio is. It is more than 1/3 of our revenue for the quarter. So that's a great milestone to have. And just kind of -- we're getting started on PCI Express Gen 6. Today, there is really only one Gen 6 GPU out there, which is the Blackwell GPU from NVIDIA. As more come, we will see continued strength in this portfolio.

For the next quarter, we have guided up 17% at the midpoint of the range of $360 million in revenue. And again, that we anticipate a broad-based growth. All of our products, Aries, Taurus, Scorpio will grow. But certainly, the highlight again would be the Scorpio family where we do expect by the end of the year, it should be our largest product family. So very excited about that and continue to work on new product families and announcing new constellations throughout the year.

That was a great overview. And continuing on to that, I mean, it's been a very strong growth profile, as you articulated, right. Calendar '25, you drove your revenues 115%, earnings 140%. Over the past year, your calendar '26 revenue and EPS outlook has been revised higher by 70% and 55%, respectively. Total revenue growth of 80% this year and 3-year revenue and earnings growth CAGR of 40% to 50%, right? And this is probably conservative, right? Help us break down the drivers of upside over the past year, but more importantly, what are the strong growth drivers for the team looking out over the next sort of two to three years?

Yes. So the last year has been very good. As I mentioned, all of our product families grew very well. As continue to grow, there were a lot of questions asked about, "Hey, what is the longevity of the Aries family?" Happy to report that we grew from '24 to '25. We are growing from '25 to '26, and we are also on track to grow from '26 to the outer years as well. So that product family continues to grow very well as we deploy it not only for scale out applications but also for scale up applications and also different form factors, starting with the chip applications where you take the chip and solder it on the board to cable applications where we supply small cable modules that eventually go into active electrical cables. So that has gone very well.

And just continuing with the theme of PCI Express, we introduced our Scorpio P family last year. And that ramped very nicely even though it was a partial year for Scorpio P, we ended the year with 15% of our overall revenues from the Scorpio-P family. So that traction will continue this year. On top of that, this year in 2026 will be a lot of growth coming from Scorpio X family. And we anticipate that -- together combined, Scorpio becomes the largest product family exiting the year. And within the Scorpio family, Scorpio X will actually overtake Scorpio P. That just goes to show the amount of TAM that's available for Scorpio X, which is very large, about $10 billion as we anticipate and it's a completely greenfield opportunity. So very excited about the growth potential there.

Our Taurus continues to grow. It was the largest growing product family in Q4. We will see more growth with our lead customer this year and we are anticipating further growth as we diversify from one customer at 400 gig to many customers at 800 gig, which should also happen towards the end of the year.

We are finding new opportunities for Leo. Lee is our CXL memory controller device that started out for general purpose compute applications at Microsoft Azure, in particular. We are finding new use cases for KV Cache. So that's progressing as well. And we are also very excited about some of the new product families that we are working on, namely our custom line of business where we talked about NVLink Fusion as a new opportunity for us to participate in the NVLink ecosystem, that's fantastic. We also talked about a custom version of Leo being designed in at hyperscaler for KV Cache offload.

And then last but not the least, optics. So today, most of the deployments are all focused on copper, and we benefit immensely from our content in copper, generation on generation, we've increased our content from starting like sub-$100 when we went IPO to crossing $1,000 in some of the content rich applications with our existing product families. Now you add to that optics with NPO, CPO out in the future. we are set to increase that content significantly beyond $1,000. So all in all, very happy with where we are, and we are investing -- we are not shy about investing to make sure that we secure our place in the future.

If we step back for a second and we think about your customers' customers and we think about the AI workloads, right, a big inflection happened in the second half of last year. Inferencing workloads overtook training workloads. And ever since then, inferencing workloads have continued to grow at an exponential rate, as measured by things like token generation and so on, right? And this created new silicon opportunities, where inferencing has created new silicon opportunities, it's created new storage and memory tiers. It's created more demand for high-performance CPUs. Obviously, storage and CPUs communicate via PCIe right in the sweet spot, right, of your technology and Astera's product leadership. That's one example, right? Your CXL solutions, as you mentioned, targeted KV Cache applications may be another example. But help us understand like how the transition for more inferencing-based workloads, especially Agentic-based workloads has potentially helped to create new opportunities for the team, maybe even potentially expand your SAM opportunity.

Yes, certainly. So we've always maintained that our solutions are applicable for both training and inference. And that rack level becomes sort of the single unit of compute that is going to be used for both training and inference. So that continues. With the advancements in Agentic, I would say our strategy doesn't need to change. Our strategy is still the same, which is to address connectivity bottlenecks for these training and inference based systems. But the applications have become more broad.

So in particular, maybe I'll point out 3 things, right? If you look at the inference workloads in particular. For inferencing, especially as we go into these long context vendors and all of us are probably kind of guilty of uploading big documents saying, "Hey, here is a big document tell me what we should do about it" or a large piece of code and write me some more code. As the context vendors go up, the bottleneck shifts towards memory. So utilizing memory, giving more memory bandwidth has become more important. And this is where we see additional applications for our Leo products. In particular, this customer opportunity we talked about where hyperscaler is looking to deploy KV Cache, offload application over CXL attached memory. So those type of applications are very good, and they're additive to the applications we already have for general purpose compute. So that's one thing I'll mention.

The second one is just a different type of inference workloads themselves. So for example, if you look at a mixture of expert workload for inferencing applications, they have a certain traffic pattern, which likes all GPUs to talk to all other GPUs because you don't know where the expert, which GPU the expert resides in. And so what you need to do is in your switching portfolio, you need to deliver full peer-to-peer bandwidth. And our Scorpio family was designed bottom-up to deliver this exceedingly high bandwidth across all of these GPUs. Not only that with the addition of features like hypercast and in-network compute, we can really offload the GPU. We can free up a lot of I/O bandwidth on the GPU by moving that feature into the switch. We can also free up some of the compute bandwidth by moving that feature also into the switch. So that becomes a very, very important part of the overall architecture that our customers are deploying.

And last but not the least, with Agentic. And in general, the sort of this requirement from the frontier labs to lay their hands on compute, whichever way they can find, you find a lot of diverse configurations coming into the mix and customers trying to deploy different configurations. And as you said, Harlan, PCI Express is really the nervous system of a server and we are in a leadership position with our PCI Express products. So we find additional opportunities where people want to connect different combinations of mix and SSDs and GPUs and whatnot over PCI Express, and that's a fantastic opportunity for us.

Yes. Another dynamic that we -- that has -- we've seen changing in the marketplace is this adoption by custom -- your hyperscale and cloud customers and the frontier model builders to want to develop their own custom silicon XPU solutions, where we call those ASIC XPU solutions, Google TPU, Amazon Trainium, Meta's MTIA and so on, right? And our view is that by calendar '28 unit-wise, the market will be 50% merchant off-the-shelf GPU 50% custom ASIC implementation, right? Our view is that the move to custom ASIC implementations is a net positive for the Astera team, more reliance -- because these custom guys are going to have more reliance on standard-based protocols for their scale up, scale out connectivity, networking requirements, but wanted to get the team's view on Astera dollar content capture opportunity sort of merchant versus custom ASIC XPU.

Yes. So we take pride in calling ourselves kind of the Switzerland of connectivity. We definitely address both merchant solutions, GPUs as well as ASICs. Having said that, there are two factors that really determine what our content is. One is the number of XPUs, ASICs or GPUs that you are going to use to deploy for a particular workload. The workload isn't changing. We are still going to ask the AI to do what we are going to ask it to do. But in the back end, how many of ASICs or GPUs deploy really impacts our content. In the case of ASICs, typically, they are not as performant as the leading NVIDIA GPUs. So they end up deploying more ASICs to solve the same problem, and therefore, we get more content with the ASIC platforms.

The second thing is what is used for scale up. So if somebody uses PCI Express or PCI Express-like protocols, that provides additional opportunity, much more -- actually a multiplicative opportunity for us to participate. So that's the other place where we benefit from ASIC deployment.

Now having said that, in our Scorpio X family is a prime example of that. Having said that, we are participating in GPU ecosystems as well, both when a customer takes a GPU and deploy them in their customized racks, we participate there in scale out. That is actually what led to our growth in 2025 with Scorpio P series and Aries Gen 6 deployments that happened as well as well as additional opportunities we are participating in now as part of NVLink Fusion, for example. Over time, what we would like to do is to address all of the AI platforms that our customers are using both the proprietary ones where we can participate like NVLink and NVLink Fusion, PCI Express and UALink as customers evolve from going from 1 protocol to the other. And over time, Ethernet and other protocols as well.

So we spent some time talking about your customers' custom silicon sort of initiatives. But as you mentioned in your prepared remarks, I mean, Astera themselves has a multibillion-dollar opportunity in front of them doing your own custom chip solutions for some of your customers, right? The team has engaged in a number of custom chip programs. You have your NVLink Fusion partnership. With NVIDIA, that's going to enable heterogeneous GPU, XPU, rack-scale solutions. You also mentioned you've got two custom CXL memory controller solutions for inferencing-based KV Cache offload applications, right? How many custom programs is the team working on? What is the dollar content opportunity per program? And when do you see these custom programs like ramping into production?

Yes. This custom business is very exciting for us. And the way this got started is really because of the trust that we developed with our customers. We've been working with our customers now as part of Astera Labs for about 8 years. And this time, we've earned the trust from our customers where they share with us what their platform is going to look like, what our solutions need to look like to participate in those platforms. Lately, what has happened is these platforms are becoming so kind of co-designed with software, hardware, the whole rack infrastructure that our customers who want customized solution to get just that much more performance out of their AI infrastructure. And they're coming to us to develop these custom silicon. And it's an exciting opportunity. And frankly, we just don't have enough people, enough resources in the company to address all of these applications.

The two that we have talked about so far, one is NVLink Fusion, which now gives us the opportunity to participate for the first time in the NVLink ecosystem. And we are collaborating with NVIDIA and the hyperscaler customer to allow the hyperscaler customers compute to be deployed over envying ecosystem. So that's very exciting. The team is fully engaged. We expect the ramp for this product to happen in 2027 alongside the ASIC ramp as has been announced by our customer. So that's one.

The second example was the custom Leo. This is a little bit different. We did not do a new chip design for this application. As I mentioned, our chips are software-heavy, we'd like to do the -- we like to architect our chips to use as much software as possible. So we were able to use some of those features in the chip to create a custom version that was more suitable for this application. So that's another example. These are the two that we have talked about publicly. As time goes on, we'll give other indications of these custom programs.

One thing I do want to mention is that these custom programs are not like ASIC opportunities, the typical ASIC opportunity where the customer has done the front-end design and the ASIC vendor is responsible for the back-end and manufacturing. These are from concept to manufacturing type designs where a customer comes in and says, I want this product. And we do the full front-end design, the back-end design and take it all the way through manufacturing. So we -- that allows us to capture more value than you would be able to capture in a just kind of a pure-play ASIC model.

Before we go into some of the specific merchant customer family, I just wanted to see if there are any questions in the audience. If you do have a question, please wait for one of our microphones to get to you. Any questions from the audience? Okay. Let's continue. So let's move to your Scorpio -- Okay, go ahead.

Yes. Just curious about the Cerebras. That's sort of like a bigger -- like a huge wafer like a pizza type of I mean where you have like more connectivity directly on semi, like a system-on-chip offering. If that does kind of ramp up, how does that change the kind of the PCI memory link that needs to be there?

Yes. I mean, first of all, what Cerebras has done is fantastic. Congratulations to that team on a very successful IPO. It's a very different model of deploying compute. And as they become more successful, that might evolve to be another platform. But at the end of the day, you still need to get data in and out of that system. And that is where we play. So the traditional scale up where you connect many of the GPUs together may not be applicable because all of these XPUs or ALUs or whatever they call it, are there already on a wafer. However, in order to get data in and out of that, you still need a framework, and that will be an opportunity for us.

Any other questions? It feels like also like for some of these more Cerebras-like opportunities, at some point, they will be needing to leverage like big storage repositories, right? And that will be a big opportunity for you guys as well?

Yes.

Okay. So on the Scorpio family of scale-up and scale-out fabric switches, two families, as you mentioned, Scorpio P, PCIe-based switches, Scorpio X, scale-up switching family you started ramping these products 2025 by the end of the year with Scorpio driving like 15% of your total sales, right? Just to your Scorpio, like you said, is going to cross over to be one of your largest product segments.

Scorpio P, let's start off with that. In addition to NVIDIA, you've been working with two other Tier 1 hyperscalers outside of this lead customer. Is the team on track to ramp Scorpio P to these additional customers second half of this year? And has the overall just overall customer engagement continue to expand?

Yes, we've been very pleased with our sort of progress on Scorpio P, which is targeted towards scale-out implications. We launched the product last year, and it became quickly our fastest-growing sort of product line within the company, representing about 15% of total company revenue last year. Much of this growth was really driven by our lead customer across multiple programs from there. And what we've continued to do is invest in the product line. Just recently, we announced the 320-lane scale out application sort of switch, which we continue to receive sort of strong feedback from our customers. And as we look to the second half of the year, we are on track to be shipping to our two hyperscaler customers and that will ramp in the second half and continue into 2027 and beyond. We are attracting a sort of high market here $4 billion sort of market opportunity on the scale-out switching applications by 2030. And given our sort of strong product position as well as the customer traction, we expect to continue to grow nicely on the P Series products.

No, perfect. And then on your X Series, Scorpio X Series products, right, this is your scale-up switching solution, started preproduction in Q4 of last year, now ramping initial production shipments with high volume in the second half with your lead customer for their flagship custom XPU solution. In addition to your lead customer engagement, the team has been talking about engagements with 10-plus customers. When do these other sort of customer programs start to fire for the Astera team?

Yes, we're very excited about the Scorpio X product family and the growth opportunities ahead of us. This is certainly a greenfield TAM opportunity for us. We have been shipping our more radix solutions. And as you mentioned, our higher radix solutions will start shipments here in Q2 and continuing to the sort of back half of the year. We have talked about having 10 customer engagements, which continues to sort of grow. This is spread across hyperscalers plus AI platform providers. We do expect to turn a couple of these into design wins by the end of the year, which will then lead to sort of revenue opportunities in the 2027 time frame. So very excited about the sort of growth opportunities that Scorpio X offers us.

That's perfect. And then if we think about the next generation, thinking about the scale-up road map architectures and we think about next generation, then we start to think about standards-based scale-up opportunities, right? And UALink standards base is the next big opportunity to unfold for Astera. It's primarily a 2027 opportunity, but you need to have your solutions to customers kind of second half of this year. Your flagship customers here. We all know are AWS and AMD, Trainium 4, M500, I believe, are the target platforms. Can you guys just give us an update? I mean, are you on track to ship eval products second half of this year. And more importantly, outside of your two lead customers, again, has the overall customer engagement pipeline expanded around UALink and UALink adoption?

Yes. So UALink continues to grow very nicely, actually. There's a lot of traction for two reasons. One is UALink was designed from the bottoms up to do scale up. It is a really, really efficient, high throughput, high bandwidth, low latency protocol meant for scale up. And also because it's an open protocol, so it encourages a vibrant open ecosystem. So we have seen continued progress in there. The UALink consortium has put out specifications. Several vendors have adopted it. So there is IP available, there is verification IP available. So in general, the ecosystem is doing quite well. And Harlan, as you correctly pointed out, two companies that have publicly stated their support for UALink is AWS and AMD. And both of these GPU and ASIC are going to ramp in 2027. And so our product development is on track to intercept these deployments as they ramp their ASIC and GPU platforms, respectively.

Now on top of that, there's actually a lot of traction from other customers. Some have already decided to use UALink outside of these two. Unfortunately, they have not yet publicly stated their position, so we cannot do that until they do so. But there is a lot of traction. The market for UALink and PCI Express combined, we think, is about $10 billion in TAM, and we are very well positioned to address that not only because of the trust that we have with our customers.

But look at what we are doing with the Scorpio X family now. We are -- with the 320-lane devices, we will learn a whole lot in the coming months on what it takes to build a good scale out switch. As you know, it is beyond just the speeds and feeds that you need, right? 200 gig and the number of lanes, et cetera. Features like in network computer and hypercast are very important to optimize how a scale-up network operates. And these are the things that we will learn and we will transition these into our UALink switch. And from a customer standpoint, these feature sets and how they use it, how they deploy it in their software is seamlessly ported from the current generation of Scorpio X devices to what we do for UALink using our COSMOS software.

Yes. And on the -- that's a good point. On the new product front, my sense in covering the team before and through the IPO and through now is that it feels like at any given point in time, you're ramping a product and you probably have two or three next-generation products in the design win pipeline, right? And the introduction of -- on the last earnings call of your high radix 320-lane Scorpio X, Scorpio P family products is sort of a reflection of that, right? Higher lane count simplifies large-scale topologies, improved performance with single hop latency. But -- and you mentioned it a couple of times here, right? You've integrated features like in networking compute and hypercast technology. Help us understand like how these new futures actually boost AI inferencing performance for your customers?

Yes. So if I just can stay and talk about the new introduction of the Scorpio X 320-lane, there are a few things that happened with this introduction. One is it's a high radix, highly [ in-count ] switch. What that means is you can connect multiple GPUs together with a single hop and you get the advantage of having really, really low latency that natively comes from PCI Express and PCI express-like protocols and the lossless net working meaning when a GPU says, "I want to get data from this other GPU", it is guaranteed to arrive in a fixed amount of time. So that's the benefit of having a large ratings and large lane count and a high throughput peer-to-peer bandwidth, et cetera. that is part and parcel of what we have with the Scorpio X family.

The second thing is hypercast. So when you go for a mixture of export type of inferencing workloads, the so-called exports reside in multiple GPUs. And so for any given inference query, you may be lighting up different experts in your system. So that creates a lot of one too many and all to all type of traffic. And what hypercast allows you to do is instead of one GPU having to individually send messages to four other GPUs, it just sends a message to the switch and say, "Take this message and send it to four other GPUs." And to switch now does the work of sending the message to other GPU. So therefore, offloading the IO bandwidth from the GPU and they can use that for other purposes.

Similarly, when you go for a training workload you have the ability or you have the requirement of multiple GPUs to solve part of the problem and then combine the results together. Again, instead of a GPU going to the other GPU and getting their piece of the puzzle, they just send one message to the switch, the switch does the combination and sends it right back to all the GPUs. So now you're able to upload communication overhead, you are able to offload compute overhead from these devices.

And then last but not the least is the COSMOS framework that I talked about. In the past, we were using COSMOS to customize our solution, optimize it, provide a whole bunch of diagnostics and telemetry type information to make these systems running smoothly. Now with the introduction of hypercast and in-network compute, both are hardware features, but they get exposed through COSMOS. So COSMOS software also becomes much better and much more embedded into our customer software and that will have a lot of benefits to us in the future as well.

Perfect. We've got a question here. Can we get the mic? Yes, thank you.

Thank you very much for your time. Do you have a major optical design ramping in 2027 with your new fiber coupler but it seems that yet the broad volume time line for CPO optics seems pushed to post-2028. So I was wondering, does this mean the upcoming generation of Ultra accelerator platforms are still fundamentally designed around the copper-first architecture?

Yes. So today, all of scale-up happens worldwide happens on copper and the reason for that is copper is reliable. It's just the lowest cost of ownership, the lowest power when it comes to scale-up applications. And so we think that the overall copper and optical will continue to coexist for a very long period of time. The connections at the rack level will continue to be copper and certainly us and others in the industry will continue to support that because that's what our customers are asking us to do.

Having said that, as these cluster sizes expand from one rack to multiple racks, 3, 4, 5, 6 racks, then copper will run out of steam. It's just physics. You cannot run from one rack, four racks over on copper at 200 gig and 400 gig. So that's where optics will come first. And different customers have different approaches on how to deploy it. Some customers want to go straight to CPO. And I think those type of applications, you will see CPO for scale-out happen as early as 2027 and then CPO for scale up happening 2028 and beyond. And we will certainly participate in that with our -- the aiXscale glass coupler technology. It is very good for high ratings implementations of scale-up and scale-out and so some of our customers are using -- are wanting us to just sell them this component. So we will have those component level sales.

Other customers want to go to NPO first because NPO does not stress the supply chain of the XPU or the switch. The XPU and the switch remained completely electrical, but you have an NPO component that allows you to go longer distances when such reach is required. And we are 100% engaged with our customers on the NPO road map and we should start to see the first revenues for NPO happen in 2027, but again from rack to rack type of links.

And for both of these for both CPO and NPO, you need an optical engine, and I'm happy to say that we started our development of -- towards getting to optical engines nearly two years ago. We only put a team together to work on analog mixed signal technology for custom [ Piles ] that are needed to drive silicon photonics. So we've been working on that. We are in a very good place with our electrical IC component of our optical engine. Electrical IC will then drive a photonics IC or a PIC. We have a full team of PIC designers working on Astera PIC solution as well. However, we are also open to our customers choosing a particular optics or a PIC solution because many customers are pretty religious about what modulation technology, what fab they want to design. So we have designed our EIC and our glass coupler to be photonics agnostic to the first level. And so we can work with TSMC coup or we can work with GlobalFoundries or TowerJazz or whatever the choice might be.

And then, of course, there is a glass coupler of technology that we acquired through the aiXscale acquisition. So the 3 of these things put together make up an optical engine that we can deploy with the switch as well as on the XPU, which is a fantastic multibillion-dollar market opportunity.

Yes. The aiXscale acquisition was actually quite interesting, right? Because when you talk to some of your partners and others out there that are working on these optical-based solutions, fiber coupling is the most challenging and critical barrier, right? EIC, PIC, I think that's relatively more mature technology. But when you listen to the CPO and NPO guides, it's all about fiber coupling reliability, scalability and so on. So I think the Astera team was smart in acquiring the fiber coupling technology first. I guess the question there is, what did aiXscale bring? What is the differentiation that they bring in terms of this very critical fiber coupling capability?

Yes. So I think maybe two things I'll mention. First of all, as you said, this fiber coupling is a mechanical aspect, right? Electrical IC, you can produce them by the bushel with TSMC is very good in their process nodes. Similarly, silicon photonics, at the end of the day, is a wafer-level technology. whereas the glass coupler is an individual piece that needs to be aligned and glued in place and so on. And so the aiXscale team has developed some very cool technology on how they do this alignment, how to do the mechanical attachment how they can make this a detachable technology so that if something were to fail, then you don't have to throw away the [ lens ], you just replace the part that fails. So I think that's one benefit that aiXscale brought.

The second one is, as I mentioned, this high radix capability. So when you want to connect, let's say, 72 and in the future, 144, 228 GPUs together, you want to be able to escape from the switch in a way that you can bifurcate these links into the number of GPUs. So that's a very different technology that is needed to build very high-density glass coupler, which is what aiXscale does.

Perfect. And then on the financials, long-term gross margin profile, 70% gross margin, 40% operating margins, you'll be driving close to that exiting this year by our estimates. That's almost two years ahead of our views back at the time of IPO. But as you layer on new initiatives like your custom silicon programs, optical scale-up initiatives as you roll out your standards-based PCIe UALink-based products, puts and takes around the sustainability or even outperformance of your long-term margin profile.

Yes. As you mentioned, our long-term operating model is 70% gross margin, leading to 40% operating income remains sort of unchanged. We expect to continue to grow faster than market given the investments that we've made and we'll continue to see the new products layering into the model as well as the customer diversification, as we talked about earlier.

In terms of the sort of gross margin, we are targeting 70%. We have been performing slightly above that and here past couple of quarters, but there will -- we do see that the margins will trend down towards our targeted model, driven by the warrant impact as well as the mix impacts and we'll see the modules versus silicon. We'll also see the additional impact of the additional switching use cases that will come into the model from there and the diversification of the portfolio. But 70% gross margin is very rich for a semiconductor sort of business and really is reflective of the value that our customers place on our solutions. We'll continue to invest at the right level in the business to make sure we can continue to grow the top line, and that will lead to the 40% best-in-class operating margins for the company from here.

That's great. Well, we are about out of time. Jitendra, Des, thank you for participating, and the setup for this year looks very strong. Look forward to being part of it. Thank you.

Thank you, Harlan.

Good afternoon. My name is Audrey, and I will be your conference operator today. At this time, I would like to welcome everyone to the Astera Labs First Quarter 2026 Earnings Conference Call. [Operator Instructions] I will now turn the call over to Leslie Green, Investor Relations for Astera Labs. Please go ahead.

Good afternoon, everyone, and welcome to Astera Labs First Quarter 2026 Earnings Conference Call. Joining us on the call today are Jitendra Mohan, Chief Executive Officer and Co-Founder; Sanjay Gajendra, President and Chief Operating Officer and Co-Founder; and Desmond Lynch, Chief Financial Officer.

Before we get started, I would like to remind everyone that certain comments made in this call today may include forward-looking statements regarding, among other things, expected future financial results, strategies and plans, future operations and the markets in which we operate. These forward-looking statements reflect management's current beliefs, expectations and assumptions about future events, which are inherently subject to risks and uncertainties that are discussed in detail in today's earnings release and in the periodic reports and filings we file from time to time with the SEC, including the risks set forth in our most recent annual report on Form 10-K. It is not possible for the company's management to predict all risks and uncertainties that could have an impact on these forward-looking statements or the extent to which any factor or combination of factors may cause actual results to differ materially from those contained in any forward-looking statements.

In light of these risks, uncertainties and assumptions, all results, events or circumstances reflected in the forward-looking statements discussed during this call may not occur, and actual results could differ materially from those anticipated or implied. All of our statements are made based on information available to management as of today, and the company undertakes no obligation to update such statements after the date of this call, except as required by law.

Also, during the call, we will refer to certain non-GAAP financial measures, which we consider to be an important measure of the company's performance. For example, the overview of our Q1 financial results and Q2 financial guidance are on a non-GAAP basis. These non-GAAP financial measures are provided in addition to and not as a substitute for financial results prepared in accordance with U.S. GAAP. A discussion of why we use non-GAAP financial measures whose difference is primarily stock compensation, acquisition-related costs and related income tax effect, and reconciliations between our GAAP and non-GAAP financial measures and financial outlook are available in the earnings release we issued today, which can be accessed from our website through the Investor Relations portion of our website.

With that, I'd like to turn the call over to Jitendra Mohan, CEO of Astera Labs. Jitendra?

Thank you, Leslie. Good afternoon, everyone, and thanks for joining our first quarter conference call for fiscal year 2026. Today, I'll update you on AI infrastructure market trends, our Q1 results and recent announcements. I'll then turn the call over to Sanjay to discuss Astera Labs' growth profile. I'd also like to welcome Des, our CFO, joining this call for the first time. Des will cover our Q1 financials and Q2 guidance.

Since our last earnings call, AI infrastructure spending has clearly accelerated. Hyperscalers, AI labs and sovereign entities are signaling the industry's build-out is still in its early stages, underpinned by strong monetization and ROI. We expect these strong secular trends to be a tailwind for Astera Labs' growth over the long term.

Astera Labs delivered strong results in Q1 with revenue and non-GAAP EPS above our outlook. Revenue for the quarter was $308 million, up 14% from the prior quarter and up 93% versus Q1 of last year. Revenue growth was broad-based, spanning across our signal conditioning and fabric switch product portfolios as we continue to diversify our business profile with new design wins across multiple customers and product categories.

Our PCIe 6 business across both AI fabric and signal conditioning was strong in Q1 with revenue expanding to more than 1/3 of our total revenue. We have now shipped millions of PCIe Gen 6 ports to date, demonstrating the robustness and maturity of our PCIe portfolio. Product smart cable modules for Ethernet AECs continue to perform well as new program designs ship in volume, while others ramp to mature levels across GPU, XPU and general purpose systems.

On the scale-up fabric front, our initial design wins with Scorpio X-Series in smaller radix configurations shifted from preproduction shipments to initial volume ramp during the first quarter.

Building on this momentum, today, we announced the expansion of our Scorpio product line of AI fabric switches for both scale-up and scale-out use cases. Scorpio X-Series portfolio now supports up to 320 lanes for high radix scale-up networking, and Scorpio P-Series PCIe 6 portfolio now spans 32 to 320 lanes for diverse system topologies, making it the broadest in the industry.

Our new flagship Scorpio X-Series 320 lane has been purpose-built to maximize AI economics by leveraging hardware-accelerated hypercast and in-network compute engines to boost collective operations by up to 2x. In-network compute offloads critical accelerator to accelerator communication and computation directly onto the switch, dramatically reducing the networking overhead during large-scale training and inference. These hardware capabilities are delivered through enhancements to our COSMOS software, which can now integrate deeper into our customers' software stacks, providing not only diagnostics and telemetry, but also directly improving AI platform performance.

Scorpio's advanced hardware and software capabilities are a result of Astera Labs' deep system-level understanding of AI architectures and close customer collaborations, creating a durable competitive moat.

We are excited to report that we are now shipping initial volumes of our new 320 lane Scorpio X with production volumes ramping in the second half of 2026. Scorpio X-Series also has widening interest in design activity with hyperscalers, edge AI inference providers and enterprise infrastructure builders to address high-bandwidth, AI clustering use cases. Scorpio P-Series continues to grow through 2026, and we expect initial shipments to at least 2 additional major hyperscalers towards the end of 2026 with broader deployment in 2027.

On the optical front, we made good progress during the quarter as we continue to work through the qualification process at a large AI platform provider with our ultra-high precision optical fiber coupler product, which we expect to ship in volume starting in 2027. We are actively expanding our volume manufacturing capabilities to support the ramp of both scale-out and scale-up GPU applications. Beyond the early commercial traction of our merchant connectors, our high-density fiber coupler technology will be a critical piece of our long-term optical road map for NPO and GPU applications.

Finally, our LEO memory controller is on track for an early ramp of CXL attached memory with Microsoft Azure M-Series virtual machines. And during the quarter, we captured a new custom design win for a KV Cache offload application with shipments expected in 2027.

As we look to the second half of 2026, robust demand reflects secular AI infrastructure spending, deep customer partnerships and expansion towards higher-value solutions within our portfolio. This trend is quickly increasing our silicon dollar content opportunity beyond $1,000 per XPU within AI racks and positions Astera Labs to outperform our end market growth rates. As a result, we expect strong revenue growth to continue through 2026 and into 2027, driven by the proliferation of AI fabrics and the industry's transition to PCIe 6, 800 gig and 1.6T Ethernet connectivity.

Based on the momentum we are seeing in 2026, we are strategically investing to drive strong continued growth. Our acquisition of aiXscale Photonics has created immediate design opportunities, and our Israel design center is fully integrated and working with customers on new programs. We have expanded our product portfolio and increased dollar content per accelerator while diversifying our customer base with additional design-ins. We are making progress within large market opportunities, including optical engines and interconnects, UALink fabrics and custom solutions for NVLink and AI inferencing.

Most of all, I'm proud of the stellar team we have built through worldwide hiring and thoughtful acquisitions, the progress we have made and the results we are delivering together. With that, let me turn the call over to our President and COO, Sanjay Gajendra, to outline our vision for growth over the next several years.

Thanks, Jitendra, and good afternoon, everyone. Today, I will provide an update on our recent execution, followed by an overview of the meaningful market opportunities that will fuel Astera Labs growth over the next several years. Astera Labs' mission is to deliver a purpose-built intelligent connectivity platform with a portfolio of standard, custom and platform-level solutions across copper and optical interconnects for rack-scale AI infrastructure deployments.

As AI deployments advance to production at scale and operational efficiency, infrastructure teams face a new set of constraints, multitrillion parameter models, agentic workflows, multistep reasoning distributed across heterogeneous compute infrastructure to name a few. The industry needs connectivity solutions purpose-built to address these workloads. Higher ratings to simplify topologies, intelligent fabric capabilities to reduce communication overhead, open and platform-specific optimization and data center-grade diagnostics to maintain uptime when a single fault can cost millions of dollars in idle compute.

Let me now walk through our approach to address the evolving needs and our future strategy, starting with our standard products. We continue to see strong momentum across both AI fabric and signal conditioning portfolios. We strengthened our mission-critical position with the introduction of our flagship Scorpio X-Series 320 lane scale-up fabric switch and the overall expansion of our Scorpio Switch portfolio. The Scorpio X-Series 320 lane high radix AI fabric switch replaces multiple legacy switches to enable large scale-up cluster sizes in a single hop and reduces overall latency. Several new features such as in-network compute, reduced time to first token and tokens per watt performance. The newly expanded Scorpio P-Series PCIe switch portfolio now spans from 32 lanes to 320 lanes to enable diverse accelerator optionality and system topologies.

Our AI fabric portfolio is poised to expand further into 2027 with the introduction of UALink-based products for AI scale-up platforms. In early April, the UALink consortium published a new specification, which defines in-network compute chiplets, manageability and 200-gig performance. UALink 2.0 delivers these advancements with an open vendor-neutral approach and confirms that scale-up switching is not simply hardware, but an AI-aware fabric actively helping the system compute and drive performance.

This evolution plays into Astera Labs' strength as demonstrated by the industry-leading feature set that are being deployed through our Scorpio portfolio expansion today. The maturity of the ecosystem is also accelerating with customers and suppliers working tightly to deploy initial programs in 2027.

On the signal conditioning portfolio, our [ AT ] products will expand to support PCIe 7 and our Taurus portfolio into 1.6T Ethernet, positioning us at the forefront of the next connectivity upgrade cycle.

Turning to our optical business. Astera Labs signal connectivity business is driven by the rapid shift of AI systems towards rack scale architectures and higher compute capabilities where scaling performance increasingly depends on high bandwidth, high radix, low-latency interconnects. These requirements will expand our AI connectivity opportunities across both copper and optical interconnects.

Astera Labs is well positioned to lead this transition by extending its proven value chain approach from copper into optics. Over the past couple of years, we have been systematically investing to broaden our internal capabilities across advanced analog and mixed signal design, DSP, electronic ICs, photonic ICs and optical packaging capabilities while also deepening our supply chain relationships. Together, these capabilities will enable high-volume deployment of a complete scale-up optical engine.

We are focused on 3 areas pertaining to scale-up optics. One, high-density detachable, refillable fiber attached solutions using the core technology from our aiXscale acquisition. We expect to ship these connectors in volume starting in 2027; chipsets in support of NPO that will enable multi-rack AI clusters starting in 2027; and eventually fully optically enabled Scorpio X fabric switches with CPO supporting larger domains, higher egress densities and bandwidth.

Next, let me talk about our custom solutions business that also continues to make meaningful progress as we work to develop new products and close on new designs. Once again, tight collaboration with hyperscaler customers, coupled with a diverse set of foundational technology and operational capabilities have been essential to our initial success. These opportunities represent a new multibillion-dollar market opportunity for Astera Labs.

First, we are engaging with multiple customers to enable NVIDIA fusion scale-up architecture for hybrid racks. Our strong historical execution, delivering intelligent connectivity solutions for NVIDIA-based systems positions us well to develop and design within these new custom programs.

Second, we are seeing new custom solution opportunities within the memory space for KV Cache applications. We are happy to report that we have won a new design, leveraging a customized version of our LEO CXL controller to maximize performance within these AI use cases. Overall, we are pleased with the initial traction we have seen on the custom solutions front and have conviction that this opportunity set will continue to broaden and become a meaningful business for Astera Labs over the next few years.

Finally, we continue to demonstrate solid momentum with our platform business as we ultimately look to expand beyond our add-in cards and smart cable modules to enable broader rack scale solutions for customers. As we have grown from an I/O component supplier to an AI fabric solution provider over the past couple of years, customers are looking for Astera Labs to bring additional value to the AI rack at a system level.

In conclusion, Astera Labs is at a key inflection point in the company's journey as we begin to ship production volumes of our scale-up AI fabrics. We are also making great strides towards broadening our business across new product categories, including optical and custom solutions as our partners look for us to deliver more value in next-generation systems. Therefore, we will continue to strategically and thoughtfully invest as we position Astera Labs to deliver growth rates above our end market benchmarks over the long term.

With that, I will turn the call over to our CFO, Des Lynch, who will discuss our Q1 financial results and our Q2 outlook.

Thank you, Sanjay, and good afternoon, everyone. I'm pleased to be joining you today for my first earnings call as the CFO of Astera Labs. I look forward to partnering with Jitendra, Sanjay and the rest of the leadership team as we continue to drive long-term value for our shareholders.

Today, I will begin by reviewing our Q1 financial results and will then discuss our Q2 guidance, both presented on a non-GAAP basis.

Revenue in the first quarter of 2026 was $308.4 million, which was up 14% versus the previous quarter and up 93% year-over-year. We saw revenue growth across our signal conditioning and switch fabric portfolios, supporting both scale-up and scale-out connectivity for AI fabric and reach extension applications.

Our Scorpio product family performed well in Q1, driven by strong demand for PCIe Gen 6 switching applications and continued expansion of designs across various platforms. During the quarter, Scorpio X-Series products began shipping in initial production volumes. Looking ahead, we expect Scorpio X-Series shipments to increase in Q2, along with initial shipments of our new Scorpio X 320 lane and then ramp to full volume production in the second half of 2026.

Aries revenue grew on strong early adoption of our PCIe 6 solutions for both scale-out and scale-up signal conditioning. In total, PCIe Gen 6 revenue across AI fabric and signal conditioning contributed more than 1/3 of total company revenue in the quarter. Taurus also delivered solid results driven by broad adoption of AEC to extend reach in both AI and general purpose compute platforms.

Non-GAAP gross margin for the first quarter was 76.4%, up 70 basis points sequentially, primarily driven by a lower mix of hardware sales across our signal conditioning portfolio. Non-GAAP operating expenses for the first quarter were $123.9 million, reflecting continued R&D investment to support our expanding product road map, including a full quarter of our aiXscale acquisition and a partial quarter of our newly formed Israel Design Center.

Within Q1 non-GAAP operating expenses, R&D expenses were $96.2 million, sales and marketing expenses were $12 million and general and administrative expenses were $15.7 million. Non-GAAP operating margin for the first quarter was 36.2%. We will continue to invest strategically to drive above-industry revenue growth over the long term while maintaining strong and durable profitability.

For the first quarter, interest income was $11.6 million. Our non-GAAP tax rate was 11%, and non-GAAP fully diluted shares outstanding were 181.2 million shares. Non-GAAP diluted earnings per share for the quarter was $0.61.

We ended the quarter with cash, cash equivalents and marketable securities totaling $1.18 billion, flat versus Q4 as cash from operations of $74.6 million was offset by cash paid for acquisitions.

Now turning to our outlook for the second quarter. We expect revenue to be between $355 million and $365 million, up 15% to 18% sequentially, driven by continued strength across our AI fabric and signal conditioning portfolios. Aries revenue growth is expected to be driven by continued strong adoption of PCIe 6 across AI platforms, supporting both scale-up and scale-out connectivity. Taurus growth is expected to be driven by increased volumes for AI scale-out connectivity. And in AI fabric, we expect robust growth driven by the continued early-stage ramp of our Scorpio X-Series products for large-scale XPU clustering applications as well as continued growth in our P-Series solutions and customized GPU platforms.

We expect second quarter non-GAAP gross margin to be approximately 73%. This outlook includes an estimated 200 basis point noncash impact related to a recently executed warrant agreement with one of our customers. We expect second quarter non-GAAP operating expenses to be between $128 million and $131 million. Interest income is expected to be approximately $11 million, and we expect our non-GAAP tax rate to be approximately 12%. We expect our Q2 share count to be 184 million diluted shares outstanding. Overall, we are expecting non-GAAP fully diluted earnings per share to be between $0.68 and $0.70.

This concludes our prepared remarks. And once again, we appreciate everyone joining the call. I will now turn the call back to our operator to begin Q&A. Operator?

[Operator Instructions] We'll take our first question from Harlan Sur at JPMorgan.

Great job on the execution by the team. As your customers compute workload inflection from training to inference in the second half of last year, essentially very focused now on monetization, right? We saw that as inferencing workloads evolved one shot to reasoning to now agentic, right. This created new silicon opportunities, right? It created new storage tiers. It created more demand for high-performance CPUs. Obviously, storage and CPUs communicate via PCIe, like so right in the sweet spot of your technology and product leadership, right, that's one example. Your CXL solutions targeted at KV Cache applications, maybe another example. But kind of help us understand how the transition to more inferencing-based workloads, especially agentic-based workloads, has potentially helped to create new opportunities for the team and potentially expand your SAM opportunity.

Harlan, thank you. This is Jiten. Let me try to take a stab at that. You point out very correctly that inferencing has created a lot of focus in the industry and a lot of additional opportunities. The good news is that at Astera, we've been focused on these AI applications from the start. And we helped the training workloads when the training workloads were still the mainstream. And now we are helping the inferencing workloads equally well.

The KV Cache offload is a great opportunity where we mentioned earlier that we picked up a new design for a custom application. For KV Cache offload, that's really a key part of AI inferencing. I also want to draw your attention to the newly introduced core PX 320 lane family that supports in-network compute and hypercast. Both of these are extremely important technologies to reduce the networking overhead and then deliver additional performance for training as well as inferencing.

Not only that, we enable these hardware accelerated modes through our COSMOS software, which now not only gives our customers the ability to do diagnostics and telemetry, but allows them to uniquely improve the performance of their system for their inferencing workloads using these unique capabilities that we have worked in tight collaboration with our customers.

We'll move to our next question from Blayne Curtis at Jefferies.

I'll echo the congrats on the nice results. Maybe you can -- in terms of the Scorpio ramp, I know last quarter, you talked about it being 20% of revenue. It's a big ramp. I'm assuming that's the biggest driver into June. I was wondering if you can kind of frame just how big that is. And then I'm curious, particularly this 320-lane product that's ramping, like what are the milestones? And what's left to do? You've sampled it, but to get that to production in an AI server, I'm just kind of curious what's left there.

Blayne, it's Des. Thanks for your question. We've been very pleased with the performance of our Scorpio product family. It's certainly been a large driver of our growth in the sort of first half of the year. We continue to expect to see Scorpio P continuing to ramp driven by scale-out opportunities. And in Scorpio X, this is really a greenfield opportunity for us associated with scale-up connectivity. The small radix solutions are ramping today, and we do expect to see the layering in of the high radix configurations in the second half of the year.

Given the size of the opportunity and the associated dollar content, we would expect to see that Scorpio will become our largest product line by the end of the year, which is strong performance for a product line that was only 15% of total company revenue last year. And as we go throughout the year, I would expect to see X-Series revenue exceeding P-Series. But overall, we're very pleased with the performance of the Scorpio product family and the outlook of the business.

Blayne, to your second point about other milestones, we are already shipping, as Des mentioned, the Scorpio X, the newly introduced Scorpio X family. And you'll be able to see and touch and feel this at COMPUTEX where we will be demonstrating this live in our booth.

We'll move next to Joe Moore at Morgan Stanley.

You talked quite a bit about your optical strategy. Can you -- I guess, can you talk about the time frame where you see optical scale-up becoming more relevant? And do you have the building blocks that you need to progress from copper to optical in that space? Do you need tuck-in type technologies? And do you need to invest a lot more? Just a general sense of what it's going to take to transition from copper to optical over the next several years.

Thanks for the question. This is Sanjay here. Yes. So we have been working for the last couple of years building all the foundational things that are required for optical enablement, all of the mixed signal technology that's required, all of the electronic IC as well as we have did the acquisition with aiXscale that brought in the pluggable connector as well as the PIC technology. So in general, I want to say we have made tremendous progress in preparation for the optical opportunities that are coming up on us.

For us, in terms of time line, what we believe is that the NPO-based opportunities, or the near package optics, would be the first one to ramp, and that will start happening in 2027. We will also be ramping our pluggable connector technologies for CPO, mostly for scale-out next year, 2027, with more of the mainstream deployments for CPO happening in the 2028 time frame.

So in general, for us, between the components that we are building that go inside the NPO, the detachable connector technology for folks that have their own CPO solutions as well as our own Scorpio X devices that will come in to support both NPO variants and CPO variants, we believe it's been all coming together nicely for us. One key consideration, of course, that we've been working is the supply chain and getting all of the commitments in place so that we can not only provide the technology that's required for NPO and CPO, but also make sure that we are able to ship to revenue. And I think overall, there's quite a bit of work and progress that we have done and enabling us to start ramping in 2027.

We'll take our next question from Ross Seymore at Deutsche Bank.

Congrats on the strong results and guide. I just want to talk about a small part of your business today, but something that sounds like it could grow a little faster than we thought before, and that's specifically your LEO product line. Given the dominance or resurgence of the CPU demand and memory being such a large cost and bottleneck these days, how is the demand trajectory and growth potential changed in your view on your ability to kind of do the cooling and the sharing on the memory side and CXL in general?

Yes, we are definitely seeing increased traction for CXL, not only for the general purpose compute application where we started, but also for AI inferencing, as we touched upon earlier.

Just kind of staying with general purpose compute first. We are seeing additional demand from our customers. We are on track for deploying this with Microsoft Azure for the M-Series instances at the data center. So that's in private beta now expected to go into general availability end of the year. And we see additional customers also kind of following suite for this particular high-memory type application. In addition, we are also excited by the new KV cache offload or AI inferencing opportunities where some of our customers have already designed us in. In fact, we picked up our second design win of a custom application for CXL earlier this quarter. And we are working with our customers, which is an additional new hyperscaler. We are working with them on at-scale performance tests and expect that one to ship for revenue in 2027.

We'll go next to Tore Svanberg at Stifel.

Congrats on the record quarter. And Des, welcome on board. I wanted to follow up on what you said about Scorpio mix as we approach the end of the year, especially in relation to Aries because obviously, Aries is now ramping in TS Gen 6. Next year, obviously, there's going to be a lot of mix networking topologies. So I understand Scorpio will be the sort of biggest product by the end of the year. How should we think about '27 between Aries and Scorpio because obviously, there are significant drivers for both.

Tore, thanks for the question. Yes, we've been very pleased with the growth rates of our Scorpio product family. As I mentioned earlier, really excited about the continued growth opportunity ahead of us. That said, we still expect to see strong growth within the Aries product line. We expect to continue to grow our sort of leadership position there. We expect to see strong growth given the PCIe 6 portfolio. It's just the fact that Scorpio will continue to be our sort of largest and fastest-growing sort of business within the company here.

Next, we'll move to Ananda Baruah at Loop Capital.

And yes, congrats on the great execution here. I guess the question, guys, would be, what's a good way, particularly with all the additional context you've given around Scorpio X and Scorpio P lanes progressing through the back half of '26. As we move forward post '26 and clusters get bigger and presumably high-rated switches have more ports, should we expect Scorpio X and Scorpio P switches to continue to increase the lane count? And if so, what -- is there any useful anecdotal way to think about like how that may occur? Or should we just think that, that sort of can continue in some perpetuity?

Thanks for the question. We can talk for an hour just on that topic. But let me say this. The AI fabric switches have become a very important part of our overall strategy, and we are investing heavily in not only the current generation that we have announced, but also upcoming devices. We are going to continue to focus on PCI express because that is a large part of the business today. But we are also working on UAL products that will form the basis of the next generation of these devices.

In terms of the lane count, et cetera, we work very, very closely with our customers to understand what their deployment profile is going to look like because it's really important to target the right lane counts and ratings for these devices because if you don't, then the customer sizes get limited and if you over-index, then you come up with a solution that is not competitive. Fortunately, we have very, very good partnerships with our customers, and they are telling us what the deployment looks like.

And I also want to add to that, that as the sizes increase, not only is it important to have a switch, actually, it is also important to have the right media types for the deployment. So for our family of switches, we will continue to support copper connectivity as we have so far. But as Sanjay mentioned earlier, increasingly, we will enable optical connectivity as well, starting with NPO with the next generation of switches and then going to CPO. And it is very important to understand that as a switch company, it gives us a perfect opportunity to deploy optical solutions. And that's something that we will completely leverage and make sure that we support an end-to-end connectivity with our switches, including copper, NPO and CPO.

We'll take our next question from Natalia Winkler at UBS.

Congrats on the results. I was wondering if you can add a little bit more color on the NVLink Fusion opportunity for you guys. Specifically, how do you see from a standpoint of portfolio, maybe where it would be most interesting for you? And also from the standpoint of competitive landscape given some of the partnerships that NVIDIA has for the NVLink Fusion as well.

Yes. Thanks for the question. So in general, if you look at our business, you can broadly divide that into 3 categories. There is a standard business that we're doing; the custom business; and then, of course, the module and the solution business that we have. Clearly, an area that we see tremendous opportunity for us going forward is the custom solutions under which we are developing the NVLink Fusion type of devices. And this actually is proving to be pretty interesting. We do have several opportunities. We're very deep in engagement for an initial design win in collaboration with NVIDIA and then a hyperscaler. So that project is going well. So we do expect that to start contributing revenue in 2027 as some of the GPUs that are designed for this kind of use case, which is called as a hybrid rack situation, where the GPU or the XPU still talks native protocols, which could be a protocol like PCIe or UALink and others. But then when they need to leverage and cross over and talk to an NVLink type of ecosystem, then they would need a product that's based on NVLink Fusion that we are developing.

So in short, I would say that we are very deep in engagement from a development -- silicon development standpoint. So we do expect that this will start providing some meaningful revenue in 2027 and then growing from there.

The second part of your question was competitive situation. I mean, obviously, this is an ecosystem that NVIDIA is creating with NVLink Fusion. There are others. But for us, the main thing is that we have been engaged with real customers, real applications. And to that standpoint, we will continue to focus on that and do what we need to do and not get distracted by any competitive press releases.

We'll go to our next question from Sebastien Naji at William Blair.

Congrats on the strong results. My question is on the Scorpio business and maybe a little bit of a follow-up to one of the prior questions. But with your announcement of the new 320-lane Scorpio switches for both the X and P-Series, how should we be thinking about ASPs for the higher rated solutions? Is it right to think that your dollar content is correlated directly to the lane count? Or is there another way to think about your dollar content? Just any details there?

Yes. So in general, what I would say is the bigger the switch, the higher the ASP. That's the way industry works. But also, please keep in mind is that these switches are more like AI fabric class device, which are a lot more than just the number of lanes, right? We talked about in-network compute. We talked about hypercast. We talked about several features that we have that are unique and critical for deploying the AI clusters, whether for training or more and more for inference application where things like latency becomes super important.

So when it comes to ASP, obviously, it's a combination of how -- what features are enabled and not just based on the port count. But we do see that our content continue to increase. And to that standpoint, we are expecting and going forward with the design wins we have, over $1,000 worth of content per accelerator. And that is, of course, significant, and it's growing rapidly for us if you consider the path that we have taken so far from offering retimers now to offering complete AI fabric. And with the future products with optically enabled switch and so on, you can only imagine that this content would grow from a dollar per accelerator standpoint.

We'll go next to Quinn Bolton at Needham.

Let me offer my congratulations as well. I guess you mentioned just the KV Cache offload custom design. I'm wondering if you might be able to put any sort of numbers around it in terms of like dollar content per CPU or dollar content per gigabyte or terabyte of memory that's attached. Is there a way we can think about how that opportunity or how to size that opportunity?

Yes. So these are going into new inference applications. So there are multiple use cases and platforms that we see for this. So in that context, this would be a significant opportunity for us to execute and deliver on.

In terms of exact dollar association with this, I would say it's probably a little bit early because some of the platforms and architectures are being finalized. But in general, for us, like was highlighted earlier, inference and KV Cache is a significant opportunity. We have the IP, not just for memory, but also for things like KV Cache acceleration and so on as part of our portfolio right now. So to that standpoint, we will increasingly develop products that provide more function and capability to ensure that memory is available for KV Cache use cases.

And I will also say that the ASP also would be -- continue to be pretty meaningful when you think about the cost of the memory. In other words, these controllers will always fade compared to the amount of money that people are paying for the memory itself. So in some ways, what I'm trying to say is that these are not ASP challenged, and we will continue to make sure that we extract the most value out of these products.

We'll move to our next question from Karl Ackerman at BNP Paribas.

This is Sam Faerman on for Karl Ackerman. So you mentioned near packaged optics as a preliminary solution to CPO. From Astera Labs point of view, do you believe customers view XPO as a viable option to extending pluggable optics? And does Astera plan to participate in the XPO NFA?

That's a great question. We clearly work very closely with our customers to understand what solutions they are looking for. XPO is a pluggable technology that has come about recently, and we will certainly participate in that. But not all of our customers, at the moment, are looking to intercept XPO. So the customers that are looking to intercept with NPO, we will certainly support them because it gives you a way to have a very high egress density without the constraints of front plate density. The customers that want us to work directly on CPO, we absolutely will work with them. As Sanjay mentioned earlier, we are engaged in that opportunity that should ship here in 2027. And for customers that are looking to do XPO, we will engage with them as well. But right now, our focus has been on NPO and CPO so far.

We'll take our next question from Suji Desilva at ROTH Capital.

Welcome, Des. Just a bigger picture question. I mean you mentioned the word custom quite a bit on this call more than in the past. When you first IPO-ed, hopper was there in areas that was fairly standard. Are we past kind of the point or evolving to the point where standard products are not as applicable to you because each platform is different? And should we think of all products having some customization? Or where is the line there? Just trying to understand.

Yes. I'm glad you asked the question. So if you think about infrastructure and AI use cases, they all are bespoke, and they all are unique between platform and between customers. But having said that, if you look at the software-defined architecture we have with our products, these are devices even our standard products, Aries and Taurus Scorpio and so on, they provide a ton of customization that customers leverage through the COSMOS interface. And COSMOS allows them to not only monitor but also customize. And now with the new devices that we announced today, they can do a lot more from a performance and other key offload feature enablement and so on.

So all in all, customization has been our story through software-defined architectures and offered through our standard products. But when we talk about our custom business, the business model is different, right? So we are developing a product for a given customer under a business model that would include NREs and other ways of paying for the development and of course, the product revenue that comes when the product starts shipping.

So in general, what we see is that as we're getting into bigger devices, whether it's for fabric class or other connectivity type of technology that goes beyond sort of what we have done so far, having the custom solution portfolio is important, and we are approaching that with our customers by also offering a variety of foundational technology that we've been building for the last couple of years.

So in general, we would see custom being an important growth driver for us. But at the same time, please think about our business in a way where the standard products will continue to be a very important part of our overall portfolio. We will do custom, but we won't be -- we will be very systematic about it. We will not take any opportunity that comes our way because sometimes the custom business can lead into -- could be so unique to one customer and with a lot of risk and margin and so on. So to that standpoint, we will try to make sure that we are systematic and thoughtful about the opportunities that we pursue on the custom side.

We'll go next to Mehdi Hasini at Susquehanna.

This is Bastian filling in for Mehdi. Congrats on the quarter and welcome, Des. I guess I wanted to follow up on UALink. Can you share an update on the adoption process and the expected time line for UALink-based switches? And what do you expect the dollar content to be? How should we think about the difference between kind of the PCIe switching pricing and the UALink pricing?

Yes. So I think within the last 3 months or so, maybe 6 months or so, we've had a couple of announcements from our hyperscaler customers on what the intercept is. Both Amazon as well as AMD have said that their respective ASIC and GPU will launch sometime in 2027, and we'll certainly be prepared to intercept that launch with our UALink switch.

In terms of the comparison of our UALink switch to PCI Express, maybe a couple of things to state. First, as we go into the new generation of devices, both the complexity as well as the speed of these devices is going up, sometimes in case of lane count, other times in case of radix. So the value that we are able to charge for these devices will be substantially higher in terms of what we are able to do for the PCI Express switches.

The second thing that I'll mention is the media attach also tends to change. So we may go from kind of majority PCI Express to a blend of majority copper to a blend of copper and NPO with the next-generation switches that also gives us a meaningfully large opportunity in terms of revenue and the TAM that we are able to address. Finally, leading up to a CPO, which is a really rich opportunity with a very large TAM that we are able to address all because we have the platform in the form of Scorpio X switches.

And we'll move next to Tore Svanberg at Stifel.

Yes. I just had a quick follow-up on capacity. So your inventory days, I think, came in at 75 days. It seems like a little bit at the lower end. But I guess, are you sort of feeling good about being able to continue to at least double revenues here and next year based on the capacity commitments you have today?

Tore, it's Des here. Yes, based upon our current view of demand, we do have supply in place through the end of the year, and we're very comfortable with what our inventory sort of holdings are here. Like others within the industry, we continue to see pockets of supply challenges. But what we've done here is really a nice job of diversifying our back-end supply chain, and we've been able to make sure that we have the sufficient supply in place to make sure we can meet the revenue commitments. So no concerns just now, and we continue to work with our supply chain partners for supply going into 2027.

And that concludes the question-and-answer session. I'll turn the call back over to Leslie Green for closing remarks.

Thank you, Audrey, and thank you, everyone, for your participation and questions. Please do refer to our Investor Relations website for information regarding upcoming financial conferences and events. Thanks so much.

And this concludes today's conference call. Thank you for your participation. You may now disconnect.

Hi, good morning everybody. So Suji Desilva, systems analyst at ROTH. Usually, I have to go back and look at somebody's bio when I'm meeting them, but I've known Nick most of his career. We started talking to each other when back at NetLogic and Broadcom and MaxLinear. He's been around the semis industry for a while, and he's had the fun opportunity to help Astera Labs come public recently, a very exciting story in the AI space that we've helped. ROTH has been along for the ride with and privileged to have you here. Nick, it's really great to talk about Astera's story and the success level because the numbers just keep surprising us, and they're pretty amazing. So thanks, Nick, for your time, of course. Welcome.

So the -- maybe just -- I keep -- so one of my companies told me this morning that the model technology they built 2 years ago is already outdated and wouldn't really be useful in today's effort. That's how quickly this is all changing. So maybe you can talk about where you sit and how you're seeing that these models are evolving so fast that your technology is helping people keep up. And we'll get into kind of what Astera does in this kind of through the questions, but maybe we could start there and what you're seeing because that's the flywheel you guys are helping support, I think.

Yes. Thanks, Suji, for the invite as well, and I appreciate having a mic. I feel like I'm channeling my inner rapper right now, something ...

You got a station.

Yes. But yes, I mean, the treadmill is running really, really fast. And fortunately, for Astera Labs, we get to have a ringside seat to the speed and the cadence that we've become accustomed with. And we've been fortunate enough to be able to stay on that treadmill and not get knocked off as things continue to go very quickly.

And as an example, I think when we came public just a little over -- right around 2 years ago, an AI server was defined as an HGX Hopper-based system that had 8 GPUs on it, and that was the extent of an AI server 2 years ago. Fast forward to today, and now we're scaling up over 72 GPUs within a single rack and looking to branch out beyond that. And then you had Jensen at GTC last week talking about scaling up across over 1,000 GPUs.

So Astera Labs in terms of positioning within that infrastructure is in a very unique position such that we are the nervous system of these boxes, and we're helping to connect very important endpoints such as GPUs, but also CPUs, networking, interface cards, memory, storage and all of these things are talking together and to each other through typically protocols such as PCI Express. And we've benefited just due to the increasing speeds of these boxes and the increased size of the clusters, you start to run into a lot of problems when you talk about both speed and distance.

So we've had a nice ringside seat. Execution has been great over the last couple of years. And what it's afforded us on a go-forward basis is a seat at the table with hyperscalers defining what the next generation and the next generation after that is going to look like. What are the challenges going to be in those boxes from a connectivity standpoint and how can we help to service those.

So to your point, the secular trend is very much alive and well. The demand for just raw compute is going off the charts. And our customers and our partners are all trying to solve how to deliver that compute in the most productive and efficient way possible. So that's where we're really linking up with these guys. So yes, I mean, we're well positioned. We benefit from the secular growth trends, but we also continue to bring out new products and deliver new solutions so that we can grow faster than the market.

I promise we'll go through the product Astera offers. And the next question is the question before that, that I'll lead to that. What is it about -- you've been in Ethernet and Broadcom. You've seen the cloud development. What is it about these workloads and infrastructures that is so demanding of the connectivity part of this equation that you see that you have to adapt your products?

Yes. I mean it really is kind of coming back to the speed and size or speed and distance. So every generation, we're pushing forward starting at -- for Ethernet, for example, going from 25 to 50 gig per lane, just starting to kick off 100 gig per lane, going to 200 going to 400. Every single time you go to that next generation of speed, it limits the amount of distance that you could travel and send signals with high fidelity.

And while that is happening, you're going from boxes that have 8 GPUs up to 72 up to hundreds and hundreds and hundreds of GPUs. So the distances are expanding at the same time that the speeds are going up. So this creates a tremendous amount of bottleneck around the connectivity portion of the system when compute needs to happen at incredible speeds, but you need to take and have a pipe that's going to be able to transport those signals back and forth.

So I would say that, that's the hugest piece or the largest piece of the problem. I think the other piece that ourselves and others are trying to solve as well is no 2 clouds are the same. No 2 hyperscalers are the same. They all have different ways of approaching this problem. They are all building their own internal solutions that have different characteristics and attributes. They're also leveraging merchant solutions from the likes of NVIDIA and AMD as well.

So from a connectivity standpoint, you really do need to leverage your focus on open platforms, open systems, open protocols. You need to have flexibility built into your architecture. You need to have a software framework that's able to adapt to a multitude of different processors, a whole slew of different brands of memories and network interface cards and CPUs, and it all needs to work together regardless of how it's mixed and matched.

So it's not as simple as just bringing a chip to the table that's performance and it can support speeds and push distance. It also needs to be able to interact with a whole variety of disparate endpoints and do it in a way that's invisible to the customer. At the end of the day, these guys are fleet managers sitting in a data center presiding over tens of billions of dollars worth of infrastructure. And to monetize that, you need it up and running and performing at the highest percentage that you can.

So we can sit in the connectivity nervous system and be the eyes and the ears of these systems and provide valuable feedback and insights to these fleet managers to notify them if power surgers are coming online, links are going down, all types of different analog attributes. And therefore, they can take that data and go solve problems before they become big problems. So that's how we're trying to help solve this piece.

I guess the key metric for these folks is the uptime really that that's the currency of AI is what you're saying. So as engineers, you get to name your product whatever you want. And so you can use Marvel characters or these guys have chosen astrological names. So we'll start with the Aries product, which is the retimer, which seems like a fairly straightforward product, yet it had profound benefits. And maybe, Nick, you can touch in the discussion of Aries on Copper versus a lot of the talk about optical. We're going to have that theme maybe throughout this conversation, but we can start there.

Yes, sure. So Aries is kind of our flagship solution that we cut our teeth with kind of 2, 2.5 years ago, designed across 100% of NVIDIA-based platforms as they ramped Hopper into high volume into the marketplace. And even until today, we ship Aries into every single major U.S. hyperscaler customer on the planet and every single merchant GPU provider on the planet as well.

So very broadly adopted, very successful product line that's driving hundreds and hundreds of millions of dollars of revenue. And it's going to continue to grow going forward. As we just discussed, as distances continue to increase and speeds go up, you're going to need retimers in places you didn't need them before. So we expect attach rates over the long term to continue to increase.

And as you move from each subsequent generation of product to another, talking about PCI Express Gen 5 moving to Gen 6, ultimately moving to Gen 7, ultimately potentially move to UAL. Each of those products are going to be more performant, more complex, more capable than the previous generation and will command an ASP increase for those as well. So yes, this is definitely kind of the foothold of how we got our start. Effectively, Aries has been deployed across every major data center deployment for AI over the course of the last couple of years. And we're firmly entrenched from that standpoint to have COSMOS and Aries be widely deployed and battle tested across the board.

In terms of optics, and we can, of course, kind of start to unpack that a little bit as we talk about Scorpio as well. But our viewpoint is that optics is going to be a big TAM adder for us. We have a very large $25 billion TAM to prosecute today just on the copper and electrical side alone. But when you start to bring optical into the mix, you're going to need additional components around the components that we already supply. And those components are not cheap.

So probably a gating factor for massive adoption tomorrow. But as you need them, as distances get further and further, further away and you start to cluster and scale across multiple racks and not just one rack, we'll be definitely looking to sell optical engines around Scorpio, optical engines to our XPU customers to bolt around their XPUs and then potentially on a discrete basis, different pieces of an optical engine to folks that want to leverage our technology.

Now let's move on to the Scorpio product, which is going to be -- which the PCI switch as opposed to retimer alone, and that is ramping in volume revenue contribution significantly in the second half of this year. What is the critical element that a PCI switch is needed for and provides because it's an ASP uplift for you as well on a unit basis? So all those thoughts of yours?

Yes for sure. So the application and the use case spans well beyond what a retimer does. Of course, the retimer is very important, and it's helping to move signals back and forth with high fidelity and to keep systems up and running and be productive. But then when you start to move into a switching class of solution, you're talking about much bigger die size, more capabilities, more functionality, you're now the traffic cop that's sitting in between all these very important endpoints, whether you're talking about CPUs, GPUs, networking, memory and storage.

So from a portfolio perspective, we've divided Scorpio into 2 main buckets. We have the P-Series, Scorpio solutions that are for what I would call head node connectivity, being that traffic cop within a compute tray, talking in between all those important end points. And then we have an X-Series solution for Scorpio as well that specifically interfaces just between GPUs or just between XPUs for a scale-up application.

So what we saw over the course of the last 12 months is Scorpio as the fastest and quickest growing product line in the history of the company going from effectively 0 in the Q1 time frame of 2025 and ultimately ended up shipping over $125 million worth of Scorpio P-Series solutions last year.

So a very quick growing product, and that was at primarily one customer on one big platform. So it kind of shows you the potential if you knock down a couple of more. So very positive and optimistic about how P-Series is going to continue to ramp. What we just discussed on our latest earnings call is that we expect ramp P-Series Scorpio solutions at least 2 additional U.S. hyperscaler customers starting probably at the tail end of this year, but more into 2027 across both merchant GPU platforms and internally developed XPU based platforms.

And then the X-Series solution is going to start layering on in the back half of the year for scale-up applications. We have a large customer that's going to leverage X-Series for their rack scale solution for scale-up and that's going to further kind of expand our content portfolio as well. And beyond that customer, we've talked about 10-plus additional engagements for X-Series of which maybe a couple more will start to ship at the very end of this year and kind of layer on top of that lead customer. But we're still very much in the early stages of this game. The market is just developing, we're going from basically a market that was $0 last year to a market that could be as big as $10 billion or more by 2030.

Great. And then, yes, so the second half ramp is for the X-Series, the P-Series is already ramping then. And maybe you could talk a little bit about Taurus and the AC market, the cabling or connectivity has become an important element, I guess, box-to-box or maybe you can help us clarify there's tons to bring optical all the way to the chip and there's some cabling there as well. So maybe you can help kind of clarify that one.

Yes, sure. So similar to what I discussed on the PCI Express side, you have the same issues in the Ethernet where distances are increasing and speeds are increasing as well, so you need active solutions or active electrical cabling solutions that have retimers on each end of the cable that are allowing those signals to be passed up over longer distances at high speed without losing their fidelity.

So we had a great year with Taurus last year. Again, we just started ramping tourist in the back half of 2024, so that was effectively 0 at that point and did over $100 million of Taurus revenue in 2025. So getting up to being 15% plus percentage of total revenue for the company. And that's, again, primarily one customer across a variety of platforms at that customer. But I think the catalyst there is as we transition in 2026 towards 800 gig port switching speeds that similar to PCI Express, speeds go up, and you're going to start needing cables and retimers and places you didn't need them in the prior generation.

So our expectation is that we are going to expand beyond the lead customer in 2026 and ramp additional designs at new customers on 800 gig starting in the back half of which should continue to drive some nice growth for that tourist product line.

Great. And I don't recall you guys being terribly acquisitive, but you did find an asset you liked recently aiXscale you'll pronounce it. And there's an entree into optical. Maybe you can talk about what it was about that acquisition that appealed to you? And maybe kind of how it hints at your road map that kind of incorporates optical in that TAM.

Yes. So I think, again, kind of -- I mean just to kind of set the stage a little bit, going back to IPO road show. I mean we've always talked about looking out over the next 5-plus years and expanding beyond our copper roots and getting into the optical arena. So we started out with Aries in a very kind of narrow set of solutions to Leo memory controllers and bigger solutions, switch fabrics, which are now much bigger, higher protein solutions.

We provide chip only solutions. We provide modules, add-in cars, system-level solutions. And then ultimately, we'll have a copper portfolio and we'll have an optical portfolio as well. So it's not something that we just kicked off in the last -- since the earnings call, we've been planning this for quite some time and putting the pieces in place to be able to participate in this incremental market opportunity.

And as you look at the solution, we're focused on the scale up side of optics because that's where we play today, and that's where we see the biggest TAM. And we are very well positioned with Scorpio X-Series as an anchor socket within these back-end topologies to be able to not only provide electrical and copper-based solutions, but to optically enable Scorpio as well for these multi-rack applications where you'll need both electrical and optical solutions.

So there's a couple of pieces of the puzzle that need to obviously be developed and solved. We've been working on some of those organically already for the last 18 months. But as we got close to customers and started really kind of diving into them, this connector technology, this bridge in between the optical engine and actual fiber media itself is a very nuanced and critical piece of the puzzle and that something that we felt like we were well positioned to kind of organically develop.

So we acqui-hired a small team in Germany called aiXscale, as you pronounced correctly, which a few people do, that brought this connector technology, which is basically a small piece of glass that interfaces between the optical engine and the fiber that will be used within our CPO solutions down the road, it could be used in optical engines that bolt on to XPUs and it can even be sold discretely. And we talked about on the earnings call that we actually have a large customer qualifying just the stand-alone discrete glass coupling piece, which hopefully could drive some volume starting next year as well.

So lots of work to be done here, but comfortable with the position that we have with Scorpio. And we still have some runway here. I mean I think Scott were talking 2028 and beyond before we start to see mass adoption but tightly partnering with customers now to define what needs to take place, what needs to be launched at what time to be able to intercept this market and it's going to be a big TAM matter. This stuff is not cheap.

So everybody keeps saying, what is optical takeover? And this is not a magical moment where optical takes everything because it is expensive. It is less reliable. It is more power hungry. So there are gating factors for mass adoption, but in places where you absolutely need to span a certain distance that can't be achieved with copper, you'll do it with optical, and that will start to phase into the kind of the mix over the course of the next couple of years.

A company like yours can give us insight. As you look at your kind of R&D kind of allocation, how much of it is an optical team and a copper team? And how much of it is really synergistic where we overestimate how separate these are?

It actually is a large piece of leverage to synergistic R&D effort that's being taken advantage on both sides of the aisle. Now of course, there are nuanced pieces of the optical side of the house that will need their own resources and their own team to be able to support and develop that type of stuff. But there's a lot of core competencies that we're going to be able to leverage on the copper side specifically that are going to be able to translate into optical.

So when we think about our -- how we are going to grow the team and we doubled our head count in 2025. We're looking to grow that significantly again in 2026. We've done 2 acqui-hires in the last 2 quarters to kind of accelerate that process. They're going to be slotted into all these different projects and all these different opportunities. Like I said earlier, we have a good track record of execution.

Now our customers are coming to us to help them solve the connectivity problems of tomorrow and the next day after. So we need resources in place to be able to support and not overpromise and deliver against these big mandates. So the team is going to continue to grow. The OpEx is going to continue to grow. We never really break out what portion is working on what, but I would say that there is a lot to leverage across both domains that we will take full advantage of.

I hope that helped clarify a bit for the audience as kind of copper and optical are being talked about a lot. There is more similarity and difference. I'll throw it out to the audience for questions to see if there are any for Nick here. Any questions from the audience?

[indiscernible].

Yes. So that's probably more applicable to like the Aries and the Taurus business lines where we're doing the signal conditioning and reach extension use case. So today, using copper at current speeds probably take -- go up to 6 to 7 meters. Beyond that starts to get tough, and that would be the kind of breaking point where you have to start bending towards optical, roughly speaking.

Thanks. Any other questions from the audience? Maybe Nick while the audience is gathering its thoughts. You can talk about your work with the NVLink and the Fusion program. That seems like a unique sort of technology opportunity into the NVIDIA platform, maybe tells what that is.

Yes. I mean it's exciting for us. I mean, we've had a long-standing relationship with both NVIDIA and AWS, so very happy to be kind of invited to the party to be able to collaborate with these 2 juggernauts in the space. So for those who don't know, NVLink Fusion is effectively a program that NVIDIA has extended to folks that build their own accelerators or their own XPUs to be able to leverage the NVLink backplane scale-up topology that's been ramped at scale has proven and to be the best in the world at this point.

So there's a big opportunity there for guys like us to be able to provide the translation and be the bridge in between these 2 worlds, right? Because the XPUs don't speak NVLink. They speak their own language, whether they're based on PCI Express or Ethernet, but they need to interface with the NVLink scale-up protocol.

So there's a lot of work being done on a 3-way basis between ourselves, AWS and NVIDIA to be able to solve for what that translation looks like. And that will be productized in a solution that sits next to every XPU that's able to take the language coming out of the XPU and then push it to an NV switch on the back end for scale up. So for us, it's a huge kind of incremental opportunity in terms of kind of revenue and a new program and a new socket.

And I think even beyond that, our expectation and hope is that it's not going to stop at just AWS, there'll be more folks looking to leverage NVLink Fusion to scale up their processors. And we would love for NVIDIA to be -- continue to be a matchmaker for us and to link us up and pull us into some incremental programs and we're starting to see some of that engagement today. So more to announce and kind of sign posts to lay out over the next couple of quarters, but it's been pretty encouraging.

Any other questions from the audience? In the back.

[indiscernible]

Yes. So yes, if you go back even just as recent as 12 to 18 months ago, the PCIe retimer business was probably 90% to 95% of total revenue. Exiting 2025, it was kind of down into the 55%, 60% of revenue and that business still grew 70% last year. So you can do the math on how fast everything else was growing to be able to reduce the percentage of exposure on the PCIe retimer side. So we continue to think PCIe retimers grow going forward, but we think Scorpio and some of the other businesses grow faster. So it should continue to come down as a percentage of total revenue.

In terms of the transition from Gen 5 to Gen 6, still pretty early in the game today, really the only PCIe Gen 6 capable GPU in the markets Blackwell, being supplied by NVIDIA, and that's driving 100% of our PCI Express Gen 6 business. But undoubtedly, there will be incremental GPU suppliers, CPUs, ultimately, all the peripherals that will kind of continue to get dragged on to the next generation. And I think crossover will probably not happen until 2027.

But good thing for us is that, again, on a like-for-like generation-over-generation basis, we see about a 20% uplift on the ASP side. So even if units were to stay constant, which I don't think they will, I think they'll grow, we will see an uplift in the revenue profile just because of the ASP increase.

Well, it was nice to have a business growing 70% decrease in the mix. That's -- it's a nice problem to have. So with that, we'll thank Nick for his time and everybody, and thanks for coming, Nick.

Thanks, I appreciate it.

All right. Welcome back, everybody. I'm Joe Moore from Morgan Stanley semiconductor research. And very happy today to have the management team of Astera Labs, Jitendra Mohan, CEO; and Mike Tate, CFO. Thank you guys for being here.

Maybe we could just start out with just a general view of AI. I think the market -- I think the investors realize that 2026 is very strong, but there's a lot of questioning about duration of that strength into '27 and beyond. It seems like everyone in the industry is kind of worried about supply for the next 4 years. Everybody in Wall Street is worried about next year fading. Can you just talk about generally what you see from the AI ecosystem?

Just like we were talking about backstage, it seems like sky is always falling for one reason or another. But I think the worries are a bit exaggerated. I really think, just again, from a layman consumer standpoint, we are still in very early innings of AI. We use AI internally in the company for some of the work that we do and the potential for using AI is just immense.

So I think -- and you can also see this from the CapEx investment decisions that some of the bigger hyperscalers are making. And specifically, as it pertains to a Astera Labs, there is still a lot of growth that we have ahead of us, whichever direction the market is going. So I do think that if you look back 5 years from now, 10 years from now, we'll be in a much bigger installations of AI and so on.

Will that ride be all linear? I don't know. But I think the goal that we have as a company is to always do better than what the market is doing. If the tide is lifting, we want to rise higher, faster. And if it's going in the other direction, we still want to beat the market.

Great. That's helpful. So maybe we could walk through some of the products and what you're seeing there. The company came public on the back of 3 different product portfolios and you've added more since then. So maybe start to start chronologically. But Aries, the retimer business grew 70% last year. Can you help us think about the forward content opportunities that's kind of signal integrity seems to be a critical issue.

Yes, absolutely. And this is what we have said, as these systems become more complex, the data rates become higher, it's just difficult for the signal to go from one point to the other. So the amount of single conditioning that you need will continue to increase. So the retimers are not going to go away so long as we have higher data rates, more complex system and copper from a media choice. And that's what we see happened last year. Our single conditioning portfolio, Aries, in particular, grew 70% year-on-year. We expect it to continue to grow this year as well and then out in the future.

So that growth is really coming on the strength of also the deployment of this portfolio, not only in scale-out, which is when we went public 2 years ago, majority of the revenue was scaled out. Now we also participate in scale-up with these PCIe retimers, both on a chip-down stand opportunities where you're driving a backplane or a front foot as well as active electrical cables. So it's been a great ride so far, and we continue to see this portfolio growth.

And can you talk about the competitive dynamics? You were first mover in PCIe Gen 5 retimers. There's been a number of competitive announcements since then, but it seems like you've still done a very good job holding market share. Can you talk about that? And then also your progress on Gen 6?

Yes. Actually, interestingly, we were the fourth company to have PCIe retimers when we first launched them. But we won because of the architecture that we chose, which is based on a software-first approach. And then we have our COSMOS software that we use to optimize, customize these solutions as well as get diagnostics information that our customers then use in their own operating stack. On the back of that, we actually became the dominant supplier for PCI Express Gen 4 and Gen 5. And then we started the PCIe Gen 6 journey in March of '24, actually right coincident with us, our IPO debut, where we first introduced Gen 6 retimer. But at the time, we said, look, we are the first one out of the gate, but remains to be seen what happens.

Since that time, actually, we have done quite well. We are probably the only company that is shipping PCIe Gen 6 retimer solutions in volume. Over these 2 years, we have learned a lot, and we have incorporated those learnings into our COSMOS software that makes our solutions stronger. But I would also say that more and more, it's not about having a point solution. It's great that you have a retimer, and I'm sure over time, it's a big market. Other companies will be successful with their retiming solutions. But we have retimers. We have electrical cables, we have gearboxes and we have switches. So we have a full PCI Express portfolio, and this is what customers look for when they are trying to pick a vendor for their PCIe solution needs.

Great. And then last quarter, your fastest-growing portfolio was Taurus, which is the active electrical cables, which is not a market where you were the leader coming in, but you've done very well. Can you talk about that opportunity?

Yes, we're very excited about the progress we're making with Taurus. We started shipping Taurus in the second half of 2024 for 200-gig switches. And then just recently, we started ramping the 400-gig switches. And like you mentioned, it was our #1 grower in Q4. These 400-gig switches are still building momentum, so they will provide growth in 2026. And then we're starting to now bring our 800-gig switch solutions into the market. And this is -- for the first time, we're going to diversify beyond our largest customer, so that will incrementally give us more growth going into 2027.

Okay. Helpful. And then the third market that you came public was Leo, the CXL business, a little bit more nascent, I think, but can you talk about that opportunity as well?

If I planned better, I would have punted this question to Mike. So Leo is our third product family. Clearly, it has not met the expectations that we had for growth for a variety of reasons, both the availability of CXL capable CPUs, as well as all of the investment dollars going into AI.

Having said that, we are actually -- it's looking quite good for the family in general, with our lead customer qualifying this for data center applications. So a CXL-based server instance is now available in private beta on Azure Cloud. And as this use case diversifies, we expect other hyperscalers to follow suit as well. So general purpose compute will be the first deployment of CXL, and it will happen later this year and continue to ramp into 2027.

We are also excited about the applications of CXL and AI. So there are some of the smaller AI platforms that are both exploring and actually, we have some design-ins for KV Cache-type applications for CXL. So those are still being explored, are not meaningful enough to comment in terms of a 2026 revenue ramp. But as these become more commonplace, we'll give more signposts for '27 and beyond.

Okay. I mean that's kind of new, right? It's been the general purpose technology primarily, yes. Okay. And then you've talked about the Switch family, which you introduced after the IPO, Scorpio being the biggest market that you address. And I know there's different P-Series, X-Series, we can talk about both. Can you talk generally about why you're so enthusiastic about switching?

Yes. So the switch product happens to be an anchor socket when it comes to designing an AI system. So clearly, if you look at our hyperscaler customers, they will start with the XPU. And with the XPU, they have two choices. For scale-out, it is almost always PCI Express base. So this is an application that is service our Scorpio P-Series family. And then they have a choice to make for their scale-up as well and scale up when they use PCI Express-based protocols, that is service by our Scorpio X family. So both of these are very large TAMs. We anticipate Scorpio P to be about $4 billion of TAM over time. And we have great traction. It is -- Scorpio P-Series is the fastest-growing product line in the history of the company. We introduced this in Q2 or started generating revenue in Q2 of last year. And just in 3 quarters, it's already crossed 10% of our revenue. So we're very excited about that.

We also announced two additional hyperscalers, U.S.-based hyperscalers that have designed this in and is going through qualification with revenue ramps happening next year. So that continues to be a very exciting application. We're also actually even more excited about the Scorpio X family, which is applicable for scale-up, and scale-up just happens to be a lot richer application with a lot of links and so on. And we anticipate that TAM to be $10 billion with a combination of PCI Express and UALink.

We already have a design win in there. So we're very excited about that. It started to ramp in 2026, will continue through the back half of the year. And we have over 10 engagements for Scorpio X family that we announced a couple of quarters back, and now it's grown even more than that. So we fully anticipate having multiple customers designed in for Scorpio X family by the end of the year.

Okay. Great. For the scale-out for Scorpio P, how pervasive will that be? Is that multiple platforms? Should we think of it as NVIDIA-centric, ASIC-centric? How do we look at that?

Yes. So Scorpio P-Series, the P stands for PCI Express. So PCI Express is very widely applicable. In fact, PCI Express makes the nervous system of all of these AI servers, whether it's compute or network cards or storage they all connect to each other over PCI Express. So Scorpio P-Series has wide applicability. And as we mentioned, we started with a Grace Blackwell-based deployment when customers wanted a customized deployment of Grace Blackwell, they chose to go with Scorpio P-Series. A similar application is we are seeing with other hyperscaler customers as well.

And the reason for that is very simple. When we designed the Scorpio P-Series, it was a design that was in grounds-up for AI applications. It was optimized to deliver the performance that is needed for AI applications to understand the AI workloads and the AI traffic. So it's a very purpose-built device and the customers like it for that reason.

Okay. And then with scale up, really big opportunity completely greenfield. And I feel like we've sort of -- we went like 2 weeks from talking about that to talking about how competitive it's going to be and who's going to win what. It seems like a lot of opportunity. I think most people know who your lead customer is there, a pretty exciting one. But you've also talked about these 10 additional engagements. How big is that opportunity for scale up for you over time?

I think over time, Scorpio X will probably be our largest product family just because it is the fastest-growing product family, and it is the largest TAM. We have 10-plus customers. These are made up of hyperscalers, AI platform providers, both in the U.S. as well as overseas. So there is just insane amount of traction that we are seeing a lot of customers actively designing in, some have designed in and are ramping as we just talked about. Others will continue this ramp later part of this year. And then we are also excited about the duration of this ramp. So some of these designs just go into production in 2027 and will continue to ramp through 2028. So a very long legs before the industry and some of these customers transition from PCI Express to UALink.

Maybe you could talk about that transition. And just there are a lot of standards here where we were talking about NVLink, NVLink Fusion, PCI Express, UALink, Ethernet, ESUN, there's a lot of different directions this can go. Can you talk about your ability to attack those opportunities?

Yes. Just maybe just to kind of lay out the land. There are, I would say, three different protocols that customers are using. First camp would be a proprietary camp, which is things like NVIDIA's NVLink, for example. Second one would be PCI Express and UALink and I'll tell you why they are in the same camp. And third would be Ethernet and all the different variants of Ethernet.

Largely, what we have found is customers are in one of two camps, either they are using what is called a memory-centric protocol, NVLink, PCI Express, UALink are examples of that, or they're using an Ethernet-centric protocol. And what we have learned now is that customers are going to stay in the swim lanes that they are in, not because of anything other than the fact that their software stacks are completely tuned for a particular protocol.

So if you look at what our plays is in each one of these, NVLink Fusion until recently was proprietary, we really did not have any play in it, nobody did. But with the advent of NVLink Fusion which was announced May of last year, and we recently announced a participation in that with one of the lead hyperscalers. Now we do have some presence within NVLink Fusion.

PCI Express and then as the transition to UALink, is really our bread and butter. UALink, in particular, was purpose-built for scale up. It is just designed for optimizing for traffic that's required to connect GPUs together. A lot of interest in UALink. And certainly, we will participate in that just as fully as we are participating in PCI Express with a complete portfolio of signal conditioning devices, switches and so on.

And then Ethernet, the customers that have optimized their software for Ethernet, they are going to stay with Ethernet. They are not moving either. And our participation there would be with our scale-out devices, both on the Scorpio P-Series as well as the Taurus retimers.

Okay. That's helpful. And then why would you characterize UALink similarly to PCI Express as memory-centric versus -- yes, I thought there was more of an Ethernet kind of basis for some of the UALink.

Yes. So there are two things. One is the protocol and the other one is the signaling layer. So now what has happened is all the signaling layer has become the same, whether it's NVLink or Ethernet or UALink, they all use 200-gigabit SerDes and 200-gigabit physical layer from Ethernet. So from that standpoint, they're all the same.

The difference really comes from the protocol. UALink was built ground up for scale-up protocols. So understand that there is GPU sending data to another GPU, trying to do all reduce operations or other AI-specific operations. Whereas Ethernet was designed as a message passing protocol. So you need to run some additional things on top of base Ethernet, which adds to more latency or reduced throughput.

The other very important factor between UALink, for example, and Ethernet is UALink is a completely open standard. It is a greenfield deployment, Astera wants to participate in it. Other start-ups will participate. Other public companies will participate. And as a result, you end up with a nice vibrant ecosystem for our customers to choose from. And that is something that is very important for our customers.

On the other end, proprietary standards are indeed proprietary. Even if you look at Ethernet, which is a nice public standard once you start adding these other protocols on top of it. And because of the dominance of one player in that industry, it starts to again look a little bit not quite an even playing field.

So our customers are telling us to focus on PCI Express and UALink. So that's what we are doing. At some point in time, when we have that opportunity addressed and our customers ask us to do other protocols, we are always open for that.

And I guess, can you talk about the UALink opportunity? There's been AMD is using UALink, but it's sort of funneled through Ethernet. You've had Amazon talk about UALink and Trainium4, but also NVLink Fusion in Trainium4, so there's some confusion there. It seems like there's a pretty good future for UALink, but can you talk about where we're going with that standard?

Yes, yes, absolutely. I think UALink has a bright future. When we look at the overall scale-up TAM, we estimated that to be about $20 billion in 2030. And then again, our estimate is that it's split pretty evenly between UALink and Ethernet. And so you don't have to convince everybody to go UALink. As a matter of fact, we do think that all of these different protocols will coexist. And we have a strong enough opportunity. Almost half of that's a $10 billion TAM to address just for UALink alone. And so we are absolutely focused on that. And then once we have solved that problem, then we look at the other $10 billion that's open.

And then your involvement in NVLink Fusion is pretty interesting. NVLink within the NVIDIA platform is their technology, what is the role that you can play in NVLink Fusion?

Yes, that's -- it's a great development where previously, as I mentioned, NVLink was a closed ecosystem to NVIDIA. NVIDIA build NVSwitches, and their GPU stock to NVSwitches directly. There is an interest from the industry to use NVL72 infrastructure, not just the switches, but the whole rack with be liquid cooling and all the nuts and bolts and power components that go along with it. The problem is none of the other external to NVIDIA XPUs stop NVLink. So NVLink Fusion allows external XPUs, ASICs to talk to an NVIDIA ecosystem. And the way this is done is to use a chip that translates the native protocol of the ASIC or of the XPU and translates that into NVLink, so that you can then talk to NVSwitch.

I think it's very important to note that NVIDIA trusted Astera and as did the hyperscalers, to sort of open up the innards of both sides of the protocol to build this chip. And this chip gets attached, one is to one with every XPU. So as a result, if you look at the content that Astera has on a, call it a hybrid rack, where you have an XPU talking to NVSwitches or a native rack, where you have an XPU talking directly to a native switch, is about the same for us. So we are just very happy to participate in both NVLink Fusion ecosystem as well as PCI Express or UALink.

Yes. Okay. Great. Thank you. So the [indiscernible] version of all this is a $20 billion TAM that you see emerging in 5 years, completely greenfield, you think at least half of it is accessible to you?

Half of it is accessible to us on the products that we are building today. We have the capability to address the other half as well, and we will do so as our customers ask us to do.

Okay. Great. Maybe we could shift a little bit and talk about optical. Most of your technology is based on copper, scale up today is entirely based on copper, but a lot of focus on the move to optical. Can you talk about that and your role in optical going forward?

Yes, absolutely. So I think you have to, first of all, look at the portfolio, not as copper versus optical. There is a core functionality that we need to deliver as part of the controllers or as part of the fabric switches, which might be memory control technology or it might be the ability to switch packets and so on. The choice of copper versus optical is a media choice. And so the products that we have today that are available for copper, over time, we will offer them in optical flavors as well.

So for example, if you look at our Scorpio X family, today, it supports only copper as a media or the interconnect choice. And if you are able to reach to your endpoint, which is an XPU directly on copper or through a retimer or cables, et cetera, that's what you will do. Our customers absolutely want to use copper for as long as possible and as much as possible, simply because it is the most reliable, it is the cheapest and it's the lowest power technology.

However, as these clusters scale and data rates go up and you need to connect multiple racks, you do need to go to optical. And we intend to fully participate in this optical opportunity that is in front of us. And we've been working on optical for optical technology for a long while actually. We just announced it gained more momentum in this last earnings because of the acquisition that we have done of an optical company. And our play in the optical space will start with an optically-enabled Scorpio switch. So just like we have the switching technology coming up with copper, it will be available in optics. The way to do that is to design an optical engine and marry it with the switch. That's where we will start.

But that is not the full solution, right? So now the switch can talk optics, but you need something on the other side as well. So this is where we also will have our optical engine available to attach to the XPU. So that becomes a second kind of product in the road map. And last but not the least, it turns out the technology that we got through the acquisition of Xscale, a glass connector, is actually a really key piece of the overall optical engine puzzle so much so that our customers are quite interested in just buying that. And so we will offer that also as a component sale.

Okay. And you've talked about this being a much larger potential opportunity when it migrates to optical, we've heard some really big numbers in terms of optical content of those scale-up systems down the road. Can you talk a little bit about that?

Yes. So I think there are two things. One is very easy to see, which is just that typically, if you are making a copper link, there may be no retimers in that link at all, if you can avoid it. And even if you do have to add one, you might add one retimer in the midst. So when you go to optical, you have to have a minimum of two, because one converts electrical to optical, near the switch, let's say, and another one converts the optical back to electrical, near the XPU. So already you have double the attach. And the ASP of each one of these is an order of magnitude higher than what you would get, let's say, for a retimer class device.

So we look at it as a huge opportunity. If magically all of our retimers were to convert to optical, that will be a wonderful opportunity. But at the same time, this is also the challenge for our customers. Somebody has to pay for all of this ASP increase and the attach rate increase. And that is the reason that our customers want to stay with copper for as much as possible.

Makes sense. You did an acquisition, a smaller company, aiXscale Photonics, what is the state of your optical be? Do you need to invest a lot to have these capabilities? Do you need more acquisitions? How do we think about that?

So maybe if I were to just sell lay out the optical engine for the audience here. Optical engine consists of three pieces. The first one is an electrical IC, and we started investing in this foundational technology that will lead to electrical ICs way back. So it's not a decision that we took yesterday or the day before. That's on one side. So this electrical IC is what talks to the switch or what talks to the XPU. This electrical IC then drives silicon photonics. I'll come back to silicon photonics in a second.

Silicon photonics is also a wafer-level technology. At the other side of the silicon photonics wafer, which converts electrical into optical is a glass connector. This glass connector is used to gather up all the photons and then launch them into a fiber. So this is the technology that we got through the Xscale acquisition. And we focused on this because we really believe that the bottleneck to deploying optical at scale for scale-up is really this connector piece, because everything else is a wafer-level technology. TSMC and others have been producing these wafers by the bushel. Same is true for silicon photonics as well. But putting it all together and doing optical alignment and glue and so on, that is really, really tricky to ensure reliability of these links. And that's also a technology that would be hard honestly, for us to develop internally, which is why we made the acquisition. And just with complete hindsight, it was a great acquisition.

Coming back to the silicon photonics. So this is the third piece, and it's in many ways, very religious choice, from our customers. Some customers absolutely want to go down a particular path because that's what they believe in. Other customers are more open. They might want to have a backup solution or a primary solution and so on. So our strategy is not to get ourselves pigeonholed into one technology, but to leave this optionality open for as long as possible, both to service multiple customers and also if a customer were to do an about turn that we'll be able to address that.

So we will offer Astera Labs silicon photonics, a choice of modulators and fabs and so on. But if a customer insists that, hey, you must use this other silicon photonic solutions, then we are certainly open to that as well. The philosophy at Astera has always been to support the customer with whichever choice that they have made as opposed to forcing our choice on our customers. And so far, we are getting very good reception from our customers. We are engaged with the requirements that they have. We understand the time lines that they want us to intercept and we are more than capable of intercepting those time lines.

Okay. Great. So maybe shift to talk about the financials a little bit. You've had really good growth even before the Scorpio X really starts to ramp in the second half. You guided to a good March, and it seems like you've got strong growth. Can you talk about what's kind of bridging you into that ramp in the second half?

Yes. We're seeing growth from all our product lines, Aries, Taurus and Scorpio particularly right now. For Scorpio, we launched P first, so the P-Series grew in 3 short quarters to be 15% of our revenues. And as we make it through 2026, we're going to add two new U.S. hyperscalers to that mix. So we have a very good line of sight of growth into 2027. Scorpio X is a much bigger segment of the switching market. We just started shipping initial volumes in Q4 and go to the higher volumes in Q1 and Q2. But the way it's working out, it will be a much more material ramp in Q3 and Q4 as well. Ultimately, Scorpio X, given the size of that opportunity, it will cross over P and Scorpio in total will be our biggest product line in short order.

Okay. That's helpful. Gross margin, you've talked about some mix-related pressure there. You also have some issues around the warrant. Can you maybe talk about.

Yes. As we broaden our product portfolio, we're going to see a wider range of gross margins. When we went public, 90% of our revenue was coming from Aries as a discrete semiconductor device. And now on both Aries and Taurus, we also -- we sell modules. So we do hardware. So we take our retimer device and put it on a PCB board with passives and other hardware componentry. So it's hard to mark up the ASPs as high and get the margins to the level of the discrete semiconductor devices. So those are ramping as a percentage of our revenues.

Also, with the Scorpio family, we will also see a wider range of margins depending on the use case on the number of lanes the device uses. So there will be applications where the device will have more lanes and the customers using in the device. So those will be a bit lower than corporate average.

And then lastly, the Amazon warrant that we have, and we've had one in the past, you take the value of the warrants that vest and you hit it against revenue, which is dollar-for-dollar impact to gross margins, which we're estimating to be about 200 basis point hit starting in Q2.

Okay. Helpful. And then last quarter, you had significant revenue upside, but also there was a big step-up in R&D associated with that. How should we think about that? Obviously, there's a lot of things that you can invest in, if you're able to continue to put upside in revenue, will we see the R&D line come up along with that? Or just how do you think about that?

Yes. I'd say, right now, we're not optimizing to the operating margin. We're optimizing to seize the moment with the TAM that's in front of us. So I would say we've underinvested last year, and we've had an opportunity with these last two acquisitions that we did, one was a technology acquisition, and one was more of an acquihire. If we find those opportunities, we will take advantage of those. There's just so much opportunities in front of us.

But that being said, even though we will continue to invest significantly in R&D, the revenue opportunities are very sizable. And we do expect to continue to have operating margin leverage from this new level that we're at right now. And our model is to deliver 40% operating margins, and we think we'll track back to that depending on revenue growth versus the level of R&D investment we're making.

Yes. Okay. We certainly developed a very strong track record as a public company when it comes to numbers. So congratulations on that.

Customer concentration, it sort of comes with the territory when you are in these markets, selling to hyperscalers. How do you think about that? And how -- to the extent that your biggest customer is getting pretty big, how do you think about diversification opportunities away from that?

Yes. I mean we're focused on the data center. So it's a natural occupational hazard to have customer concentration. Our lead customer is the most aggressive in adopting our technologies first. So they were the first to adopt the Taurus and the first to adopt Scorpio. That being said, we're seeing a nice diversification path for Taurus and Scorpio beyond our lead customer. We talked about 800-gig Taurus. Now we'll broaden it out beyond our lead customer. And then with Scorpio, we have two new U.S. hyperscalers, and we have over 10 scale-up customers. So we see a path to see very good diversification. But it doesn't mean our lead customer is not going to grow. They're going to grow significantly. So...

Yes, it's hard to outgrow them in too fast. It's an important aspect of your business is the relationship with hyperscalers and the trust that they have to develop products for them and things like that, something that was notable even before the IPO that you had those relationships. Where does that come from? Where does that faith come from? And does that open up a lot of opportunities for you guys?

Yes. I would say that's actually one of the greatest strengths that we have as a company. And this started way back, I think just when we were getting started. We made a prediction of the type of components that the industry and our customers will use. And I don't know how much they trusted us at that time, but they say, yes, go ahead, build what you want to build. We built it, and it turned out that they really needed it, and we had the perfect solution when they needed it.

And so that was really a start of a great relationship where we have listened to our customers, we've understood what it is they want to do, and then we have consistently delivered. So this combination of moving along at the same pace at which our customers want us to move and be very consistent in our delivery track record has earned a trusted adviser position with our customers.

And so now what happens is we get involved with them 2 or 3 years before they actually need the product because that's how far they plan their deployments in advance. And a lot of things have to come together. And we get a seat at the table to understand that. And especially now with the Scorpio class devices that, by the way, we engaged like more than 2 years ago with our both the lead customer as well as the AI platform provider. We get just such a good view into what their road map is and what products we need to build. And extremely importantly, what we don't need to build and what features we don't need to put into our own products.

So that is a fantastic advantage that we have. And this is really what has led us to add the number of product families that we have added to the company. We now have four. And certainly, we are working very hard to introduce more product family into the industry, all based on customer feedback.

That's very helpful. Let me pause there and see if we have questions from the audience. One up here. Will you wait for the mic.

I have two questions. The first one is for the CXL solution. So you said that there is some market traction for this product. I mean like for the customer using it, it's more used for the standard product or they need to do the customization for each customer. So this is the first question for the CXL.

The second question is, you talked about right now, Astera want to be a full stack solution for the optical engines. So I mean for this product, it will be only used for the pluggable optics or it can also be used for the CPO, which are major trend right now?

Very good. So let me address the first question first, which is how do we customize our solutions for our customers and specifically for the CXL product family that we have. So there's nothing special about the CXL product family in the sense that it is a standard product that can be used by multiple customers. But just like we do with our other products, there are a lot of software hooks inside of this product. So if a customer wants to use a particular type of memory or a particular combination of the DRAM itself and the power regulators and whatever else, we can do that.

More importantly, if our hyperscaler customers want a certain amount of qualification criteria or serviceability, reliability features when it comes to memory, we can customize it for them. And so our strategy has always been not to spin silicon based on different customer needs, but to customize it based on the COSMOS software that we have. And once we have done this customization for the customer and they incorporate it into their own software stack, it becomes a very sticky solution. So that's your first question.

The second question was the optical engine that we are talking about, is it for pluggable or for CPO? Actually, we are targeting the scale-up optics first. And scale-up optics will rely on either an NPO, near-package optics or CPO, co-package optics, and that's what we are targeting first. Pluggable optics has been a domain of scale out. And while it's a big market, there already are a couple of large players in that space. So we are not targeting it just yet. But at the right time, we will go after that opportunity as well.

Question up here.

As you move customers through 800 gigabits and towards 1.6 terabits, at what point do the -- does the architecture for this copper become the constraint? And when you hit that wall, is the answer going to be more conditioning? Or do you see a genuine pull from your customers towards optical?

Yes, there is not a kind of a proverbial wall that one day, everything is copper and the next day it all becomes optical. Optical is already widely deployed in the data center. All of the long links that are today running 100-gigabit per second and transitioning to 200 are already optical. But it is largely deployed in scale out.

When it comes to scale up, the density of these connections is so large that are traditional pluggable optics does not work. And so our customers want to stay with copper for as long as possible. Specifically to your point, at 200-gigabit per second, we believe that all the connections within the rack level will continue to be copper. There is no need to go to optical. Copper is the most reliable, cheapest and the least power solution that you can have.

Now once you start to increase the domain or cluster size from one rack to multiple racks, what might happen is within the rack, continue to stay copper, but the rack-to-rack interconnect, which can span several meters may have to go to optical. So we believe that's where the first intercept will happen for optical. And eventually, when the data rates are even higher or you've solved all the cost and reliability issues with optical then you can have a pure optical deployment, which is optical both within the rack as well as across the rack. But I think that day is still far.

Have those decisions have been made when you talk about optical scale-up still several years away. People -- there's a lot of start-ups saying, we have this business, we have that business. It seems like all of that is still kind of speculative at this point.

I would not say it's speculative, but the hard decisions have not been made. Simply because nobody has demonstrated deploying CPO at scale. And showing 100 or 1,000 or 10,000 units is one thing, but having these things deployed in the millions such that they do not fail is a very difficult problem for the industry to solve. So we must all together solve the reliability problem, and then we must also solve the cost problem. At a minimum, these two must be solved, and we are working on our solutions. And as we discussed, we want to have a flexible approach so that if course corrections are needed, we are able to make them.

Great. Well, that brings us to the end of our time. Jitendra, Mike, thanks so much for the time.

Thank you.

Thank you.

Good afternoon. My name is Karly, and I will be your conference operator today. At this time, I would like to welcome everyone to the Astera Labs Q4 '25 Earnings Conference Call. [Operator Instructions] I will now turn the call over to Leslie Green, Investor Relations for Astera Lab. Leslie, you may begin.

Thank you, Karli. Good afternoon, everyone, and welcome to the Astera Labs Fourth Quarter 2025 Earnings Conference Call. Joining us on the call today are Jitendra Mohan, Chief Executive Officer and Co-Founder; Sanjay Gajendra, President and Chief Operating Officer and Co-Founder; and Mike Tate, Chief Financial Officer.

Before we get started, I would like to remind everyone that certain comments made in this call today may include forward-looking statements regarding, among other things, expected future financial results, strategies and plans, future operations and the markets in which we operate. These forward-looking statements reflect management's current beliefs, expectations and assumptions about future events, which are inherently subject to risks and uncertainties that are discussed in detail in today's earnings release and in the periodic reports and filings we filed from time to time with the SEC, including the risks set forth in our most recent annual report on Form 10-K.

It is not possible for the company's management to predict all risks and uncertainties that could have an impact on these forward-looking statements, or the extent to which any factor or combination of factors may cause actual results to differ materially from those contained in any forward-looking statements.

In light of these risks, uncertainties and assumptions, the results, events or circumstances reflected in the forward-looking statements discussed during this call may not occur, and actual results could differ materially from those anticipated or implied. All of our statements are made based on information available to management as of today, and the company undertakes no obligation to update such statements after the date of this call, except as required by law.

Also during this call, we will refer to certain non-GAAP financial measures, which we consider to be an important measure of the company's performance. These non-GAAP financial measures are provided in addition to and not as a substitute for financial results prepared in accordance with U.S. GAAP. A discussion of why we use non-GAAP financial measures and reconciliations between our GAAP and non-GAAP financial measures is available in the earnings release we issued today, which can be accessed through the Investor Relations portion of our website. With that, I would like to turn the call over to Jitendra Mohan, CEO of AStera Labs Jitendra?

Thank you, Leslie. Good afternoon, everyone, and thanks for joining our fourth quarter conference call for fiscal year 2025. Today, I'll provide an overview of our Q4 and full year 2025 results followed by a discussion around the current trends within the AI infrastructure market. I will then turn the call over to Sanjay to walk through Astera Lab's near- and long-term growth profile.

Finally, Mike will give an overview of our Q4 2025 financial results and provide details regarding our financial guidance for Q1 2026. Astera Labs delivered strong results in Q4 with revenue at $270.6 million, up 17% from the prior quarter and up 92% versus Q4 of last year. For full year 2025, revenue was $852.5 million, up 115% versus the prior year. Growth within the quarter and for the year was broad-based, spanning across our signal conditioning, smart cable module and switch fabric product portfolios as we continue to diversify our business profile with several new design wins across multiple customers.

Secular trends remain robust within the AI and cloud infrastructure space, supported by exceptionally strong spending commentary coming from the top U.S. hyperscalers. with Google and AWS alone, guiding nearly $400 billion in total CapEx spending for 2026. We are benefiting from this increased spending both in the near term and long term. Furthermore, the market opportunity for our intelligent connectivity platform is substantially larger than we initially anticipated, encompassing multiple product lines, physical media types, form factors and protocols for both standard and custom applications.

Starting with Scorpio, our PCD family continued its volume ramp at our lead customer with growth coming from both existing and incremental platform designs. For the full year, Scorpio PCDs exceeded our target of 10% of revenue and remains the only PCIe 6 fabric shipping in volume in the market. Looking into 2026, we anticipate continued growth for Scorpio P-Series at our lead customers as well as commencing shipments into at least 2 additional major hyperscalers on their next-generation AI platforms.

Moving to Scorpio X-Series. We expect to incrementally grow revenue in the first half of 2026, followed by a transition to high-volume production in the second half. We continue to make excellent progress with additional engagements looking to leverage PCIe for scale-up networking. As previously communicated, we are engaged with 10-plus customers for Scorpio X family. And our current expectation is that we will ship initial quantities of Scorpio X series to support new customer platforms in the second half of 2026 with volume ramp set for 2027.

Solid traction continues to develop with respect to UALink with a vibrant ecosystem, including product announcements, broad IP availability and compliant methodologies being finalized. Recent public roadmap announcements from AWS and AMD, along with other ongoing engagements indicate a broad adoption. UALink remains the highest performance, lowest latency, fully open solution for AI scale of connectivity, and we will be ready to intersect the initial customer platform ramps in 2027. Our Aries portfolio continues to perform well. with [ PCI6 ] solutions contributing robust growth during the quarter, and the overall portfolio growing nearly 70% year-over-year in 2025.

The demand for Aries is driven by increasing deployments of custom AI accelerators at large hyperscalers. Our Aries Gen 6 products are the industry's only PCIe 6 DSP retimer solutions shipping to customers in high volume today, and we are well positioned to maintain our leadership role in the market. We remain very early in the PCIe 6 transition cycle and anticipate additional customers will launch PCIe 6 6 capable AI accelerators and systems throughout 2026 and into 2027. As a result, we expect our Aries product line to continue growing in 2026 and beyond.

Taurus was our strongest performing product family during Q4 as new programs began shipping in volume to support designs across both AI and general purpose systems. In 2025, we saw Taurus revenue grew by more than 4x year-over-year, driven by a breadth of 400 gig designs that will serve as a baseline for continued growth in 2026. We look for the transition to 800 gig switching platforms to be the next catalyst for market expansion, driving further growth opportunities for Taurus.

Finally, we made good progress with our Leo CXL memory expansion products in 2025 and look to build upon that in 2026. We are excited to announce our partnership with Microsoft, Intel and SAP to enable customers to evaluate CXL memory expansion capabilities for their specific workloads within Microsoft Azure M series CRs machines. This program represents the industry's first publicly announced deployment of CXL attached memory, and we expect initial production volumes to commence in the second half of 2026.

Overall, we are proud of the progress we have made in 2025. We added new product lines to service more sockets and address customer applications, increased dollar content per accelerator, diversified our customer base with new design-ins and scale our operations. This progress and a strong track record of technology and operational execution has helped us forge tight relationships with key AI and cloud infrastructure providers. As an example, an update on our relationship with Amazon can be found in our 8-K filed today.

Looking at the combination of growing AI infrastructure deployments and the increasing complexity of high-speed interconnect architectures is poised to drive significant growth for the AI connectivity space. We estimate our served addressable market opportunity will expand by more than 10x over the next 5 years to reach $25 billion. This market opportunity spans our existing and announced copper-based product families, including Aries and Taurus signal conditioning solutions, Scorpio AI fabric switches, Leo CXL Memory Controllers and our recently announced custom solutions for scale up connectivity.

While these numbers and opportunities are substantial and exciting, there is a significant amount of work that needs to be done. Astera labs is deeply committed to building an A+ team with an execution mindset and capabilities to essential to support our customers' technology roadmap and maximize our share of the large market opportunity ahead of us. Therefore, we are strategically investing in the expansion of our team and capabilities to execute against a broadening set of revenue opportunities generated by our customers.

We took an exciting step in this direction this week with the announcement of a significant expansion in our global engineering operations through the establishment of an advanced design center in Israel. Our investment in this talented ASIC engineering team substantially increases our resource pool and will help accelerate development of cutting-edge, high-bandwidth and custom AI fabrics and emerging AI inference technology. I would like to take a moment to thank our global team of Astera, our partners and our vendors who work tirelessly in 2025 to deliver world-class AI connectivity solutions to the market.

Their steadfast focus and effort have placed Astera Labs in a position of strength heading into 2026 as we continue solving next-generation AI connectivity challenges. Finally, we announced today that Mike Tate will transition from the CFO role into a full-time role as a strategic adviser reporting to me. Mike has been instrumental in Astera Labs growth and development since inception, and we are very grateful for his many contributions. We're also very excited to announce Desmond Lynch will join Astera as our new CFO effective March 2. Desmond brings great semiconductor financial experience to the company, and we look forward to drawing on his expertise as we enter our next phase of growth.

We're also thankful that Mike will continue to work full time in his new role to support the company and ensure a smooth transition to Desmond. With that, let me turn the call over to our President and CEO, Sanjay Gajendra to outline our vision for growth over the next several years.

Thanks, Jatin, and good afternoon, everyone. Today, I want to provide an update on our recent execution, followed by an overview of the meaningful market opportunities that will fuel our growth over the next several years. Astera Labs mission is to deliver a purpose-built intelligent connectivity platform with a portfolio of solutions, including silicon hardware and software for rack-scale AI deployments.

Over the past several years, we have been building our portfolio, expanding our capabilities with foundational IP, growing our talent pool and demonstrating the technical and operational execution, which has helped us to establish multi-generation partnerships with leading AI platform and cloud service providers. Looking ahead, we plan to deliver technology enhancements to our core portfolio of AI fabric, signal conditioners and memory controllers, while also expanding our breadth of capabilities to new categories, including custom connectivity solutions, products to address memory bottlenecks in inference applications, optical engines and other optical solutions for scale up and scale-out networks.

Let me now provide an update on our future product strategy. Starting with AI fabrics. We have been thrilled with the initial traction of Scorpio P-Series and X-Series over the last 18 months. Looking into 2026, we are poised to see diversification with our P-Series solutions for head node connectivity at new hyperscaler customers in addition to the volume ramp of our X-Series solutions for scale-up networking.

Exiting 2026, we expect Astera Labs to continue being the market leading provider of PCIe 6 switching solutions and will become a leading provider of merchant scale-up AI fabrics. Our early engagements have been crucial to our understanding of the nuances of deploying complex AI fabrics at scale while also identifying new and innovative approaches to expand our roadmap. Furthermore, hyperscalers are demanding flexible connectivity solutions optimized for their unique architectural approaches and application needs i.e., one size does not fit all.

These requirements lend themself to our software-defined architecture for silicon products, which will now extend to wider radix configuration multi-protocol support in-network computing and ultimately, the incorporation of photonic switched to accelerator link. These new requirements coupled with larger XPU cluster sizes, are expanding the merchant scale-up switching market opportunity, which we believe will grow to roughly $20 billion annually by 2030.

Our current road map across PCIe, UALink and platform-specific scale-up topologies put us in pole position to service at least half of this market merchant silicon opportunity in the near to medium term with aspirations to address the entire market opportunity over the next several years. Moving to our signal conditioning portfolio, we saw tremendous growth in 2025 with both Aries and Taurus. Looking into 2026, we anticipate strong additional growth fueled by robust secular trends and technology evolution.

Across the portfolio, we are well positioned to benefit from forthcoming protocol specification upgrades for PCIe and Ethernet that will double bandwidth capabilities and therefore, drive additional reach extension content. Given the market size and growth rate, we'll continue to heavily invest in these portfolios with current development stretching out to PCIe Gen 7, UALink 200-gig and 1.60 Ethernet applications.

We remain well positioned as market leaders in both areas, and we look to leverage our AI Fabric engagements to drive additional opportunities within these categories. During Q4, we announced an expansion of our product portfolio to include custom connectivity solutions to address next-generation AI infrastructure featuring heterogeneous compute resources. Our initial prospects in the custom solutions space will help to enable NVIDIA's NVLink fusion scale-up architecture for hybrid racks. And we are seeing opportunities to support additional hyperscalers to provide interconnect flexibility and optionality.

Through close collaboration with hyperscaler customers and leveraging a broad range of foundational technologies and operational expertise, Astera Labs is well positioned to provide a broad set of solutions tailored to custom applications. Next, we are working closely with key customers to define, develop and build optical connectivity engines for scale-up networking. These silicon photonics solutions will ultimately help to enhance both our AI fabric and signal conditioning portfolios as XPU cluster density scales.

We believe that transition to optical connectivity for scale-up applications will be additive to the overall AI networking market sites with copper and optical link coexisting from a system standpoint. This could more than double the merchant scale-up switching opportunity. Additionally, the discrete high-density connectors added through our XK acquisition are seeing strong interest and are being qualified for scale-up applications. Along with our expanding portfolio of AI platform solutions, we continue to make meaningful progress towards further diversifying our cloud infrastructure customer base.

In 2025, we have seen customer activity and engagement accelerate across all product categories as AI and cloud providers look to Astera Labs to help solve their next-generation infrastructure challenges. Many of these engagements have converted to design wins and will meaningfully broaden revenue across multiple hyperscalers as we exit 2026.

In conclusion, Astera Labs has arrived at a critical inflection point. Strong fundamental momentum has been generated over the past several years with solid execution helping to build mature, multi-generation customer relationships. Robust secular trends within rack-scale AI infrastructure and criticality of intelligent connectivity solutions are catalysts for a material expansion of our market opportunity. These factors give us the confidence to reinvest in ourselves our customers and our partners to drive the deployment of AI infrastructure. We look forward to scaling our team with the continued strong emphasis on execution and invest to deliver on our rack scale vision throughout 2026. With that, I will turn the call over to our CFO, Mike Kate, who will discuss our Q4 financial results and Q1 outlook.

Thanks, Sanjay, and thanks to everyone for joining the call. This overview of our Q4 financial results and Q1 guidance will be on a non-GAAP basis. The primary difference in Astera Labs' non-GAAP metrics is stock-based compensation, acquisition-related costs and its related income tax effects. Please refer to today's press release available on the Investor Relations section of our website for more details on both our GAAP and non-GAAP Q1 financial outlook as well as a reconciliation of our GAAP to non-GAAP financial measures presented on this call.

For Q4 of 2025, Astera Labs delivered quarterly revenue of $270.6 million which was up 17% versus the previous quarter and 92% higher than the revenue of Q4 of 2024. During the quarter, we enjoyed revenue growth from our Scorpio, Aries and Taurus product lines supporting both scale up and scale out PCIe and Ethernet connectivity for a wide range of AI rack-level configurations. Scorpio P-Series demand for PCIe Gen 6 switching applications was robust during Q4.

Scorpio X Series shipped preproduction quantities during the quarter. Aries demonstrated growth during the quarter with Aries 6 revenue growing strongly as we began shipping PCIe Gen 6 SCMs for scale-up topologies in high volume. Taurus displayed strong growth during the quarter driven by the ramp of new 400-gig programs for scale-out connectivity for both AI systems and general purpose platforms.

Q4 non-GAAP gross margin was 75.7% and was down 70 basis points from the September quarter levels, primarily due to a higher mix of hardware sales. Non-GAAP operating expenses for Q4 was $96 million or up $16 million from the previous quarter due to the continued expansion of our R&D organization, including the aiXscale acquisition that closed during the quarter. Within Q4 non-GAAP operating expenses, R&D expenses were $70.7 million. Sales and marketing expenses were $11.1 million and general and administrative expenses were $14.2 million.

Non-GAAP operating margins for Q4 was 40.2% down 150 basis points from the previous quarter. Interest income in Q4 was $12 million. Our non-GAAP tax rate for Q4 was 13%. Non-GAAP fully diluted share count for Q4 was 181.2 million shares, and our non-GAAP diluted earnings per share for the quarter was $0.58. Cash flow from operating activities for Q4 was $95.3 million, and we ended the quarter with cash, cash equivalents and marketable securities of $1.19 billion.

Now turning to our guidance for Q1 of fiscal 2026. We expect Q1 revenues to increase within a range of $286 million and $297 million, up roughly 6% to 10% from the fourth quarter levels. For Q1, we expect Aries growth to be driven by a variety of AI platforms across both scale-up and scale-out connectivity. Taurus growth is expected to be driven by increased volumes of 400 gig designs for AI scale-out connectivity. And Scorpio growth will be primarily driven by the continued deployment of our P-Series solutions for scale-out applications and initial volumes of our Scorpio X-Series for scale-up switching.

We expect Q1 non-GAAP gross margins to be approximately 74% with the increased mix of our hardware-based solutions in the quarter. We expect first quarter non-GAAP operating expenses to be in the range of approximately $112 million to $118 million. As previously outlined, our customers continue to present us numerous large revenue opportunities for AI connectivity solutions. The planned increase in operating expenses will enable us to capitalize on these opportunities.

This guidance includes expenses related to an aquihire transaction we closed this quarter, which helped us rapidly scale our recently announced Israel Design Center. Interest income is expected to be approximately $11 million. Our non-GAAP tax rate should be approximately 12%, our non-GAAP fully diluted share count is expected to be approximately 184 million shares. Adding this all up, we are expecting non-GAAP fully diluted earnings per share to be approximately $0.53 to $0.54.

Lastly, I would like to welcome Desmond as our new incoming CFO, and I believe is a perfect fit for the company at this exciting stage of the company's growth trajectory. I look forward to continuing to support the company in my new role, while also ensuring a smooth CFO transition. This concludes our prepared remarks. And once again, we appreciate everyone joining the call. And now we will open the line for questions. Operator?

[Operator Instructions] Your first question comes from Blayne Curtis with Jefferies.

Congrats on the results and congrats, Mike, on the new role. I just want to ask you, obviously, this warrant $6.5 billion is a huge number. So I might already know the answer, but I wanted to ask you about -- it seems like one of the biggest debates is still the acceptance of UALink for these next-gen designs. You mentioned 2 lead customers mentioning it. I'm just kind of curious, as people think about your UALink switch opportunity, particularly your largest customer versus the custom connectivity and maybe them using NVLink. Just kind of curious with this deal, is there any better visibility if you can kind of think about that mix between hybrid boxes and native UALink for these [indiscernible] customers.

Thanks, Blayne. Maybe let me start and then Mike can chime in on the warrant itself. So yes, clearly, AWS announced at re:Inven that the Trainium 4, which is slated to ramp in 2027, will support UALink which was a very positive endorsement of UALink as well as support for NVLink Fusion. Subsequently, AMD has also announced that their MI 500 series will also support UALink again in 2027. So these are 2 very good public announcements in support of UALink and there are several other discussions that are ongoing. UALink ecosystem is humming. We've got great availability of IP, a lot of vendor announcements and so on. And so we will be ready with our UALink solution to intercept the ramp that happens in 2027. Now for NVLink Fusion, this also represents a meaningful opportunity for us. And then before we jump into what the opportunities, I do want to call out the fact that both Amazon, the hyperscaler as well as NVIDIA have chosen Astera as a partner.

And that's a very important statement in terms of the trust that they place in the Astera Labs. So the opportunity itself is to take the native protocol that the XPU or the ASIC speaks and translate that into NVLink. This is a sophisticated question, and we have a solution that we will deploy to address this and given the fact that the solution attaches to the XPU on a 1:1 basis, we anticipate the overall revenues to be in line with [ a switch ] opportunity where we might be selling a UVLink switch.

So all in all, exact mix of how much NVLink Fusion would be deployed versus a native solution would be deployed remains to be seen. But for us, the opportunity is roughly the same for both.

Yes. And just to point out, we did file under [ 8-k ] warrant agreement with Amazon today. It demonstrates our strong relationship with Amazon. Under the terms of the warrant agreement, we're issuing 3.3 million warrant shares that best upon the achievement of performance conditions comprised specified tranches of payments to purchase up to $6.5 billion of our Smart Fabric switches, signal conditioning products and also our optical engine solutions.

I actually wanted to ask you on the last part. I thought it was interesting you've been talking more about optical, it's part of that warrant agreement. Can you maybe just talk about -- you talked about it doubling your TAM maybe timing on that?

So we think that -- when we think that the optical for scale up, the timing should be somewhere in 2028. We do believe that the initial deployment for optical technology CPO, in particular, might happen with scale out and that might precede the deployment of scale-up.

Your next question comes from Joe Moore with Morgan Stanley.

You talked a little bit about the OpEx increase. I guess, it's a pretty big step function up. Can you talk about was the acquisition part of that? I know it was small. And then as you look out to these optical scale-up aspirations, I assume that's expensive. Just kind of any incremental sense of why the OpEx is coming up so much so quickly?

Yes, Joe, over the last couple of quarters, we've been coming -- having a lot of advanced dialogues with our customers, and they're presenting us significant revenue opportunities that we really felt now is the time to really invest in. As we spoke on the call, the TAM is much bigger than we originally expected, just when we measured it just 12, 18 months ago, so we are increasing our investments to pursue these opportunities.

Last quarter in Q4, we did close the aiXscale acquisition. So now we have a full quarter in Q1. And then just recently, in this quarter, we closed another aquahire, where we got a very sizable capability to help us scale up our new Israel design center, where we just also brought in a very exciting capable leadership as well. So this is all to pursue these opportunities. Our customers are pushing us to to develop for them.

Okay. And then you talked about UALink I know there was also a fair amount of noise over the course of the quarter about ESUN Ethernet scale up. Can you just kind of talk about the handicapping those 2 technologies and how you see those technologies coexisting going forward?

Yes. So the scale up networking, Joe, it remains a very large market, and it will include proprietary approaches such as NVLink or Google's [indiscernible] and will also include the merchant and standard approaches such as PCi Express, UALink, Ethernet and ESUN. What we are seeing is that hyperscalers are going to leverage the type of solutions that their software stack is designed for. So for example, if a customer is using a memory-centric protocol like NVLink or like PCI Express, they are likely to continue to use that and transition to UALink as those solutions become available.

At the same time, the customers that are using Ethernet are likely to stay with Ethernet and then maybe move to ESUN and when ESUN becomes available. This is overall a very big market, and there is a lot of room for different solutions will coexist. And we do indeed think that these solutions will coexist. We are primarily developing solutions where our customers are asking us to, which happens to be PC Express UALink and now increasingly on NVLink fusion.

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

On Scorpio, I'm curious if it achieved that 20% of sales, I think, milestone that you had set in Q4. And if it did, is this kind of $200-ish million -- with this $200 million run rate, where do you see the outlook for Scorpio at a high level for 2026?

Yes, Scorpio continues to perform very well. It just launched for the first time in Q2 of this year, and it did break above our 15% for the year, and it grew very nicely in Q4. This is all on the Scorpio P primarily, which is scale-out switching. We did say we just started to ship initial volumes here in Scorpio we'll have increasing volumes as we enter into 2026 with a much more material ramp in the back half of the year. So Scorpio by far is our biggest TAM right now.

So it's growing at a very fast clip as a result of that.

Okay. And for my follow-up, Jitendra, how do you see your content as NVIDIA moves to the Vera Rubin generation versus the Blackwell generation, right? And how do you see whatever the NVIDIA racks are deployed in the Vera Rubin, do you think they let them capture more of their proprietary content? Or do you still see enough opportunity for Astera, wherever Vera Rubin is installed?

Yes, great question. So as we have established now that the opportunity for Astera arises when our customers do custom deployments of the [indiscernible] well or in the future Vera Rubin reference design. We have very minimal opportunity with the reference design itself. The initial ramps that Mike just referred to are what happening as part of these customized deployment of Grace Blackwell platform by our lead hyperscaler customers. And as the hyperscaler customer announced at their public event, they want to continue to deploy Vera Rubin also as a custom deployment. So we will certainly do our best to make sure that we are part of that solution as well as the designs transition from Grace Blackwell to Vera Rubin.

Your next question comes from Tore Svanberg with Stifel.

Congratulations on the results and Mike, congratulations on your new role. I had a follow-up question for Sanjay. Sanjay, when you talked about the SAM by 2030, $20 billion, you'll be able to address half of that today. But you said PCI, UALink and then platform-specific scale-up topologies. And I'm just curious, is the world changing a little bit where there's less reliance on standards and there's more platform-specific scale-up initiatives? And is that also why you're stepping up your OpEx as much as you are this quarter?

So if you think of a scale-up topology, you're interconnecting accelerators. And to that standpoint, the accelerators could be a merchant silicon or could be an internal custom ASIC. And because these are homogeneous planes and everyone is trying to eke out the maximum amount of bandwidth and performance on the connectivity side, so in general, what you will expect is there is quite a bit of customization that will be needed.

Now with Astera, we are unique in the sense that our fabrics are designed to be software defined in the sense that they can be updated to do certain things that are custom and optimized for the specific scale-up topology. So what we have tried to do is not do one silicon for every opportunity, but to be able to leverage the same piece of silicon, but to be able to customize that in many different ways.

So in some ways, we are making sure that the investment that we do is [ banish ] and the differentiation comes from software rather than keep doing a unique silicon for everyone. Now having said that, the thing that we are seeing is a tremendous influx of opportunities. And that's largely coming from the fact that we have spent about 12 to 15 months, sort of being in the scale-up domain, and we have learned a lot.

There's a lot of unique things that become critical when you're designing kilo fabrics and that learning has enabled us to better present our solution as well as the feature set that we're incorporating in our new product lines, and those are gaining interest and support from several new customers. And to support that is where we see a need to step up our R&D meaning the time to invest is now. And this will help us as we think about the longer-term growth of the company as well as our position in the scale of market.

That's great color. And also a follow-up for you to Jitendra. Just to clarify, I think you said Scorpio X. Preproduction is still first half of this year, then you start the ramp with your lead customer second half. But I think you also said that you expect to have some preproduction with additional customers beyond your lead customer in the second half with ramps in '27? I just want to make sure that I got that right.

Yes, that's correct. There is a lot of traction for the Scorpio X family for customers who are trying to use PCI Express in the memory-centric protocol further scale up. We are feeling so many different costs, and we expect some of those to get qualified towards the end of this year and then ramp in '27.

Your next question comes from Ross Seymore with Deutsche Bank.

And Mike, congrats on the new role, you'll be missed. I guess my first question is on the OpEx side of things. You're basically spending about $100 million run rate more than than you were prior. I guess that the revenue opportunity is larger. But can you give us a little bit of idea on what the time to revenue would be in this when you say now is the time to invest. I know the 2030 numbers are big, but is that when today's investment pays off? Or should we expect things sooner than that?

Just any more color on the OpEx would be helpful.

Yes. There is a range, but the technology that we're developing does have a longer lead time, but there are new opportunities that can be turned into silicon relatively shorter and then you have the qualification process with the customers. But you could see from start to revenues in 18, 24 months on the earlier side. But also keep in mind, we've been looking at optical from the inception of the company, knowing that at some point, it was going to be very important as a connectivity supplier.

So we've been putting the pieces together internally and including the new acquisition that we had last year. But we're building something internal that other companies are paying billions of dollars for it to get externally. So -- and by doing this, we're doing it the right way, and we're building [indiscernible] way along with the input from our customers. So our development is closely in line with input from our customers as well.

I guess, Mike, as my follow-up on -- to the Scorpio family. I believe you said it crossed 15% of sales in 2025. So I just want to clarify if that was true. But perhaps more importantly, sort of bogeys as far as the growth rate this year. I believe in the past, you talked about it would cross over and become your biggest product line at some point this year. Is that still the case? Any updates on those sorts of timing and magnitude?

Yes. So yes, so we originally set up for a 10% bogey, we did cross above 15% for 2025. And again, that's all just P-Series, X is for scale up is a much bigger, larger TAM for us and that we're starting to shift initial volumes in the first half, but the more material step up in the back half. So the commonization of those 2 will put us on a trajectory for it to be our biggest product line. But Aries and Taurus and Leo are all growing as well. So it's hard to know exactly when we cross over.

But definitely at some point, it will. It will -- it's going to drive very good revenue growth for us.

Your next question comes from Sebastien Naji with William Blair.

Congrats on the results. Just in terms of the more customized solutions that you're building, is it right to think that your average ASP for the solution or the content opportunity should go up meaningfully versus some of your existing products? And is there anything to call out in terms of maybe a different margin or different profitability profile for those solutions?

And to confirm, you're referring to the Scorpio X?

That's right.

Or the customs, sorry, I didn't hear the first part.

Yes. Sir, for the customized solutions that you're building for some of your hyperscale customers?

Yes. So the custom solutions that we build like, for example, the NVLink Fusion opportunity that we noted. So the engagement model, of course, tends to be different. But the attach rate will also tend to be higher. So I think from a volume standpoint, it will be at a certain rate. But the ASP is there will be some considerations that have to be applied because there will be blocks that will come from the partners that we work with. So end of the day, I think, like Jitendra highlighted, when it comes to our revenue content, both based on a native switch versus a custom solution, things sort of even up where on the native switches, it tends to be one switch being shared across a few accelerators whereas in the custom solution type of products, you are doing it on per accelerator, which means that you will have higher attach rate, although the ASP might not be at the same level.

Sebastian, in general, a helpful thing that we've found is every generation of XPU, our content has grown up so far. And we continue to head in that direction to having more dollar content per XPU generation.

Got it. Okay. That's really helpful. And maybe as a quick follow-up just on the Taurus line. Is there a way to think about how much of the strong growth you're seeing there is coming from just strong underlying market growth versus Astera's ability to gain share in that market?

Yes. So it's -- I want to say both, but definitely the speeds have gone up from 400 to 800 gig. The need for active components, whether it's onboard or within cables in the form of AEC is growing. So that fact is not changing. Our business model is slightly different, as probably you're aware, we don't do the whole cable. We do the modules that go inside the cable assemblies and rely on our cable partners to provide the at-scale deployment with multiple multiple vendors supporting the same opportunity. So to that standpoint, what I would say is that we are seeing that transition happening.

You can see that with some of the numbers that we shared today. Our Taurus revenue has gone up Q4. And generally for the whole year, it's about 4x. So we expect that trend to continue. Again, we won't be called in on day 1 because of the business model that we have. But as volume picks up even for 800 gig, what we're expecting is that Astera will come in strong with multiple cable suppliers using our module to support the high-volume brand.

Your next question is from Sean O'Loughlin with TD Cowen

I have a question and congrats on nice results. And Mike, congrats on your retirement and hopefully, a little bit of an easier new role at the company here. I wanted to ask on the warrant agreement. And maybe get some background on sort of, one, there's a period of exercise to 2033, I believe. And then there's a dollar amount associated with it to the extent that you're able to talk about sort of what drove the 2 companies to reach the agreement? And is this -- should we think about this as maybe like incremental to what you already had expected to be doing with that customer over that time frame? And then anything you could give on linearity that would be great.

Yes. I can't go into much more detail than what we publicly disclosed. In the 8-K, we did file the full warrant agreement. So you can get some of the more material terms out of that filing. It does demonstrate the strong relationship that we have with Amazon. This is a follow-on warrant. We've had a previous warrant agreement in place. And what happens is the warrants are earned as revenue milestones are achieved, which is the $6.5 billion that we outlined.

This -- to account for the warrant, you do take a noncash charge for the value of what Vest,and that goes directly against revenue and effectively directly against gross margins as well. So as the warrants are achieved, we are kind of modeling a noncash hit to gross margins of about 2 points a quarter starting kind of in the Q2 time frame. The warrants have a like you outlined for 7 years.

Great. And if I could ask another, unfortunately, financial follow-up. On the NVLink Fusion agreements and whether it's with Amazon or whether it's with any other hyperscaler that might adopt that topology. What is the -- what does that financial arrangement look like? Is that a license that the hyperscaler is paying directly to NVIDIA and they sort of grant you the IP? Or is that something that gets incorporated into the ASP and the margin profile of that product?

Yes. obviously, we can't answer that question just given all the NDAs and other things that we need to be respectful of. In general, what I would say is that it opens up a completely new set of opportunities for us to play in NVLink ecosystem, which we did not had a play so far. So to that standpoint, this is additive to everything that we are talking about. In terms of the exact business model, I think we'll let hyperscalers or NVIDIA provide more color.

Your next question comes from Karl Ackerman with BNP Paribas.

Yes. Mike, you know that some of the increase in OpEx is being driven by your investments in optical and follow our acquisition of aiXscale. Do you anticipate your customer opportunity for integrating your optical [indiscernible] technology is larger for a switch portfolio than your signal conditioning and SCM products. And I have a follow-up.

Yes,. I would -- maybe let me take that and then Mike can chime in afterwards. I would say that the opportunity for including optical into scale up is actually a very large opportunity, probably larger than the signal conditioning opportunity that we have of the order of what we are saying for scale of fabric connectivity. We have not quite sized it up exactly, but it is a very large opportunity. and we are working very closely with our customers to understand what their requirements are and what their time frames are, and we'll be ready to intercept those.

We believe for scale up, in particular, that's likely to be in 2028.

Also just to maybe add to that, right? I know optical is an important area for us to invest, and we are doing that with some unique architecture and capabilities. I'd also say that the increase in investments we're doing like Mike outlined, these are also servicing opportunities on the fabric side, which are much more based upon existing engagement and as customers are seeing the value of what we offer, they're coming back with requests for additional lane count or Radix configurations or features and things like that.

So the investment that we're talking about, including the Israel team that we set up, which will focus on AI fabrics, it's been done in a way that we get the near-term, midterm and long-term growth, setting us up nicely for building on the momentum we have and getting to a different scale from a revenue standpoint.

Got it. Very helpful. you indicated that you have Scorpio P design wins now with 3 hyperscalers, 2 of which are new. Are your Scorpio P designs based on custom compute designs? Or are they also being designed on custom GPU racks as as well? And then clearly, you're having traction here. So you could perhaps address the opportunity that you see with Scorpio P extending into 2026.

I know there's a lot of focus on Scorpio X, but perhaps you could spend some time on the engagements you have as well as the opportunity that you see on Scorpio P into '26?

Scorpio has been really -- I mean it's mostly going into scale out use cases -- now that we have granted a chip in reasonable volume over the last 3, 4 quarters, we are seeing additional customers, specifically like we called out 2 new hyperscalers that have adopted it. They go into production a little bit later towards the end of this year. These design-ins are supporting platforms that are both merchant GPU-based as well as custom accelerators.

And we do expect these to add meaningful revenue for us in 2027. But beyond this, again, these are the ones we called out, just given the significance of these design wins. But in general, I want to say for the P-Series ever since we announced, it's been a attracting a lot of customer interest and traction, and we do have several design wins on that in different lane coconfigurations and so on.

So that continues to be a device that has been. Now I want to say at this point, we are probably still the only 1 that's in high volume production with our Gen 6 switches. So it's serving many customers and use cases.

Your next question comes from Srini Pajjuri with RBC Capital Markets.

And let me echo my congrats to Mike. I look forward to, I guess, continuing to work with you, but it's been a pleasure over the past 20-plus years. My question, there's been a lot of skepticism about the growth of Aries retimers, and you continue to show that this -- obviously, this segment continues to grow and you're projecting growth for this year as well. if you could talk to us about what's driving that growth? Is it primarily the units? Or is it the PCIe 6 transition and that -- or is it new customers?

And then as we go into '27 '28, how should we think about any potential implications as the market goes from copper to optical?

Yes. So the -- if you think about PCI Express, PCI Express is really the nervous system inside of a server. All of the significant components, whether it's a GPU or a CPU or a NIC card, they talk to each other over PCI Express. And we established this early lead in PCI Express with PCI Express Gen 4 Gen 5 and now extending that into PCI Express Gen 6. We see not only the continued growth in the, what we call the chip-down applications, but we also see more growth of PCI Express in cabled applications where we set our smart cable modules that does give us an uplift in the ASPs.

These devices are used in scale up applications. And again, as we all understand, scale up is a very rich opportunity where we have multiple connections and a lot of them. So we are definitely benefiting from that trend with PC Express going from Gen 5 to Gen 6.

Now your question about transition from copper to optical. So most of our customers will continue to stay with copper for as long as they can. That has been their preference all this time. We will reach a stage where maybe the reach or the bandwidth is just high enough and is not able to be supported by copper. And in those cases, we do expect a transition to optical, but copper and optical will continue to coexist for a long period of time. The transition to optical will likely come in the form of first like pluggable optics may be followed by a new package optics and eventually the Holy Grail, which is the copackage optics, which all of us are working towards.

But we expect to see the first deployments of co-packaged optics for scale up sometime in the 2028 time frame.

Okay. Got it. And then the other debate, Jitendra, is obviously UALink versus ESUN. I guess, hypothetically, if the market moves more towards ESUN, I'm just curious, I mean, do you have the capability to give to ESUN? And if so, how quickly can you do that? And I'm just curious because, again, this is a debate that we can answer sitting here because it is something that's going to happen in the future. But just wondering if the market were to go to ESUN, how do you see your positioned? What kind of role do you anticipate playing in that market?

Yes. So I would say that we are much closer to the action, so we understand who's doing what, largely for the initial deployments of the new scaler protocols, whether it's UALink, Ethernet or ESUN. From a capability standpoint, there's definitely within reach. We can -- if we choose to design an ESUN based solution we can. However, as has been the trend with Astera, we listen to our customers very closely. And so far, everybody is telling us to focus on UALink and that's what we are focusing on. As you say, if things were to shift towards ESUN, we certainly have the capability and with the addition of the additional resources that we are deploying, we can definitely go in the direction of additional solutions. As a matter of fact, I would say, over time, our aspiration really is to address the full connectivity TAM and not just limit to any 1 particular protocol.

Next Question comes from Quinn Bolton with Needham & Company.

This is Robert on for Quinn. First wanted just to double click more on to the [indiscernible] product offering. Can you maybe discuss how qualifications for your 800 gig AECs are progressing? And do you expect to be shipping 800-gig AUCs to multiple hyperscalers this year?

Yes. So 800 gig is obviously starting to ramp up right now. And to that standpoint, what I can say is that we are very closely engaged with that from a qualification process standpoint. Given our business model, which is slightly different, like I highlighted early on, we come in as the volume starts expanding. So to that standpoint, you can expect a similar transition happening to our business as well with more deployments for 800 gig.

800 gig, like we highlighted before, is broad-based, meaning there are multiple customers that are using AECs for 800 gig. And those are opportunities that we expect to gain from from a revenue standpoint.

And then just quickly on Scorpio P. Just wanted to double-click a little bit more into that color. Do you expect these additional P-Series customers to reach kind of the size and scale I guess, down the line as volume ramps with them kind of in the second half of the year and beyond. I guess, 10 of these new customers be as big as your main hyperscale there?

Yes, there is a potential for that.

At least the ones that we called out now there are, like you noted many P-series opportunities and design wins we have, but specifically to the hyperscaler opportunities that we called out, these are mainstream use cases. So we do expect them to have the revenue impact like what we had so far.

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

I just had 1 quick 1 and then a longer-term one. So you announced 2 new hyperscalers for PCIe late in '26 or say in saying revenue in '27. Are those U.S. hyperscalers? Or are those Chinese hyperscalers? And then on the longer term, I think you mentioned on the call about the MI 500 series, supporting UALink. Obviously, the 400 series supports UAL, but it's over Ethernet. Are you saying that the 500 series will be native UAL? Or is it still going to be Ethernet supporting, obviously, that's a big difference in what switches you're using.

Thomas, I don't want to speak for our AI customer. But if you look at some of the publicly released information, they have said that they will continue to support both, but they've also said that they believe UALink native is the highest performance scaleup protocol.

There are no further questions at this time. I'll now turn the call back over to the presenters for any closing remarks.

Thank you, Karlie, and thank you, everyone, for your participation and questions. Please refer to our Investor Relations website for information regarding upcoming financial conferences and events. Thanks so much. Have a great day.

This concludes today's conference call. You may now disconnect.

Welcome back, everybody, to the second day of Needham's 28th Annual Growth Conference. My name is Quinn Bolton. I'm the semiconductor analyst for Needham. My pleasure to host this fireside chat with Astera Labs founded in 2017 and headquartered in San Jose, California. Astera Labs provides rack scale AI infrastructure through purpose-built connectivity solutions, the company's intelligent connectivity platform integrates CXL, Internet, NVLink, PCIe and UALink semiconductor-based technologies with the company's COSMOS software suite to unify diverse components into a cohesive, flexible systems that deliver end-to-end scale-up and scale-out connectivity.

Joining me on stage from the company are Jitendra Mohan, Co-Founder and CEO; and Nick Aberle, VP, Treasurer and Investor Relations.

Jitendra, Nick, thank you for joining us.

Thanks.

Thank you. I couldn't have said that better myself.

Thank you. Just kind of starting off some big picture questions. Astera is a sort of leading play on rack level AI infrastructure, which you call Infrastructure 2.0 it's a brief introduction for investors that may be less familiar with the company, can you provide an overview of the various products that supplies to enable Infrastructure 2.0.

Yes, certainly, I see many familiar faces and some -- and actually many new ones, so Happy New Year, everyone. Maybe what I'll do is I'll start on the other side, what our vision is, and then we'll translate that back into the products.

So the vision that we have at Astera Labs has been this way ever since we founded the company, is to really provide all of the connectivity infrastructure that our customers need. What has changed between 2017 when the company was founded to now is that the basic unit of compute has become a rack. So back in 2024, you would have a server, maybe this tall that housed 8 GPUs, and that was the workhorse of the AI industry actually still is today. But moving forward, that has changed to a full rack.

And at Astera, we want to provide the full connectivity infrastructure that goes into this rack. That consists of switches that connect all of these GPUs together, typically called a scale-up switch. It is all of the single conditioning components like retimers, gearboxes, active electrical cables that aid with the connectivity as well as the software that ties it all together.

So for us, that translates into multiple different products, both in hardware side as well as software. On the hardware, we have our Scorpio P and X family of fabric switches. These are -- Scorpio P families responsible for PCI Express connectivity, typically used in scale-out applications. Scorpio X is responsible for GPU to GPU scale-up connectivity. So that's used for building scale-up networks. And then we have our Aries retimers that are deployed either as a chip down on a board or as an active electrical cable, both for scale-out as well as for scale-up applications. Then we have our Taurus products, which are also for signal conditioning, but for Ethernet. And these are typically deployed as active electrical cables where Astera provides a smart cable module.

And last but not the least, we have our Leo products, which are a little bit different in that they address the memory bottleneck in these AI systems, allowing you to add DDR5 memory with CXL connectivity, starting with general purpose compute applications but then eventually also making its way to AI. At this point, I'm very happy to say that all of these product lines are in production and contributing meaningfully to our revenue.

Excellent. And we'll get into a lot of those products in a moment. But when you look at your product portfolio, you have PCI and UALink fabric switches, you have the signal conditioning products. What are the company's core competencies that sort of tie all these products together?

So I would say there are maybe 3 things. So first of all is our architecture. So we came up with this architecture, which is a software-first architecture where we try to do whatever we possibly can in software or in firmware. So there are many, many microcontrollers that are embedded controllers that are in our chips that are doing a lot of the processing. And the advantage that, that gives us is the solution becomes extremely flexible. So as the needs of AI change, you get a new workload, we can morph our solutions to address that optimally. We can customize the solution to our end customers' requirements because everybody tries to do their own scale-up networks or their own infrastructure development a little bit differently. And so we have the ability to optimize and customize this for our end customers.

As a result of this flexible architecture, we also have tons of diagnostics. So if you think about somebody who's deploying hundreds of thousands, millions of these GPUs and XPUs together, they want to make sure that their infrastructure continues to hum together very well. And for that, we need to create a lot of -- or produce a lot of diagnostics information, the health of the links, the health of the chips, the health of the systems, and we tie this all together, and this is the second component with our COSMOS framework. So COSMOS is a software framework that ties all of our chips together, allows our customers to customize the solution, optimize the solution, collect all of the diagnostics and really make their infrastructure hum smoothly.

The third component, which I think is probably the most significant one is not technical at all. So it is really the trust that we have developed with our end customers over the last 7, 8 years, where they are willing to share with us what their road map is, what solutions they want to bring to the market 1 year, 2 years, 3 years out. And that gives us a very significant leg up in how we define our own road map, what are the features that we put into our products? What are the features that we don't put into our products. Oftentimes, that's even more important. And certainly, what products to build and at what time frame to build that in. So we're really proud of the relationships that we've built with our lead hyperscaler customers.

That's a great overview. So you've got your software-defined hardware architecture. You've got the COSMOS software suite, you have the customer trust. All of that has enabled you to build sort of a competitive moat. Where do you think you have the greatest lead across the product portfolio relative to some of your competitors?

That's a good question. And I would say that our presence in different segments is different but we do aspire to have leadership position in all of them. Maybe the greatest lead today just from a volume or market share standpoint is in the retimers. We are the de facto choice when it comes to PCIe retimers. Very rapidly, we are following that up with the PCIe and switches, Scorpio P and Scorpio X families, where we were not the dominant player at PCIe Gen 5 family. Broadcom was actually the dominant player. But we are well on our way to establish that leadership position with PCIe 6. We were the first ones to introduce a PCIe 6 switching family. We are the first ones to introduce a PCIe retimer family. And in fact, if you think about PCIe itself, that is where we are the strongest with a full portfolio from retimers to gearboxes to cables to switches.

So if a customer wants to deploy a PCIe-based, not necessarily PCIe but PCIe-based solution, then Astera is the one-stop shop for that. And we'll translate this into UALink in the future. where all of the hardware components, software components will run faster with UALink but we aspire to do that with UALink as well. Ethernet is perhaps where we have a less dominant position. Ethernet has always been a share game. Customers like to have multiple sources. And so we compete for share with other companies in the Ethernet space.

Great. Last sort of question just about the sort of AI spending environment. There's been some concerns about NVIDIA and circularity of investments. There's been some concerns about increasing amount of debt to fund CapEx investments. Just what are you seeing from your customers, their spending plans, their road map plans? Do you expect continued growth -- strong growth in 2026? Are you seeing any sort of clouds on the horizon?

So first of all, the circular investments are not my area of expertise. My head also spins. But fortunately, we don't have to worry about it as much because I think what I am very confident of is just as an end customer myself of Gemini and ChatGPT and Nano Banana and all of the new things that are coming up. I think the end customer demand is very strong. So who funds what as -- vendor-backed financing has always been a thing in the past as well. The reason it failed in the Internet era was there was no end demand. Everybody was just sort of fueling this. Here, the fundamental difference is the end demand, I think, is very strong. Everybody who's deploying these AI system is saying that, hey, there is ROI. They are getting returns on their investment. So -- and we see that translated into our order pattern as well. So we don't see any evidence of a slowdown in 2026 or in 2027.

Excellent. I wanted to move now to the products. We'll start with the Scorpio X as it will become, I think, over the next 12 or so months, your biggest product line. Last month, AWS hosted its re:Invent show and made a couple of key announcements that I think affect the company. First, they announced the Trainium 3 platforms were moved to PCIe based switch fabric topology in the scale network. And then second, as they look forward to their next-generation platform, they talked about migration of scale-up to UALink and NVLink. So how do you see these announcements at a large customer affecting your business?

These are all very positive for us. None of these are any surprises. We have known about this for some time. But of course, we can't talk about it until our customers make the announcement. So as I mentioned earlier, on the PCI Express side, we have a full portfolio. We are very closely engaged with our customers. So we see that as really our kind of backyard in terms of the benefits to us. So long as customers are deploying PCI Express-based scale-up. And while AWS announced that at re:Invent, there are other customers, many other customers actually, we mentioned 10-plus engagements in that space. So really excited about the potential for Scorpio X. And you are correct, Quinn, that eventually, it will become -- it is on track to become our largest product line overtaking both Scorpio P as well as the retimers over time.

The other -- the announcements that were important was this is the first public admission by AWS for supporting UALink. So that is both good for us at Astera as well as good for the industry for UALink deployments as a follow-on standard to PCI Express. We do think that customers who are currently looking at PCI Express and PCI Express base scale-up networks will over time transition to UALink. So I think that is great.

NVLink Fusion was a new announcement that caused a little bit of a stir in the market to begin with. So NVLink Fusion is where you are able to use the nice ecosystem that NVIDIA has curated with the NVL 72, including the power components and the liquid cooling and all the nuts and bolts and everything and the NV switches in particular, and use that with your own compute rack.

So not only does it present a new opportunity for Astera, previously NVLink was really off limits. Now we have an ability to build a solution that attaches 1:1 with XPU. For every XPU, you need this solution. This is a new TAM. So we are very excited about it. We have one hyperscaler customer that's deploying it. Hopefully, we'll see other customers deploy it as well over time.

And is that solution something that takes UAL and converts it, the protocol conversion to NVLink? Is it more you'll have a native NVLink system and you just have chips that can go into providing that NVLink connectivity?

Yes. So the way this works is XPUs talk in their native language, whatever that may be, maybe it's PC Express, maybe it's UALink or maybe it's something else. But the NVL 72 infrastructure talks NVLink. So you need to do a protocol translation. And it is not like a retimer type product that a certain data comes in and the same data goes out. There is a lot of complex data flows that need to be managed, security that needs to be managed, links that need to be managed. So it ends up becoming a complex solution where we can command a healthy ASP for it. And with the fact that it is a higher attach rate than switches, in terms of the dollar opportunity, it becomes very similar to us in both cases.

Excellent. Okay. Great. The other announcement was Marvell's recent acquisition of Celestial AI, which is one of the start-ups in the CPO segment. And with that acquisition, Marvell and Celestial named Amazon as their lead customer for CPO solutions. And so as you think forward to a future system that integrates CPO, both on the XPU and the switch fabric, how does that affect your business? Would you look to come in with CPO capabilities yourself? Would it be a chiplet design where you can do the switch fabric and then have chiplets that provide the connectivity off chip? How should investors be thinking about adoption of CPO in future fabrics?

Yes. So at the highest level, CPO is a net increase in TAM. CPO, optical solutions are a lot more expensive than copper solutions. So if the -- as the world shifts to optical, it is good for anybody who's playing in the optical solution, and we'll absolutely be playing in there as well. That is also a reason why I believe that our customers don't want to go to optics because they have to pay a lot more for optics, including in power and in cost. So overall, it is up to the industry, including folks like us to deliver a reliable low-cost solution that enables optics and therefore, allows disaggregation of these racks. And we are already actually in close engagement with our customers on when this transition is supposed to happen, when they would like to deploy copper -- sorry, optical, and we will be right there.

Now to your second point, you cannot deploy optics in isolation. So you cannot say I have an optical engine and now you can deploy optics. You must have a switch that's capable of talking optics and you should have an XPU that's capable of talking optics. And it is our plan to work an optical engine to create an optical engine that will enable Scorpio family to be -- to have optics I/O as in addition to a copper-based IO. We announced the acquisition of aiXscale, which is a company that does the packaging piece, which we believe is the most critical piece to scale optics. This allows very efficient connection of optical fibers into the silicon photonics chip.

Now unlike Celestial, for example, where they have chosen to go down a particular path of silicon photonics, we are open to multiple solutions. If we need to work with, let's say, a Celestial type of a solution, we can happily work with that. We will have our own silicon photonics solution offering that if a customer wants to use that, they can. Or if they want us to work with a third party, we will work with anybody whose silicon photonics our customer wants to use.

And then there is the -- so you have packaging, then you have silicon photonics and then you have an electrical IC, I think what you mentioned to a chiplet. So this chiplet is also very important because it has to understand what the XPU is talking. And this is where the analogy with NVLink Fusion becomes important that if you already understand what the XPU is talking, it is easier for you to produce this chiplet, which understand the XPU on one side and can produce the electrical signals to drive the silicon photonics.

So the summary is that when the industry is ready for deploying CPO, Astera will have our own solution. But at the same time, we are also open to working with what our customers want to do in terms of which silicon photonics they choose.

There's been a lot of investor questions about CPO over the past couple of months. Where do you think -- well, I should say, where and when do you think that the market begins to adopt CPO solutions? Do you think it is adopted in scale-out networks or scale-up networks? And is it a couple of years away? Is it potentially further out? Just any big picture thoughts without trying to specify any.

Yes. This actually came up, I think it was like 6 months ago when I was in one of these conference and somebody predicted the demise of copper. So it's not -- first of all, it's not going to happen. We do think, like you said, that scale out is the first place where CPO gets deployed. And you can see that happen with the demonstrations that Broadcom and NVIDIA have done with their respective switches. And there is a definite advantage to that because the alternate there is pluggable optics. So compared to pluggable optics, CPO provides all kinds of advantages. Compared to copper, it does not. The advantage that it provides, of course, is longer reach.

So we do think that scale-out is where optics get deployed first. Whether that happens in '27 or '28 or '29 remains to be seen. But if you look at optics for scale up, that one we believe is a '28, '29 proposition. And part of it is also dependent upon the industry. If like I said earlier, if you can come up with an optical solution that's as cost effective, as reliable and low power as copper, then maybe that transition becomes faster.

Excellent. You mentioned that the 10 engagements or over 10 engagements for PCIe scale-up switches. How are those engagements progressing? Are you sort of more or less bullish on the PCI scale-up opportunity than, say, you were 6 months ago?

Definitely more bullish. The amount of traction that we have for the Score P, X family is just through the roof. I have not seen anything like this in my whatever, roughly 30-year career. We mentioned 10-plus engagements. AWS talked about their move to switch to PCI-based solutions for switching. And there are many others that are exploring this because frankly, PCI Express is a very good protocol. It is designed to be low latency. It is designed to provide memory semantics. It's a full open standard. You can get components from multiple vendors.

So a lot of the customers are working on it. Some of that is a confirmed design win for us and actually ramping in Q1, starting to ramp in Q1, maybe more volumes happen in the back half as the XPUs get deployed. On the other end of the spectrum, some of them are more in exploration stage. But the amount of excitement that customers have and the traction is just we are not able to keep up.

How long do you think the PCIe switch fabric, scale-up switch fabric lasts. I think there's some perception that this may be a 1- or 2-year cycle and then you'll get sort of broad conversion over to either UALink or ESUN. What's your view? I mean, is PCI as a switch fabric or a scale-up fabric likely to stick around for perhaps longer than investors believe?

Yes, PCI is still growing. I think it will stick around longer than anybody believes. In fact, the engagements that we already have on our PCI Express portfolio already are ramping through '27 through '28. And now customers are starting to talk about PCI 8 already. So I think PCI Express as a protocol will continue for a long time, and PCI Express a scale-up is already -- people are committed to '27 and even in 2028.

Having said that, those who are using PCI Express today are likely to turn over to UALink as that becomes established. So we think that there will be initial deployments that happen in 2027 and perhaps starting to take speed in the later half of '27 and '28. We also think that folks that are doing PCI Express today will move to UALink because they are both memory semantics-based protocols. So any software optimization that folks have done for PCI Express translates rapidly over to UALink. We just get to run that protocol a lot faster and get better performance in their system.

And for that reason, those who are using Ethernet today are likely to continue to use Ethernet. Ethernet does have higher latency and a different way of setting up the memory addressing. But if you have optimized your protocol in your XPU software to use RDMA over Ethernet, it is likely that you will stick with Ethernet until both UALink and ESUN become established and maybe there is some benefit to switch over. So for the last part, large part, we feel that customers will stick to their swim lanes. And maybe out when both standards are fully established, there might be some people that move over.

On the UALink, looking forward to the UALink switches, you're starting from a very strong position in PCIe fabric switches. As we look to UALink, you're developing UALink switches. Your competitor, Marvell is also developing UALink switches. They recently announced the acquisition of XConn to bolster their switch capabilities, I think, both on the Ethernet side as well as the UALink side. Just maybe talk about how you feel positioned in UALink switch as that -- I think both companies, yourselves and Marvell expect to start sampling switches later this year.

It's a good validation of what we have believed in and said for a while that scale-up switching is a very big space. And of course, Marvell made the acquisition of XConn to get better footing in the PCIe switching space. They acquired Innovium a while back. So with that, they have Ethernet, and they did announce UALink as well, and so did we. Now I think the difference between the 2 is that we have been in this space now for some time. We announced our Scorpio P and X families, not just announced, but demonstrated working silicon a year ago at OCP in 2024. So during that time, not only have we taken these devices to production and generated significant revenue, we've actually generated significant learnings of what works and what doesn't work. So PCI Express itself is a complicated protocol.

But when you deploy it for scale up, there are many learnings and many customizations that we have had to do based on our COSMOS software. And those learnings will stay with us and translate to the next generation of the devices, whether it's in PCI Express or in UALink. So I think that's a unique advantage that we have as we deploy the next generation of our Scorpio family for both PCI Express and work on UALink.

There is another part that is completely separate from the protocol, which is management. So we also understand what type of data and diagnostics information customers look for in order to manage these racks. And that knowledge, again, will carry over from the current generation of Scorpio X devices to the future generations.

Great. I wanted to sort of move on to the P-Series PCI switches. I think the original design was sort of in AI head nodes on sort of custom versions of NVL 72 rack. So curious your outlook on sort of P-Series switch on future AI head nodes. It looks like more recently, you're starting to see Scorpio P perhaps being used as part of the PCI fabric switch to help with onboard connectivity between XPUs on a tray and maybe a newer use case that some of us when you announced the product line weren't thinking about. But what's your overall outlook for Scorpio P?

Yes. So P-Series, P, by the way, stands for PCI Express. So P means PCI Express. So any standard PCIe application makes use of Scorpio P-Series. X is for scale-up, and it's typically a customized version of the PCIe protocol specific to scale up. So that's the difference between the 2. So Scorpio P promises to be even a wider, more diversified user base than Scorpio X, and we are starting to see that. We announced at the last earnings call that we now have a new hyperscaler customer that's deploying P. There are others that are in the hopper as well. So over time, we expect to basically land all of the major customers on the Scorpio P platform. So again, bright future for Scorpio P. Scorpio X is for the customers that are using Scale-Up.

So to answer your question, 2025, Scorpio P was the main revenue contributor. It will continue to ramp in 2026 but we will start to see Scorpio X layer in on top of Scorpio P in '26.

And you're seeing the Scorpio P both across NVIDIA -- custom NVIDIA platforms as well as hyperscaler ASICs...

Yes, sorry, I didn't answer that. Yes, absolutely. So I think with complete hindsight, we made a very good decision in developing that particular form factor of Scorpio P, the 64-lane device, which was designed for AI workloads. It doesn't do everything, but it does those AI workloads very well. And it allows you to connect an XPU that might be running, let's say, Gen 6 to the rest of your infrastructure, whether it's SSDs or network controllers and so on, that might still be Gen 5. And we are seeing more use cases across different hyperscaler customers that want to use Scorpio P-Series that device in this application.

Excellent. I wanted to switch now to the Aries PCIe retimers. First, can you speak to the emergence of PCIe Gen 6 devices over the past year, how you're positioned to capitalize on this opportunity relative to the PCIe 5 generation?

So now our PCIe 6 generation is in full production. We introduced the first sort of demonstration of that right around our IPO time in '24. So just like I talked about on the switches, it has taken us a lot of learnings to go from an initial proof of concept to full production. So I feel very confident about the PCIe Gen 6 generation. Gen 5, we have a leadership position, very strong market share. And I think we are on track to deliver the same thing with the PCIe 6 generation as well. There are many of our competitors that have announced similar products, but we are yet to see any of them go to full production or have the same robustness that our Aries family has, both by carrying over the learnings from Gen 5 plus all of the new learnings for Gen 6, that are now incorporated as part of the Gen 6 family.

As we go from Gen 5 to Gen 6, the opportunity sometimes actually is larger because the data rate is twice as high. And so the signals don't quite reach as far as they did with PCIe Gen 5. So we have the opportunity of higher attach as well as higher ASP when we go from Gen 5 to Gen 6. So on the whole, our retimer portfolio grew from '24 to '25 and it's going to grow again from '25 to '26.

One of the concerns we've heard from investors is NVIDIA's NVL reference designs have largely moved away from retimers on reference design. And so talk about what are some of the drivers of the PCIe retimer market more broadly. Obviously, the upgrade from 5 to 6 gives you an ASP lift. shorter distance probably means more retimers in general. But what drives the market going forward?

Yes. So I think there are 2 answers to that. One is just the signal conditioning answer. There is physics that if you are going to run a faster protocol, you don't go as far. And therefore, you need to add a retimer somewhere in the middle. Sometimes that retimer happens to be on our board. Other times, the retimer is actually in the cable. So just like we have our Taurus cables that are AECs for Ethernet, we have PCI Express-based AECs as well, where we supply the smart cable module and somebody builds the entire cable. So we are seeing a lot of growth in that deployment of PCIe 6 retimers.

In addition to that, the use cases that we have for scale-out continue to be there as well. So we have that. When we -- when NVIDIA moved to a different architecture on Grace Blackwell, while we lost content in the reference design, we actually gained a lot more content when hyperscaler customized the deployment of Grace Blackwell platform into their own data centers. And they have said that they're going to do the same thing for Vera Rubin. So I think that leg of growth continues as we go from Grace Blackwell into Vera Rubin. But the fact that customers are deploying PCI Express or PCI Express-based protocols for scale-up is perhaps a bigger opportunity for these devices.

Two other questions. Are you starting to see or when do you expect to see maybe broader adoption of PCIe retimers in the general purpose server market?

Yes. We already are shipping PCIe retimers into the general purpose market but it has not taken off to the same level as AI just because the endpoints are not quite there yet. So if you look at the endpoints, the network cards, the SSDs, they're still PCIe Gen 5 and the server form factor is such that in all cases, you don't need to have signal conditioning. So when the CPUs go to Gen 6, which they are now, and these network cards and SSDs go to Gen 6, which they, at some point will, that's when we will see kind of more Gen 6 retimers going into the general purpose compute side of the market.

Okay. One of the things I think we've seen is it feels like the generations between PCIe Gen 5, 6 to 7 and 8 may be accelerating, which sounds like it could create gearbox opportunities for you. Talk about that. And have you seen any of your -- you talked about not seeing a lot of competition yet from the Montage, the Marvells, Broadcoms or Credos. Do any of those competitors have a gearbox capability built into their retimers?

Again, I'm not aware of that because gearbox is a very special feature that you need because you're translating from PCIe Gen 5, which uses a different protocol to Gen 6, which is even though they are both PCI, there is actually a difference in protocol. So because we have a switch, we have a much deeper understanding and the IP on how to do this gearboxing function and translation, which is what we use for our PCIe 5 to 6 gearboxes. So unless somebody has that switching expertise, it will be a little bit difficult for them to do the same thing.

The other point that you made is also very valid, which is these generations are coming in faster and faster. So you have a need for a gearbox from 5 to 6. Well, then when everything goes 6, you don't need this gearbox. But by then, somebody has gone to Gen 7. So you might need a Gen 6 to Gen 7 gearbox. So while it's a transitionary socket, then we fully acknowledge that, you will continue to need this, it seems like for successive generations.

Great. I wanted to move to the Taurus AEC product line. In the second half of 2025, it looks like Taurus was one of your faster-growing products. Maybe just talk about what drove that strength? Was it mostly 400-gig AECs at your lead customer and sort of your thoughts on the transition to 800-gig AECs just more broadly across the industry?

Yes. So the second half growth, you're correct that it was the fastest-growing product line in Q4 was driven by one customer, which is our lead customer for some of both AI and general purpose compute applications within the customer. And it was driven by both factors, just the overall increase, just the rising price, they're deploying more XPUs and therefore, they need more cables as well as share gain that we had in that application. So of course, good news. That will continue to give us benefit for the rest of 2026.

What is also more exciting in 2026 is as the industry broadly transitions from 400 gig, where there is limited opportunity for AECs to 800 gig, where the opportunity becomes a lot wider. So we are engaged with many customers in samples and so on of deploying 800-gig AECs. So this will be a new layer of growth that we'll have for 2026 with 800-gig AECs.

And are you seeing those -- like when do you think that starts? Is that earlier in the year? Is that later in the year in terms of 800 gig Taurus?

Yes. So for us, the first half will be spent in qualification and the volume ramps will come in the second half.

Great. And then maybe just in case some in the room aren't familiar with your go-to-market strategy versus that of Credo, you make the smart cable module, you provide firmware, you don't build the full cable. Credo does vertically integrate and supplies the entire cable. What are the pros and cons of the 2 approaches?

Yes. So first of all, we are -- both Credo and us are going after the same problem statement at the same time that we are addressing in different ways. The Credo approach is that they provide the whole cable, and there are some benefits of that. They can usually run faster because they only need to work with one cable vendor and they control the whole chain. The advantage of the approach that we have is that we provide the supply chain diversity that our customers require. So when you want to go from 10,000 cables, maybe 100,000 cables to millions of cables, which is really the demand for some of these larger hyperscalers, you want to have a diversified supply chain.

And this is where our solution shines because we build the smart cable module, which is on either end of the cable for those of you not familiar but this is built to the exact specification of the hyperscaler. They know all of the components that go in there, including the DSP that comes from Astera, but also all of the power components, the EEPROMs and this and that. And more importantly, all of the security, the firmware upgrade, all of those capabilities are controlled by the hyperscaler. So now the hyperscaler does the matchmaking and says, okay, I have qualified this smart cable module from Astera, cable vendor 1, cable vendor 2, please build this cable for my application #1. And they might go to cable vendor # 3 and cable vendor # 4 for another application. So they get to have full control of their supply chain. There is no margin stacking, and they don't need to requalify the cable for every application.

So for the hyperscaler, there is a huge advantage. And as Astera, we service hyperscalers. So that's why we evolved with this model, and we feel very comfortable and confident that over time, this will be successful.

Excellent. I just want to touch quickly on CXL and sort of the memory bandwidth bottlenecks. Can you give us an update on how you see the CXL market developing? Do you still expect to ramp this application for memory expansion in 2026.

Yes. So CXL, amongst all of our products has been much slower to ramp than anything else actually. So I'll have to admit that. However, I think that 2026 is the year that we see CXL getting deployed. Towards the end of the year, Microsoft published a blog about deploying SAP HANA databases with CXL memory. We had a press release as well with the Microsoft quote. So I think this is one example where CXL really shines in general purpose compute to enable in-memory databases or large memory applications to run memory that is not bound to the CPU. So as we have said consistently, this is where the first deployment of CXL will happen. Back half of this year, we should start to see this contribute meaningfully in terms of revenues for this product family.

What I'm also excited about is some of the recent explorations that other companies are doing for using CXL in AI applications because memory is also becoming a bottleneck, especially when it comes to inference and long contacts that GPUs need to store. HBM is a very expensive memory to store those contacts in. And so some of our customers are exploring using CXL-based memory for like KV Cache applications. So again, we're not counting any revenues and such yet. But just purely from a technology standpoint, it is starting to be explored. And who knows maybe in 2027, that becomes a contributor to revenue as well.

Excellent. Just last question for me, just sort of thoughts on capital allocation and future M&A. You did the aiXscale Photonics deal. Are you looking at additional tuck-ins where you might be able to accelerate your road map?

Yes. I mean I think based upon everything that Jitendra just said, there's a tremendous amount of opportunity in the market, and this is being driven by secular trends around just higher XPU volumes, demand for AI, but also increasing complexity and this widening problem statement around connectivity and the need to drive innovation on the connectivity side. So we're seeing evidence of that through our customers and the pull from our customers to drive next-generation platforms, next-generation products. And given that opportunity, we are going to be very aggressive in the market in terms of building our team organically but then also being very thoughtful and selective around bolting on additional teams, whether it's for IP purposes or just for kind of raw horsepower to start to service some of these opportunities.

So I would say, from our perspective, strategically, we'll be looking to drive organic investment as well as M&A, mostly, I would say, from like a bolt-on perspective, acqui-hires to bolster the team and really put us in a good position to capture as much of this TAM as we can over the next several years.

Excellent. We've got maybe a minute or 2 for questions from the audience, if anyone has a question.

All right. We'll wrap here. Jitendra, Nick, thank you very much for joining us at the Needham Conference. Really appreciate it.

Thank you very much. Thank you, everyone.

Thanks.

All right. Welcome back to the Barclays Global Tech Conference. I'm Tom O'Malley, semi and semi-cap equipment analyst here. Very pleased to have Jitendra Mohan and Mike Tate from Astera Labs. Thank you for being here, guys.

Yes. Absolutely.

Thank you.

So a thematic that we've started with in a lot of these conversations is just the idea that we're in the early innings of a large AI investment cycle. Maybe talk about -- I understand that you are enabling this large trend and maybe not making decisions from a top-down level, but I'd love to kind of get your perspective on where we are in this investment cycle and then maybe just beginning with where Astera is helping these deployments actually come to life.

Yes, absolutely. So maybe just a quick show of hands, how many people drove here this morning? So we were driven ourselves. I have a Tesla. And from my garage to the parking lot, it drove all the way with 0 intervention. And so when you say -- and yet, my wife will not let me have FLT because she is afraid that the one mistake it might make will end up with me [ death ]. So I really do think that we are in the early innings of AI as these systems do need to get measurably better. And in order to do that, you're going to need a lot more compute to make these models near perfect for them to be super useful in our daily lives. And as a consumer, I can say that when I enter a search query into Google that I've been using for decades now, I'm no longer asking it a search question. I'm asking it an AI question and I expect to get an AI answer for it.

And so this is just going to become second nature to all the consumers. And today, we don't pay much for these systems, but I think we are going to get hooked, addicted and we will pay money for all of these systems. So I think there is a lot more runway here. We are truly in the early innings. And in order to make this truly successful, the amount of compute will need to go up. And as the amount of compute goes up, the level of connectivity that is required to have this compute talk to each other is also going to go up. And that is definitely a big boon for us.

Since we started Astera, we've been helping GPUs and XPUs and CPUs all connect to each other. And I would even go out and say that the connectivity has become even a bigger problem that needs to be solved to drive efficiency in these systems. So we are really looking forward to what is to come. The orders that we have from our customers don't seem to show any sign of slowdown. So if there is any talk of AI slowdown, et cetera, we are not seeing any of it in our business.

Super helpful. So I know you addressed a lot of different verticals in AI. We have cables, retimers, the areas that we spend more time on, but I kind of want to start in reverse order today because something you just mentioned where we have all of these deployments, a lot of the questions that come up today are how are we able to get these in the market. And one of the areas that comes up again and again is memory, right? And I think on the last call, really the first time since the IPO in a more significant way, you talked about CXL and a partnership that you have with Microsoft. So I think it's really useful. Maybe spend a little time talking about what you're doing in memory, in particular, and what CXL does and then maybe a little bit about the partnership that you talked about.

Yes, absolutely. So ever since our IPO, we've been talking about that Astera is all about solving the connectivity challenge and improving the efficiency of connectivity infrastructure, especially as it relates to AI. We have done that very successfully with, call it, interconnect products, which are the retimers and the switches and so on that we talk about a lot. But -- and that addresses the data bottleneck and the networking bottleneck with PCI Express, Ethernet and soon UAL, et cetera. But there is also the memory bottleneck that you point out. As these models become larger and larger, it is becoming very difficult to fit them into the HBM memory that these GPUs have. So today, the solution is you just deploy more GPUs so that you have collectively the right amount of HBM and then you talk amongst these GPUs extremely fast. And of course, we have benefited from that trend.

But as these models continue to go into trillions of parameters, I think when we went to IPO, we were talking about $1 trillion. And now I think the largest model is 3-point-something trillion, it's just insane. So there will have to be some more ways to drive efficiency into the system. And one of them is to address this memory bottleneck which CXL can do in a very unique way with both KV Cache offload. That's really a very key part of any inference workload as well as checkpointing. So those are the 2 workloads or the use cases that we see in AI applications where CXL is beneficial.

Having said that, the first implementation or first deployment of CXL will be in general purpose compute, as we have been saying for a while. And the press release that we had and the announcement from Microsoft is a very important next step in that journey where we are now deployed in the data centers in a private setting, of course, for now to accelerate database workloads for SAP HANA. And it's very easy to see the performance benefits that you get when you have more memory and you can hold the entire database in memory. And that's what we are delivering. That's what our customers are seeing, and we are very excited to see the ramp of that in the later half of next year.

And ultimately, is the goal as we move more to AI workloads, is it the physical design and the layout of the chip where you're able to expand beachfront with products? Or is it making memory that is associated with one accelerator look like memory that is a pool that all the accelerators like where is the ultimate end goal here?

Yes. So actually 2 very important points that you touched upon. I'll take the second one first, which is how does having additional memory help. So the use cases where we see in the AI context, CXL being deployed is as a second tier of memory. So everybody will have the right level of HBM because that is by far the fastest memory that is accessible to the GPU. So long as more HBM is available, people will keep using it. But that is just not enough. And so what you will start to see is a second tier of memory over CXL or even some other protocols in the future that software can make use of and the additional latency of this memory can be hidden in the AI workloads. So that's where we will see CXL play a role or second-tier memory play a role. It does require more software work, which is why this rollout has been slower. But we do see over time, this will happen.

Now the second thing that you touched upon is actually a very, very important trend about the beachfront. So as these GPUs are becoming and XPUs are becoming larger and larger, our customers are -- they don't want to put any of the I/O on a reticle size limited die. So they want to use all of that die for compute because that's really what they are paid for, and this is what you need to solve the problem. But that whole compute is useless if you cannot plug in enough data, either data or memory. So that becomes a very interesting place for us to provide I/O chiplets, which take either copper connectivity or even in the future, optical connectivity to make sure this compute, this insanely fast compute is kept occupied, again, delivering higher efficiency in the overall system. So very excited about both of these trends.

Very, very interesting. All right. Switching to another vertical, so in Taurus, we're hearing a lot in the AEC market in general today. We had a Golden Cable initiative that got launched earlier this week. Another one of your competitors does the full cable solution. You guys kind of sit in between just the single chip and the full solution. Maybe talk about your strategy there. And you don't break it out specifically by segment in your earnings calls, but you've definitely talked more positively about that trend of late. Maybe talk about where you are today from a customer perspective and kind of what you're looking for in Taurus in the next year or so.

That's a lot of questions, go to Mike.

Yes. On the revenue side, we are seeing a nice step-up in Taurus. We highlighted in the last earnings call that Taurus will be our biggest driver of growth in Q4. We're ramping into 400-gig solutions with our lead customer. And then we're also moving into 800 gig. We're working with a customer -- broadening our customer set with these designs, and we'll see those start to deploy early next year, but much more material ramps in the second half as it layers on top of the 400 gig, which we do expect to grow in 2026.

Do you think that when you get to kind of steady state in the back half of next year, there's a chance that, that 800-gig business can cross over the existing 400-gig business? Or is that something that just takes time?

We see the 400 gig continuing to grow. So it's because of that, we still see 400 gig being more significant. But as we exit the year, the turnover to 800 gig will be starting to play out.

Perfect. Okay. Let's move to the Aries platform. So in a similar vein, you guys address solutions in the 36 and the 72 both with retimers and then also with P switches and cables. That's across both your custom silicon and the [indiscernible] 36 72. Maybe let's just spend some time on the retimer business at first. Very early on with NVIDIA, that was very strong. It's the business you went public on. As you move to more system solutions, I think it's well understood you use more NVLink. You still have some content there, but it's not as big as it was historically. Can you just talk about like the cadence of the Aries retimer specifically? Because I know, one, you have a large customer that is maybe not using as much, but then you have sort of custom silicon customers and you have a refresh of PCIe that's coming. How to balance those 2 and look at that business over the next year or so?

Yes. When we went public last year in March of 2024, predominantly most of our retimer business was on the NVIDIA Hopper platform. As we made it through 2024, we saw the emergence of these XPUs. These are the hyperscalers doing their own ASICs, and that was initially for scale out, and then we started getting scale up. So we're getting retimers for both. On scale up, we get a lot of times the AEC PCI. So you get a higher ASP with the SCM models. So as you go into 2025, we've seen tremendous growth from these XPUs driving our AI growth. While at the same time, the Blackwell transition from Hopper, those revenues became more switching revenues versus -- they still have retimers, but much more heavy on the switching side.

So as we go now into 2026, we're seeing the transition from Gen 5 retimers to Gen 6, it will create another leg of growth for us. But we're -- even that being said, we still have a lot of growth on the Gen 5 on these hyperscaler deployments.

So I think -- If I just add one other thing. I think the other -- one reason why Gen 6 is still limited is just the availability of Gen 6 GPUs and XPUs. Right now, there is only one vendor, one merchant GPU that is out, and we are already ramping very significantly with them. Later on this year, we do expect the deployment of other Gen 6 capable GPUs and XPUs. So the Aries retimer business is looking good for us.

So a way that investors, I think, look at your content opportunity over time is looking at pieces of custom silicon that are using either PCIe scale-up in the back end or UAL eventually in the back end. And in the course of the last couple of earnings calls, you guys have talked about a customer set, at least in PCIe that has expanded. I believe you said at one point to low double-digit customers. So can you talk about like what type of engagement those are? Because you would imagine if you're designing some sort of scale-up network, it's a long process. It's 18 months at a minimum between ASIC design to deployment. So what are those conversations like? Maybe any help on where those customers are? Are they large guys [indiscernible] cloud? Anything you could give would be helpful.

Yes. So we mentioned, I forget last quarter or the call before that, that we are engaged with 10-plus customers on our Scorpio X family for scale-up using PCI Express or PCI Express [ lite ] protocol. The engagements are in all different stages from confirmed design-ins that customers are going to ramp in the coming years. We've talked about the Scorpio X family in initial volume this quarter and ramping in the first half of next quarter and then even more meaningfully in the second half. So very excited about that. Qualcomm actually recently since our announcement, went out publicly saying that they support PCI Express as a scale-up protocol. And the reason for that is anybody who designs their compute system and coming from a compute-centric mindset picks a load store-based memory semantic-based protocol, which is NVLink is a load store memory semantic-based protocol. PCI Express is a memory semantic-based protocol, and UALink will also be a memory semantic-based protocol.

So it's not a surprise that when people decided to pick a particular protocol for scale-up, they gravitated towards open standard, which is PCI Express. And like I said, many of them, we believe, will just move over to UALink because their software stack is already optimized for this method of communication. And when you go from PCI Express to UALink, it is just the same thing but runs a lot faster.

In terms of the engagement itself, we have ramps coming up. We have other confirmed design-ins. We also have many customers who are exploring using the PCI Express, a full swath of opportunities and engagement levels. But the level of engagement that I've seen on these products on the Scorpio family, Scorpio X in particular, is the highest that I've ever seen in the 30 years that I've been in this industry. We're very excited.

Yes. I want to go to the UAL switch, but I want to stick on the PCIe version of the switch first. So if you look at where that content is supposed to come in, you said in calendar year '26, I believe you said in the back half as well. There's a lead customer that people are very familiar with, but you're expected to kind of ramp along with that. When we went to re:Invent or heard from re:Invent that T4 was working with NVLink Fusion, I think the first reaction from a lot of investors was, oh, no, does this mean that the future of PCIe scale-up is at risk? We've had a chance to talk in between, but I'd love for you guys to have an opportunity here to talk about why does that actually benefit you guys that there are multiple SKUs versus it being a problem.

Yes. So maybe I'll say a couple of different things. First, just to go back to what you said before. In order for somebody to deploy a new scale up, they have to have made that decision 2 years ago. So the opportunity that we are currently tracking with our PCI Express, Scorpio X portfolio will continue to ramp through next year, 2026 through 2027. So any new technology, whether it's NVLink Fusion or even UALink is unlikely to become dominant until sometime after that. They will all start to show up in 2027, part here and then become dominant hopefully, the years after.

Now talking about NVLink Fusion and the different optionality that some of our customers want to have with their XPUs. I mean it is only natural to see in this kind of power and wafer constrained world that people would want to have more optionality. And the AI dynamics keep changing every 3 months. So I do want to, first of all, give you the short summary.

The Astera Labs content is growing higher with every generation of GPU and every generation of XPU. There are always calls that get made on, hey, no, no, the world is falling. People are going from Hopper to Blackwell. But as we have shown repeatedly, we are very tightly engaged with our customers, strong, strong collaboration, and we are very confident that our content goes up.

Now talking specifically about the Trainium4 announcement, they talked about 2 different types of deployments. One, let's call it a native rack, where Trainium talks its native protocol and goes directly to a switch using [indiscernible] or UALink as they announced, which is actually a very positive development for UALink. And that is going to be the highest performance, highest performance system because there is nothing else in the middle. You go directly from Trainium to the switch and lowest latency, lowest power and whatnot.

With the NVLink Fusion, though, another option opens up where now you can add another component, and this is a fairly complex component, which takes the input from an XPU, Trainium in this case, as an example, and converts that into NVLink because NVLink is not the protocol that XPUs talk natively. So this component has to do protocol translation, manage the data flows, manage security, manage a few other things and then convert it into NVLink. The advantage to the hyperscaler of using NVLink Fusion is they can use all of the work that has gone into the NVL 72 rack, which is immense. Not just the NV switches and the NVLink protocol itself, but all of the liquid cooling, power supply, the entire supply chain can be leveraged. The price that you have to pay is additional latency in this component, additional power dissipation and certainly additional cost.

So exactly how an end customer will deploy a native rack or call it this hybrid rack remains to be seen. But on the whole, our content goes up because in this system where you are doing a translation of protocol from one to the other, your attach goes on a per XPU basis. So you're taking a complex component and attaching it to every single XPU. Whereas in a switch scenario, you typically have one switch that attaches to multiple XPUs. So we are very excited about this announcement. And actually, we are more excited about the trust that both the hyperscaler customers and NVIDIA have placed in us to be able to develop this critical component.

Something that I've thought about, and I'm unsure at this point and maybe you have a good answer here is if I look at when NVLink Fusion was first announced, there were side-by-side comparisons showing either a CPU or an XPU in the scale-up architecture. And I would imagine content would scale with you guys as foreign items, either CPU or XPU come into the NVLink world. Is it possible for customers to potentially do both of those? And would that also be a content uplift for you guys at the same time?

Yes. This is changing. I mean the AI world is changing so rapidly, it's hard to keep up. This particular announcement for Trainium was to use NVLink Fusion for scale up. So this is not the link from the XPU to the CPU. And this is where our content is. When the link goes from the XP to the CPU, we don't know yet what will happen. In the past, we had a retimer content there, as you know, in the Hopper generation. In the Grace Blackwell generation, that content went away. And it really depends upon how the customer decides to architect their system. If they are willing to put the -- if they're able to put the CPU very close to the GPU, then there is no need for a retimer. But if the GPU is further away from the CPU, then you do need a single conditioning component. And of course, our retimers are an excellent choice that customers leverage.

Okay. Helpful. Back to the X switch on the UAL front. Marvell at their earnings talked about a UAL switch. The integration point for you in the market seemingly is Helios and the 500 generation. Can we talk about timing of when we're expecting to see that in market? And then also does it matter that competitors are there first? Where are you competitively?

I mean right now, there is -- nobody has a UAL switch out in the market. It's still early innings and it's greenfield deployment. As you know, Astera's philosophy is to talk about products when we have them and we can show them working at a trade show or what have you. So that's likely the strategy that we will follow for UALink as well. But as has been discussed before, UALink is a really important development for the industry and a fantastic, fantastic opportunity for Astera Labs. So we will take all of our combined knowledge that we have built by being in scale up.

See, when you are trying to build a scale-up switch, there are 2 things that are really important. Everybody talks about the protocol, whether it's NVLink or PCI or UALink or even Ethernet, which is, of course, an essential component of the switch. But there's another part, which is how do you make the switch operate seamlessly, reliably with the most amount of efficiency in a scale-up system. Keeping these complex racks up and running operationally is a huge challenge. And by being in PCI Express scale-up topology, we have learned a whole lot on what else is needed outside of the protocol. And our customers have deployed our COSMOS software in their operational stacks to configure these devices, get telemetry information and make sure that the fleet operations run smoothly. And all of that will carry over to UALink.

So we believe we are very confident in our ability to execute towards these UALink switches. We think that you will start to see some of them towards the second half of 2026 for pipe cleaning and so on, samples, et cetera. And the full qualification will happen in 2027 and then ramp ensues from after that.

Very helpful. When you look at the world, I think many of us think in 1s and 0s and black and white in terms of how deployments will occur, either the world is all going to UAL or the world is all going to ESUN, this is the end. When you look at the end state or at least what we can think of as an end state AI for the next 5 or 10 years, do you think that this is a fractured world where certain guys are using scale-up ESUN, certain guys are using UAL and it's really just whoever is married to what protocol? Or do you think that we go entirely in one direction?

Okay. So you will get many different opinions to this question. And everybody is going to talk their playbook. So I'll try -- while I'm biased, I'll try to be as objective as I can be. And the way we see it is the latter, which is it's going to be a factored world where all of these standards will actually coexist. NVLink will continue to do -- NVIDIA will continue to do what they do with NVLink. So there is going to be an NVLink ecosystem, which previously we were not a part of. But with this NVLink Fusion announcement, we are now part of that. So that's a very exciting incremental TAM. We can participate in the NVLink ecosystem, if you will.

Then there is the PCI Express ecosystem today that again, we are in full position with. And as I mentioned earlier, we do believe just because of the amount of work our customers have put in into their scale-up software that those who are using PCI Express-based systems and using load store protocols will naturally want to transition over to UALink because it's very similar. There is even a more optimized version of PCI Express and removes the line speed bottleneck that PCI Express currently has. So they will likely go from PCI Express to UALink.

Now at the same time, the folks that are currently using Ethernet and publicly only Intel has really stated that they are using Ethernet for scale up, but there are others. They have optimized their software to use RDMA over Ethernet. And they are unlikely to change as well. So I think those guys will continue to use RDMA over Ethernet. Maybe in future, they will go RDMA over ESUN, who knows. But I do think that these 3 different camps will continue to coexist out in the future.

Any plans? You're obviously doing Ethernet over cabling right now. There is a large Ethernet switching world, which is dominated by a single player today. Just given your technology stack and your engagement with hyperscalers, would you ever see your portfolio moving into an Ethernet switching platform?

I think, first of all, I would say, never say never. And you are also correct that the Ethernet side is dominated by one large player. And they make really good switches for everything else that they might do. They do make very good products. So at the end of the day, we have to look at each opportunity and see what is our role going to be? What is it that the customers are asking us to do? What differentiation can we bring from Astera and what gives us a lasting kind of sustainable advantage in that space.

Today, our customers are asking us to focus on PCI Express and transitioning that to UALink. And that's a massive opportunity for us. So that's where we are focused in. If things change, we run out of things to do or customers tell us to go into the Ethernet side, then we will approach that, but we'll approach that carefully because of the large incumbent and really decades of learning have gone into that product.

NVIDIA is talking more about [indiscernible] in next generation. The supply chain is talking late '26, early '27. Marvell just bought Celestial and is talking about the future of scale-up optics. You guys bought aiXscale, aiXscale, I believe how you pronounced it in October. It feels like this is your play in that market as well. Could you maybe talk about when you see [indiscernible] intersecting scale-up architectures? Is it something that you think proliferates from a scale-out perspective first and then gets to scale up? Anything on the timing and what that means for your business?

Yes. I apologize. Left over from COVID from a month ago. Also, it comes about when somebody asked a harder question. I don't know. So, I mean, kidding aside, we've been talking about optics now for some time. That it's actually a huge opportunity for us. Any amount of dollars that we get from a copper link, we can significantly increase that with an optical link. So if magically, everything could go optical, we would actually love it because it's a huge TAM expansion and huge opportunity for us. And it is exactly for that reason that our customers don't want to go to optics because you have to pay in power and you have to pay in dollars.

Having said that, as these systems become more complex, data rates go up, there is a desire to expand the scale-up domain from 1 rack to 2, 3, 4 racks, you have no choice but to go to optics. And we are very well engaged with our customers on what that intercept is. It is likely to be in the 2028, 2029 time frame. But also imagine you cannot go from 0 to full deployment of optics in one shot. So there is going to be some pipe cleaning deployment that happen next year in 2027, leading up to a bigger deployment in 2028.

So with this aiXscale acquisition, we feel we are very well prepared to intercept what our customers' plans are. If you look at an optical link, it makes up -- it is made up of 3 components. There is an electrical IC. And of course, electrical runs in our blood. We have no problem building electrical ICs. It does require some foundational technologies for it to be able to talk to the second component, which is photonics. So that's a photonics IC, and we have been investing in this space with analog mixed signal technology that allow us to drive a photonics IC from our electrical IC. And then there is a very important piece, which is the connector, which takes the light output from the photonics IC and couples it to fibers and then eventually gets it out.

And our belief after doing a lot of survey is that this connector is actually a really key piece and a limiter in how you can scale optics from producing millions of EICs or millions of PICs, which is relatively easy to do, but then connect fiber, et cetera, reliably. And we said that aiXscale was probably the closest to getting to high-volume production on this piece. And also, it would have been very difficult for us to do. So we went ahead and acquired them.

And on the photonics piece, we also have the capability of building our own photonics, which we will do. But we are also very open to working with other players that our customers might point us to. Just like scale up is a religion in many ways, the choice of photonics is also a very religious argument between customers. So in Astera philosophy is not to force our customers to do things our way. We would very much enable our customers to keep doing what they want to do. So if they come and say, we want you to use this photonics, couple it with your electrical IC and your packaging solution, we will be very happy to do that.

So Astera has one of the most robust growth trajectories of any company I cover, spinning off a ton of cash, new product introductions. Mike, maybe talk about as these new products come into the fold, what does that do for gross margins, one? And then two, from a capital returns perspective, I know you've been very, very clear about reinvesting in the business given all of these opportunities, but any priorities that you'd like to discuss.

Sure. We've been very clear since we went public that as we broaden our product portfolio and increase our TAM opportunities that our gross margin model is 70%. So we still believe that is where we're tracking to over the long term. And with that and the growth that we have that we can deliver 40% operating margins with that. We're overachieving on both sides of that. But as we continue to grow and mature, that is kind of the direction we're headed. As for cash flow, we're very profitable right now. We're building up a very large balance sheet. But we still think at this early stage of the company, just going public last year that we'll continue to look to potentially deploy those for strategic M&A or other strategic initiatives at this point. And as we mature and continue to develop the track record of cash flow positivity, that we'll look to ways to return to investors as well.

Feels like it's been a lot longer than a year. It's been a lot of great stuff from you guys. I appreciate you being here. Thank you so much.

Absolutely.

Thank you very much.

Well, folks, thanks for joining us. My name is Simon Leopold, semiconductor and data infrastructure analyst here at Raymond James. We're at the second day of our TMT+C Conference. It's telecommunication, media, Technology and Consumer here in New York. Thanks, everyone, for joining us.

So this morning, kicking it off with a fireside chat with Astera Lab. And we have with us Nick Aberle, who's the VP of Finance and Investor Relations. So fireside format, some Q&A that we prepared. Let's kick it off with how would you like to introduce the company to a potential investor who may be new to the story, trying to get a ramp up and understand where Astera fits into the world?

Yes. No, I think thanks for having us, Simon, and thanks, everybody, for coming and supporting us. Yes. So the company has been around now for a little over 7 years. And the way I'd like to kind of open it up is just saying that the vision of the company and the mission statement of the company is to solve data center, networking, memory and data bandwidth bottlenecks within the data center. So if you look at the portfolio that's been rolled out over the last couple of years, it's really been very much focused on relieving these bottlenecks, but not just relieving them, but actually providing a lot of proactive information and data and feedback to hyperscaler fleet managers to allow them to drive much more productive, higher utilization data centers.

So we've been very successful with that for the first couple of years here, we've been ramping. I think it's always helpful to kind of start with where we were a couple of years ago, where we are now and where we're going over the next couple of years. If you look back to 2023, the initial Aries, PCI Express Gen 5 portfolio ramped into volume production, most notably on NVIDIA-based platforms for HGX form factors, and that was really what drove a lot of the initial volume for Astera Labs. But in 2024, it expanded out beyond these kind of merchant GPU-based systems into systems that were supported and powered by custom ASIC accelerators from the likes of AWS, Google and others. So that was the next kind of foray into the market. We started selling not just the Aries, PCI Express retimer solutions, but also our Taurus Ethernet retimer solutions and modules into these platforms for -- or scale-out opportunities.

In 2025, a pretty landmark moment for the company as we introduced our Scorpio switch fabric family, which is for both scale-out connectivity and for scale-up connectivity. The ramp that we've seen thus far in 2025 has been focused on scale-out. So think connectivity between CPUs, GPUs, networking, storage and memory. And then as we move into 2026, we have a very large opportunity to ramp the X series of the Scorpio solutions, which is for scale up, and that's going to be connecting GPUs to GPUs or AI accelerators to other AI accelerators.

Beyond that, and I'm sure we'll kind of get into the longer-term vision in a minute. But obviously, everybody is moving towards a rack scale approach to building out AI and cloud infrastructure going forward. So moving from smaller systems to much longer -- much larger, much more complex units, which puts even a bigger strain on the connectivity backbone, which us and others will hope to address here over the next couple of years. But we're really going to do it from a couple of different standpoints. We're going to provide silicon modules and hardware into these boxes. We'll provide modules and hardware into these systems. We'll provide software support that allows these guys to continue to proactively monitor the systems. We're going to do it over a variety of different protocols. We've seen nice revenue ramps on the PCI Express side, ethernet, of course, but we also have some opportunities to get into some new kind of protocols as well, including UALink, which I'm sure we'll talk about shortly.

We had a big NVLink announcement last week, so we can talk about that as well. So I think the plan is here to show some broadening and some diversification beyond what the nice revenue ramp that we've shown over the last couple of years, and there's just a tremendous amount of opportunity for us to go prosecute that.

So I think you guys have coined this term AI infrastructure 2.0. And I feel like you sort of have alluded to that, but maybe sort of solidify that for the audience in the sense of sort of like what's the old architecture and what's the new architecture and what makes it 2.0?

Yes. So it's really that transition from looking at an AI server as being a unit of compute, which was a new thing. If you look back just a couple of years ago, but it doesn't take long for the industry to continue to move on and NVIDIA has been an amazing trailblazer for the space in terms of going and moving to a rack scale solution with the NVL72 based systems that we're starting to launch and ramp earlier this year. But outside of NVIDIA, I haven't seen anybody else really kind of start to build these racks at scale yet using the scale up kind of back end domains to interconnect all of these different accelerators and all these GPUs.

So for us, AI infrastructure 2.0 as a transition towards a rack scale architecture the broader industry. And like I mentioned before, bringing a lot more complexity, a lot more need and demand for high-performance connectivity solutions that cannot only manage the speeds that continue to increase and double, call it, every 1 to 1.5 years. The size of the boxes are getting bigger. So the distances in between the end points is getting longer. That puts a tremendous amount of pain on connectivity, especially when you're trying to move data very fast between these endpoints. And also just the embracing of open standards within these racks and the desire for hyperscalers and large customers to want to embrace a collaborative open ecosystem in order to drive technology velocity going forward.

So we've -- really put our arms around this open standard approach. Obviously, we have a very nice foothold on the PCI Express side, nice ramping business on the Ethernet side, but see open standards like UALink being very much utilized going forward as hyperscalers look to base their multigenerational design road maps off something that they can see and collaborate and rely on a multiple -- or a very diverse supply chain to get product from.

And so how would you describe your competitive moat? So what's the advantage in that if you think about the semiconductor space, there are so many companies that could potentially be playing and competing against you and you seemingly had a leadership position. What's sort of gotten you there? And what keeps you there?

Yes. So the initial step into the market on PCI Express Gen 5 specifically was really on the back of providing a solution that was on time that was higher performance. But I think the key differentiator was this ability to provide value-added feedback to customers in order to have them increase productivity and utilization within their systems. So when you think about the size and the scale of these data centers, the amount of boxes that are proliferating these buildings, the amount of links, I mean, just PCIe links alone that are connecting all these different endpoints. All those endpoints are coming from different vendors and configured in different ways.

There's a tremendous amount of complexity involved in that. And Aries, our flagship gold standard solution for PCIe Express retimers was -- basically allows these guys to sit and monitor each one of these millions of links that they have within their data center. So there hasn't been too many, if any, AI servers that have shipped into the market over the last, call it, 4 years that haven't had some type of Aries content within them. We have built a software framework that's based upon both firmware that sits on the chip and an API that's provided to customers that allows fleet managers to build dashboards and monitor specific attributes and feature sets that are important to them within all these different boxes that allow them to see what's happening on each different link and how to proactively maintain them before something really bad or destructive happens within these systems.

The name of the game for them, obviously, is to monetize all this equipment. There's tens of billions of dollars of market of CapEx that are being installed. And to the extent that we can even increase utilization by a couple of percent, obviously, it's more than paying for our solutions. So for us, from a competitive standpoint, that's going to be a huge differentiator going forward. All the products in portfolio going forward will support COSMOS as a software fleet management, observability infrastructure. And that's going to be something really powerful that we leverage. I feel that if you look at the installed base that we've been able to really kind of put out and entrench into the market as well, there's been a tremendous amount of teachings and knowledge that we've been able to establish and pick up through that process that somebody else, if they had the same exact chip, the same exact software would have to go recreate and have somebody give them the opportunity to proliferate widely to be battle-tested. So I feel like that's going to continue to be a big hurdle for others as well.

So maybe it would help to map some of the product because you've mentioned some of the product names. Just a little bit sort of educational as to what do they do? Because I've got the sense that Aries is sort of the anchor foundational where this company got its sort of big start and now Scorpio is sort of the cool thing that's growing fastest. But think about financial analysts, level of understanding, they need to take away. But I often sort of get these questions of prioritize these products, where is the growth coming from? What's most important? Maybe walk people through sort of the layman's version of the functions.

So I think the exciting thing about the Astera Lab story is, obviously, we've had a lot of success with the early products that have ramped into the market. We have a very broad exposure. We ship to vast majority, if not all of the hyperscalers and AI platform providers in the space. But if you look at the road map going forward and what we're trying to accomplish over the next 5 years plus, what you'll see as an increased focus on providing more value to hyperscalers.

So as you mentioned, some of these initial products, and I'll walk kind of through each of them here in a minute, is that they served a purpose that allowed hyperscalers to accomplish a certain thing, which is to monetize their data center. And that kind of got us invited back to the party to do more. So what we've seen through that building of relationships on the areas and tourist side over the last couple of years has really invited us into these conversations to say, "Hey, this is where AI infrastructure, how we build our cloud or how we build our AI infrastructure is going to look like over the next several years. We need you to solve this and that challenge". So we need to go work together to build the solutions that are going to solve those challenges of tomorrow, not what's happening today.

So I would classify -- so Aries and Taurus, we call it kind of signal conditioning solutions. So think of it as signals are moving around within a box at rapid rates, either based on a PCI Express or an Ethernet protocol. And like I mentioned before, you have a very eclectic group of different endpoints that you're ultimately trying to connect. So these signal conditioning solutions sit in the middle of these links and are able to basically take data in, repair, replicate and move it to the next area. And the reason why this is important is because the boxes are a lot bigger today than they were when you look at the general purpose compute space kind of going back 5 years ago where chips were a couple of inches apart. You could have situations now where they're multiple meters apart. So over the longer term, we would expect the signal conditioning solutions to see higher attach rates as a function of both increased distance and increased speed.

But really, the next class of solution that we're rolling into the market is these fabric solutions that do more than just connect one endpoint to another. They're really starting to be kind of called the traffic cop of these systems where they're interfacing with multiple different endpoints, all interoping with each other over a next-generation high-speed standard like PCI Express 6, which we're supporting today and believe that we're the only guys shipping in any type of meaningful volume into the market today. And having that be a high-performance solution has high-end SerDes on the front end, but then is able to handle with high port counts and high lane counts, all these different endpoints within a box. So if you look at just the value that we were kind of grabbing in the signal conditioning landscape, call it, a couple of years ago where we could get maybe $50 to, call it, $100 worth of content on a per accelerator basis, with Scorpio, at least on the P-Series, what we've seen so far is that number expand to multiple hundreds of dollars of content opportunity per accelerator within Iraq.

And as we move to scale up topologies next year with the Scorpio X-Series, you're talking more complex solutions, bigger chips, more performance chips, faster solutions, you start to see that couple of hundred dollars move into many hundreds of dollars and even approaching kind of $1,000 threshold in the kind of full grand slam if we won everything that we can provide into a box. So that's the trajectory. I'd say just we're in the business of stapling as many dollars as we can to every XPU that goes out the door. There wasn't a lot of dollars to be stapled early on, although we were very successful with our penetration. And now the next step is to kind of grow the Scorpio business into the space and get that proliferated as widely as we can.

So I feel like your stock gets knocked around with headlines on different competitive shifts among the hyperscalers and among the XPU suppliers. And I sort of sometimes scratch my head a bit because it seems like you're everywhere. So maybe it's a question of share in different types. Help folks understand sort of your exposure, whether it's Blackwell, Trainium and Trainium generations or the ASIC options in that I think you have the ability to address all, but is sort of something more important? Is something greater content? How should we think about the importance of the different architectures and options for XPUs?

Yes. So one of the big kind of step-ups in the company's business profile over the course of the last 12 months has been the proliferation of XPUs and custom ASIC accelerators into the market. And the big reason for that to be such a large catalyst for us aside from just having more XPUs to attach to or to staple dollars to is the fact that the addressable market opportunity is so much bigger in those types of environments versus an NVIDIA environment for us at the present time. Maybe that changes going forward, but we'll talk about how it's trending right now.

So the company cut its teeth in scale out, I would call head node connectivity where there's a kind of finite amount of units and demand that sit between all these different endpoints as we're trying to clean up signals and like I said before, do good observability and monitoring for the customers. But when you open up the scale-up back-end clustering application, then you're starting to see an environment that has a very rich attach rate relative to what we see on the scale-up side where you can see multiple signal conditioning solutions attached to any given XPU because they all need to talk to each other.

So the big difference, obviously, is that NVIDIA within their own environments uses a proprietary chip-to-chip protocol called NVLink, which is proprietary to them. So there's no outside vendors that could support that today. But in other places like some of the big hyperscalers, they're using open standards to drive their chip-to-chip interconnect strategy on the back end for scale up. So you see some customers today just on the verge of ramping PCI Express or customized versions of PCI Express for scale up. So it's really that difference between the merchant GPU landscape and the XPU landscape that provides us a much bigger opportunity, and it's accelerated our growth this year. And it's not to say that we don't have good content on the NVIDIA side. A lot of the Scorpio -- a vast majority of the Scorpio growth that we've seen in 2025 has been driven by NVL72 attach, Blackwell attach as we've seeing us not only be able to -- we went from shipping Aries Gen 5 solutions to shipping both Aries Gen 6 solutions and Scorpio P-Series solutions, so a multi-hundred dollar content step-up in just that environment alone.

But when you unlock that back-end scale up marketplace, the attach rates go up, the complexity of the solutions go up. And you're still getting all those opportunities on the scale-out side. You're just adding all the scale-up opportunity on above it. So we obviously are very broad in terms of how we ship out to the market. We ship to all the major hyperscalers today. But it's really on the XPU side where we see the richest content opportunity and the most dense attach rates.

So there's a section in our report with the sort of subject adding alphabet soup because there are this sort of variety of protocols that you have to deal with. And so it feels like PCI Express is kind of home base. UALink is sort of what's ramping. And I feel like the debate is around Ethernet and flavors of Ethernet and then NVLink Fusion, which, if I understand correctly, is sort of the version of NVLink that is for others. How do you think these stack up in terms of what are the market preferences and where do you have advantages or where are you threatened?

Yes. So I would say you referred to volatility earlier. I mean we do see a lot of volatility in the stock. But I mean, I think it's -- a lot of it is because there's such a tremendous opportunity out there, especially when you start to think about the scale-up domain for connectivity solutions like ours and like for others. You're talking about tens of billions of dollars of potential market opportunity that is effectively 0 today. So you're going to have this huge market landscape that's basically being born and that will ramp and grow to tremendous volume over the next, call it, 3 to 5 years. And you have folks jocking for their position to kind of put themselves in a place where they have the best opportunity to win the biggest piece of that market as possible.

And we're all including us, very tightly are working with each of these individual AI platform providers or hyperscalers or infrastructure builders to help support them and their decisions that they're going to make in terms of what path they want to go down. So as you correctly mentioned, you have a couple of different scale-up protocols that will be prevalent. I would say, today and over the next kind of year or 2 and that would be obviously NVLink and NVLink Fusion, which as the Fusion piece would be for the XPU environment. You have PCI Express or derivatives of kind of semi-custom approaches to PCI Express and then you have standard Ethernet. So those are the kind of flavors that guys get to choose from today. And I think the reasons that they choose any of these given paths could be for different reasons. No 2 clouds are the same. So they have different preferences, different focus points. If you're focused on driving low latency, you might go down that PCI Express path. If you're looking for just raw horsepower, you might go down the Ethernet path today.

But I think what we are seeing in terms of the next several years going forward, and UALink kicked this off in a big way last year was a focus on building an open standard that's purpose built for AI. If you think about PCI Express, obviously, not built for AI. This thing has been around for decades, same with standard Ethernet. So given the market opportunity, given that the dollars and the resources at play, it makes a tremendous amount of sense to kind of collaborate and bring an ecosystem together to kind of drive these technology standards going forward. So UALink will be the natural extension of PCI Express going forward. We've talked about engagement with over 10 potential customers that are looking to use PCI Express for scale up today with a vast majority of them looking at it as a longer-term step towards UALink over the longer term.

And then on the Ethernet side, you're starting to see some of the same movements, right? The announcement of ESUN back at OCP was basically the same move that UALink did a year ago, which was bringing the consortium together, let's get together and collaborate on a spec and let's get that spec out, let's make it open, and we'll have an ecosystem that can support it for diversity of supply. So it's such a big space. I think what you're going to end up seeing is different players take a different path for whatever suits their profiles the best. NVLink Fusion, if you don't have a ton of resources internally to go build your own scale up infrastructure on the back end, you want to get to market quickly, NVLink Solution could be a good solution, a good path for you as well. So I think each one of these pieces are going to have their play. We're going to try to get products in the right areas to expose ourselves to as much of the market opportunity as we can. But as of right now, just given the engagements that we see, the relationships, the multigenerational road maps that we're attaching to, we're confident that we're going to be playing with the significant -- multiple significant players here over the next 5 years.

And is the right strategy for Astera to sort of be really, really good at sort of betting on a horse to win? Or is it to support every protocol?

Yes. I mean I would say that it's listening to the customer. I mean you've listened to [ Jara and Sanjay ], the key pieces of the company's kind of core value since the beginning is to be intently focus on what the customer is asking for and understanding what that customer is trying to achieve and solve for, not just now, but going forward. So to the extent that customers are focused on going PCI Express to UAL over the next 5 years, obviously, we're going to go down that path and expend significant resources to support that move. If customers come to us and say, hey, they want to go down to NVLink Fusion path, we have a play there. We have a very tight collaboration with NVIDIA, and we can work with them very closely, the hyperscaler and NVIDIA to bring a solution to market over the longer term.

And then even Ethernet, we're not having anybody knock on our door at the moment to go build an ESUN switch, for example. But to the extent that happens, which it very well might, then we would push in that direction as well. So we're typically not building products to go chase markets and look for customers later. We're typically building products that we have a teaching customer that's going to help shepherd something to the finish line and then buy a significant amount of product at the end of the finish line.

So I want to pivot a little bit. You recently made an -- or closed an acquisition of a company. I'm going to say at aiXscale? Is that the sort of...

Yes. It's pronounced x scale. It took me a while to figure that one out as well, yes. aiXscale.

aiXscale. So it's photonics. And the reason I think this is intriguing is that one of the risk statements is when the world goes to co-packaged optics, do we need as many retimers? What does that do to your business? And so now you've made an acquisition to, I think, address that opportunity set. Can you help us understand how you think about the opportunity set, the time line, the risk statement I've made, is that a valid statement? How are you thinking about it?

Yes. No, I mean we see it as an opportunity, and I think that we've been very consistent in our messaging around optics and CPO over the course of the last, call it, 1.5 years since it's been really kind of getting a lot more attention that eventually, over time, due to increasing speeds and increasing distances that optical will have to be a solution to allow you to open up that next segment of the market.

So we've been dropping breadcrumbs over the last year around, hey, we're going to get into optical. We're spending money bringing resources in, hiring folks to focus on optical solution for the market. So I think that the aiXscale acquisition specifically is that kind of next hint of the direction that we're going, which is clearly going to be towards the scale-up application for optics and silicon photonics. Our view is it's still a little ways away. So I think there's some time here to develop and to build relationships with customers to understand what they want. But what we're hearing from those AI platform providers and infrastructure builders is that optics is still very expensive, less reliable, especially for a scale-up domain, it's going to be used only when it absolutely necessarily -- it's an absolute necessity.

So what we expect to see over the next several years is as today, we're going to be scaling up maybe 72 accelerators within racks, maybe that doubles to 2 racks for 144 accelerators going forward, which should all be able to be supported by copper. But at some point, you might want to start clustering and scaling up across several racks, 3, 4 racks. And then those types of situations when you're needing to move data more than 5, 6, 7 meters, optical is going to kind of be the only play to go down. So we're hearing from customers, multi-rack scale-up will be the application that first allows for penetration for CPO or some type of silicon photonics within the scale-up market opportunity. So we're putting together the pieces of the puzzle to go penetrate that market. So we have a lot of the technology in-house today already. There's more things that we need to build and potentially partner or buy.

But the glass coupling technology that came over with the aiXscale guys is an intriguing piece that is kind of a novel nuanced part of the puzzle that we're hearing from customers is extremely important. And for those who don't know, it's basically the technology that's going to allow the silicon photonic chiplet to connect with the fiber media on the other side to basically engage that CPO solution with the optical connectivity on the other end that this is a really tricky difficult piece to try to figure out, especially at scale when you're starting to build many, many, many units and ship and deploy many, many units.

So this is one where we didn't feel like we had the expertise in-house to go figure that out on our own. We found a great team in Germany that's going to help us do that, that's already started to kind of show that they could do it at scale. They have good IP that supports it. But a lot more to come here. We're going to have more announcements, and we're clearly going to kind of go down that path to try to get into this market in a big way when that starts to ramp.

I like it. I refer to optics as the one thing I actually understand. So it makes sort of brings it closer to home to me. But when I think about this acquisition, I sort of can envision a couple of other paths. Do you look at it as sort of a feature technology that gets incorporated into the Scorpio family? Or do you look at this as a building block to future products? How should we think about the use of it?

Yes. So I would say that we're in a good position today to be a player in this market over the long term because we have the critical piece of the puzzle, which is the switch. And today, you're basically going to scale up and connect that switch with copper endpoints, copper traces, copper connectivity, copper links. But over the long term, Scorpio has an opportunity to kind of cross over into the optical domain as well.

So optically enabling Scorpio is a key focus for us, taking advantage of that kind of anchor socket and pull position with kind of a best-in-class switching solution is going to be the heart of the product, obviously, and then putting the pieces in and around it in terms of the silicon photonic chiplet, the coupling -- fiber coupling technology from Xscale to get a full system-level solution is a key piece of that as well. So it's taking what we have today, which is the anchor socket, and this is one of the building blocks like you mentioned, to kind of build out a grander solution over time.

And then Marvell recently did this acquisition, Celestial AI. It sounds similar, but different. Do you look at this as competition, validation? How do you sort of think about what's going on across the industry?

Yes, probably a little bit of both. Again, we talked about the significance of the scale-up connectivity opportunity kind of earlier in the chat kind of being tens of billions of dollars of land grab for folks like us and others to go grab as this market starts to ramp up. And then to the extent that optical starts to play and expand that market opportunity even further, there's going to be a tremendous kind of opportunity for everybody.

So clearly, those guys are making the strategic moves that they would think that they need to put into place in order to capture that market. We're going at it in a little different way. But at the end of the day, I do believe that it's a bit of validation and ultimately, it will be some healthy competition that plays out here. Again, big markets, big customers to go capture. And we're all kind of racing to get to the finish line there.

So you know how they say time flies when you're having fun? Yes. Our time is up. I want to ask one question to close out, which is what do you consider either the least appreciated or most misunderstood aspect of the Astera story?

I mean I think just as a function of being kind of a new public company, it's taken a little time for folks to get to know us, and we've tried to establish a track record and relationship with investors. And we feel like we've been doing a good job delivering results and trying to be conservative and move the ball forward. But I think the biggest thing, and hopefully, this has started to show itself over the last couple of years and even more going forward is that we're really here to build a large iconic company over the long term. And the way that we're going to do that is to provide even more and more value to customers.

So when we were thinking about how to build out our road map, how to support customers, where to dedicate and allocate resources. It's really with our arms and hands tightly coupled with our customers try to figure out how to build this thing, not just for the next year or 2, but for the next 5 to 10. And so I think a lot of the steps you'll see from us in terms of new products, new solutions, new announcements will be very much tied to how do we put ourselves in a position to be an iconic company over the next 5-plus years.

Great. Nick, thanks for joining us. Thanks, forlks, for joining us. This is Astera Labs.

Thanks.

Thanks, everyone, for joining. I want to welcome everyone to RBC's TIMT Conference. I'm Srini Pajjuri, I cover semis. Today, we're pleased to have Astera Labs. We have Nick Aberle, who heads the IR team there.

So Nick, thanks for joining us. We have roughly 30 minutes. We're going to leave a few minutes for Q&A at the end as well.

So Nick, well, it's been a little more than a year since you guys went public, very successful IPR. Maybe we can start with when you went public, I think you were targeting a high-growth market, roughly, I don't know, $10 billion-ish TAM. Things have changed quite a bit since the market has broaden and you expanded into new markets as well. So how are you thinking about in terms of the addressable market, what's -- how to think about your SAM and TAM and growth potential?

Yes. Got it. Yes. Well, thanks for inviting us as well. Srini, I appreciate it, and thanks for all the investor support.

Yes, it's almost been almost 2 years. It will be 2 years in March. So at the time has flown by for sure. And a lot has changed, yes. So the initial market opportunity that we outlined around the IPO was around $7 billion, and that's grown substantially just over the course of the last, call it, 18 months or so. And the market opportunity for us is really a function of 2 big pieces. I mean, there's clearly a secular trend and tailwind that's at our back associated with just the build-out of cloud and AI infrastructure and new accelerators coming on the market.

And obviously, the goal of Astera is to staple as many dollars as we can to each one of those accelerators shipping into the market. But we've been able to kind of further enhance the market opportunity on a company-specific basis, by bringing out new products, developing new technologies, expanding the road map and just broadening the overall portfolio. So if you look at where we stand today, there's the initial pieces of the business that we outlined during the IPO, the signal conditioning, pieces of the business, which entail the retimer opportunities, both for PCI Express and for Ethernet. We've outlined is about a $3 billion TAM or served addressable market because we serve that today.

And both of those are poised to grow quite nicely over the next couple of years, both pieces of the business are going to be kind of graduating to the next level, going to PCI Express Gen6 for the PCIe side, ultimately, Gen7 as well, where the speeds will double again. And then same thing on the Ethernet side of the house as well. We're primarily shipping 50 gig per lane solutions today that will expand the 100-gig per lane starting next year for us and then obviously, 200 gig per land beyond that. So some pretty big drivers in that business where you see adoption especially on the Ethernet side, fairly limited today in corner cases, but much more adoption and attach expected as we move to these faster speeds.

Leo, obviously, a big market opportunity as well. That business is still small for us today from a revenue standpoint. But I think we put out a press release last night talking about new engagement with Microsoft there. So hopefully, that those will start to drop some proof points that something starting to brew on that front. When you start to think about the market opportunity there in the entire CPU landscape of, ultimately, every CPU is CXL capable how many of those we can enable that market opportunity starts to be quite large as well. But it was at OCP last year where we launched the Scorpio product line, which boosted the TAM even further, made up of a couple of different pieces, and we'll talk about each of those pieces individually in a little bit.

But the P-Series, we're calling it on or about $2.5 billion of incremental opportunity for head note connectivity. And then the X-Series, which is for scale-up clustering of accelerators we called that at the time at being around $2.5 billion to $3 billion market opportunity. And then most recently, talking about UAL being incremental on top of those kind of PCI Express or PC Express like protocols that X-Series addresses adding another couple of billion dollars on top. So yes, all folded in kind of looking at the $14 billion, $15 billion served addressable market today. Clearly, we're not sitting still either. We're developing and pushing forward, seeing some of the announcements on the optical side. So we would expect to continue to grow that TAM going forward.

That's great. So your largest business, the retimers, the chip that sits between CPU and GPU and sometimes CPU storage, et cetera. There were some concerns, I guess, a lot of concerns as the market moved from hopper generation to Blackwell, and that still continues to grow to this day, I think last quarter, it was up for this year looks like it's going to be up like 60% to 70%. And your expectation is for that business to continue to grow into next year.

That's correct. Yes. So Aries is still a big -- very big piece of our business. Yes. So if you look back 4, 6 quarters ago, it was over 90% of the total revenue. Today, it's closer to 2/3. And all the while, it's grown, yes, we'll grow close to 70% this year on the Aries business. So quite healthy growth, but also diversified kind of the business across the other product lines, just because they're growing very fast as well. So yes, I mean, there's a little bit of skepticism coming into the year about the transition to black wallet NVIDIA. Obviously, we had a big presence there with Hopper and you started to see some of the reference platforms that NVIDIA goes to market with on the rack scale side, use less retimer content, but we've also picked up a meaningful amount of retimer content in other areas.

We've ramped aggressively on the custom ASIC accelerator market side. And we've also started to penetrate the scale-up opportunity as well for PCI Express retimers. And as these racks grow and clusters grow in size, distances become more challenging for these PC Express signals to traverse with high fidelity. So you start to need PCA Express retimers either chip down on a board or at the end of AEC cables. So that's been another big driver of our business. So yes, I mean, to your point, we expect that business to continue growing into next year. PCI Xpress Gen 6 transition is just underway. Blackwell is really the only endpoint today that supports Gen6.

So we'll see a whole slew of incremental accelerators to start to adopt PCI Express Gen 6 starting next year going into '27. That will drive higher attach for retimers but also a higher ASP for us as well. I think on or about 20%, plus or minus, ASP lift on a generation-over-generation basis. So overall, very excited about the Aries business and the opportunity there.

That's great. I think this market took off of Gen AI with the Hopper generation. And then obviously, they're attracted some big competitors, right? I think Broadcom announced some products and Marvell announced some products. It looks like you have that visibility into next year in terms of your design win momentum also you're talking about ASP premium going forward?

Yes, that's exactly right. Yes. So I mean, just like any other big and growing market, you would expect competition, so not super surprising from that standpoint. And we respect all the guys that are coming into space. They're very well-known, established names in the networking arena. But yes, I mean, we have a substantial foothold in this business today. We're in the incumbent position. We're the only guy shipping PCI Express Gen6 solutions into the market today. We talked about PCI Express Gen6 making up a little over 20% of our business in total in Q3. And then to our knowledge, nobody else is shipping anything in volume today. So the head start is there. The learnings and kind of the gauntlet that we've gone through on Gen5 has built a lot of muscle for the organization in terms of understanding the nuances and the iterations that need to be made to kind of handle all these different workloads and interop between a wide variety of endpoints. So we feel very good about our position there.

Got it. So should we kind of think of retimer growth going forward kind of primarily associated with the ASIC side of the things? Or do you think there is actually more growth left in the GPU market as well?

Well, I think it's going to be both. So we've seen -- so the Gen 6 business that we're shipping today is shipping into NVL72-based platform. So that is a growing business. We would expect that to continue to expand next year. But to your point, I mean, I would say that the XPU side or the custom ASIC accelerator side is probably a little bit more fertile market opportunity for us because we can play on that scale outside in the peripheral connectivity side, in addition to the scale-up clustering side as well. So it's just a bigger dollar content per accelerate opportunity.

Got it. And then in terms of your visibility, I think the systems are designed well in advance, right, like, I don't know, 18 to 24 months. Is that the sort of visibility that you have in terms of design win moment?