Good afternoon, everybody. Let's get started with the next fireside chat. We're very happy to have two executives from Astera Labs, Sanjay Gajendra, who's the COO and co-founder of the company, as well as Nick Aberle, who's the VP of Finance and runs IR. Gentlemen, thank you so much for joining us. You guys have been an AI connectivity pure play, relatively recent IPO. It's been an exciting year, year and a half for you. Your stock price has gone up nearly 5x over that time. Congratulations on that. I think you're uniquely suited to see what's happening in AI, given all the connectivity protocols you address and the customer base, etc. I wanted to start with a high-level question, and that was simply, where do you see we are in the AI cycle? Are you seeing spending changing, accelerating, decelerating, diversifying?
At the highest of levels, where do you think we are?
Yeah, so let me kick it off. First of all, thank you for having us here. It's wonderful to be here and talk to you and the rest of the folks here. Going to your question again, we still fundamentally believe we are in that early innings of AI. I think it's based on the number of RFQs and opportunities that we are seeing. It's not slowing down anytime soon. In fact, it just continues to get bigger and bigger. The things that are playing into this, of course, is the hyperscalers. You know, their spend continues to grow. There has definitely been no slowdown in that.
The second thing that, of course, we see is there are increased investments that happen on, let's say, the internal or custom accelerators, largely because I think companies now understand where the money is with AI, what the workload requirements are that they need to optimize from a TCO standpoint. There is that realization. It doesn't mean that they don't need the general purpose GPUs from NVIDIA or AMD and other providers. It just means that there's also an understanding of what is needed. There is, let's say, a little bit more thought and optimization that's happening on those kinds of decisions. In general, I think the notion of having a robust supply chain and multi-vendor approach to addressing various aspects, whether it's on the GPU side or on the connectivity side, definitely is a change that we're seeing.
Our vision is that, you know, three to five years from now, having a rack, AI rack that is open and scalable with a multi-vendor supply chain is going to be a key requirement, just given how broadly AI will continue to be deployed, both for training workloads as well as on the inference side.
You're not seeing any slowing in the spending, and if anything, some acceleration. Are you seeing any changes? Is the size of the processor counts in clusters continuing to increase? If so, how do we think about your connectivity content scaling with that?
Absolutely. There's no slowing down, right? You went from, let's say, an 8 GPU cluster in the Hopper generation to now 72. For Rubin and others, if you use that as a benchmark, you can see that scaling quickly to 288 and beyond. I don't believe it's going to slow down, just given how the model complexity and model sizes continue to grow. In terms of connectivity, it's a fantastic place to be in because, end of the day, the GPU does what it does, or the accelerator does what it does. As the cluster sizes grow, the amount of data movement increases tremendously. As you probably know, in a GPU cluster, every GPU needs to talk to every other GPU. Even if it's a 72 GPU cluster, from a software standpoint, it needs to appear as one GPU sharing the same memory.
The way you achieve that is by creating a connectivity fabric that's super fast, super reliable, and is able to ensure that, from a software standpoint, things are happening in a way, coherency and things like that, that everything looks like one. That magic happens through connectivity. Connectivity truly is a bottleneck today. If you think about it, there's a lot of innovation, a lot of products and technology to be built. As some of you guys might have heard, even though people pay thousands of dollars for NVIDIA GPU, roughly half the time it's actually sitting, waiting, idling for data or memory. That simply means that the problem to solve is connectivity. With these cluster sizes growing, that problem is not going to get simpler. It's only going to get more complex. That's the opportunity we see in Astera Labs.
Why don't I pivot somewhat temporarily to just a little nearer term dynamic? You mentioned NVIDIA and some of the stuff about shipping to China, not shipping to China. Does that matter to you guys? Is that a little bit of a headwind if the processor companies can't ship the merchant folks? If they can start, does that become a tailwind for you, or is it just around the edges and it doesn't really matter?
I want to say, again, China is a significant market. I think they do see there is a lot of data out there. There's a lot of momentum around AI workloads and all that stuff. It's also a slightly, you know, different set of dynamics given the geopolitical overhang that we have. For us, yes, we do see some single-digit revenue we make out of China. At the same time, we are not like banking on it to the point that our strategy and the models we're putting together are skewed to it, just given, you know, you never know what could happen. For us, I want to say if something good happens, meaning geopolitically speaking and things open up, we will always see that as an upside versus planning that in our base models. That's how we've been approaching it. Let's see how things work out.
The one good news is that, you know, at least so far, connectivity is not as much under the microscope, you know, when you compare it to the processor and technologies related to that.
Right. It's more secondary, but still important. I want to get into some of the diversification dynamics. We'll get into a little bit of the customer and processor and market side of things. I guess first on the hyperscaler side of things, I think in the last year you've gone from three 10%+ customers to now, I think in the last quarter you had five or six of those. What's leading to that diversification?
Yeah, so what's leading is really a dynamic within the market in the sense that, you know, the cloud and hyperscaler customers, it is a limited set of customers that truly matter, right? When you have customers ramping, you can immediately see that in how the share ships. That's one thing. The second thing is that the appetite for new technology varies between hyperscaler to hyperscaler. There are some hyperscalers that are always hungry for new technology. As soon as it's available, they want to deploy that in their fleets. That's what you see. You have certain customers, predominantly, of course, NVIDIA and AWS and all these guys, they are the first ones to deploy once the technology comes forward.
There is an element of timing, meaning as you think about with time you'll see more people come in and there is always a distribution of, let's say, 12- 24 months difference between the first one to the next set of folks that follow. The third factor I would say is the business model. If you think about a hyperscaler like Meta, they're mostly running internal workloads: Instagram, WhatsApp, Facebook, and so on. Compared to someone like Amazon or Azure from Microsoft, they're running your traditional cloud services, meaning rental models. They don't know who is going to require what, whether they need an NVIDIA instance or an AMD instance and so on.
The point I'm trying to make is that depending on the number of customers, the limited market, the aptitude for technology, as well as the business model that the hyperscalers are driving, those things will have a factor in how our market share changes. In some ways it's an occupational hazard if you think about the hyperscaler data center market.
Yeah, it's naturally concentrated no matter what you do, and you wouldn't be winning if it wasn't. What do you see as the most of the key design wins happen at the hyperscalers for you or as part of reference designs with various processor makers?
It's a combination, I want to say. Increasingly what happens is that, you know, since many of the hyperscalers are doing their own accelerators, for us, that is an opportunity where for the non-NVIDIA ecosystem, we get to play both on the front-end network. Think of scale-out topologies, GPUs connecting to CPUs and all that. More importantly, we get to play on the back-end network where GPUs talk to each other, right? This is the scale-up portion of the network. In general, given our focus right now with Scorpio X- Series, which is a fabric device that is used for interconnecting GPUs and so on, there is a lot of momentum, a lot of engagements we're having directly with the hyperscalers. When you think about reference designs, whether it's on a GPU platform or CPU platform, it's definitely a channel and an avenue for us.
At the same time, I want to say the engagements have to happen with the hyperscalers so we can drive the rest of the decisions. You have the dominant merchant silicon vendors like NVIDIA and AMD that have been great partners with us, and we continue to be on their platforms.
On the processor side of things, late last year and earlier this year, you talked a lot about diversifying into the custom architectures away from, not away from, but in addition to the merchant side of things. How do you view the opportunity on the custom side versus the merchant side going forward? Is it a big battle between the two within your business, or can they both grow?
I think the short answer is both can grow because there is space for everyone. The class of GPUs that folks like NVIDIA are doing are in a completely different category. These are general purpose GPUs that can be used to run a variety of different workloads, right? Whereas you have custom accelerators that are being designed to solve a unique class of products. I don't think any custom accelerator is being designed to go head-to-head with NVIDIA. You will have space for both depending on the kind of workloads and also the kind of business model that a hyperscaler is driving. For us, like I noted a little bit earlier, we like the non-NVIDIA ecosystem from an attached rate standpoint because we get to play both on the front end and on the back end, right? That's why we see much more larger content.
If you think about our growth itself, we have been working towards this five-step strategy that we have laid out where the first step of our growth, if you think from the IPO timeframe, was based on the NVIDIA Hopper generation. We were on all the UBB and HGX platforms that they were shipping, right? In Q3 of last year, we transitioned to, like you noted, custom accelerators. Our revenue went up one step up starting the third quarter of last year. Now we have entered the third step in our journey. We've started with Q2, where we're not only selling retimer class devices, we're selling the switch devices for the Blackwell-based platforms. Our dollar content went up from a little less than $100 per accelerator to now multiple hundreds of dollars per accelerator from an attached rate standpoint for the design wins we have.
Step four that we're working towards is to further diversify our portfolio with what we call Scorpio X- Series. These are fabric devices for GPU-to-GPU interconnect. The Scorpio P Series, which I noted as step three, is for head node connectivity for GPU, CPU, and so on. X- Series puts us in a completely different position from a socket standpoint because the scale-up switch is one of the anchor sockets, one of the early decision factors that hyperscalers make. We have several design wins with design wins starting to ramp later this year to high volume production in 2026. We have over 10 customers that we have engaged on the scale-up fabric side.
Of course, the step five for us is the vision that we're driving for a complete rack-level connectivity solution that is based on UALink, which is a standard being defined and developed for scale-up network in an open standard format that will go into production in 2027. If you think about our story, it's one in which we are adding more and more product lines to service the same application with products designed from the ground up for servicing the AI workloads. We are expanding the customer base that we're working with, both from a traditional, let's say, hyperscaler type of customers to other tier two, tier three players that are essentially starting to deploy large clouds based on the new cloud model and things like that. In summary, I think we are well set. It's just a matter of ensuring we continue to execute.
Lots of ways I could go with that answer, but I'll pick the ASIC side to wrap up on that, and then we can get into UALink and some of those other protocols. Is the number of ASIC vendors out there that can be potential customers of yours continuing to grow? I know it's not a zero-sum game versus the merchant side of the equation, but it seems like that opportunity in and of itself is diversifying and growing really rapidly. Is that true?
Absolutely. Like I noted, I think the hyperscalers have realized that they've realized where the money is with their AI workflows, and they want to build, you know, accelerators to address that and optimize to that. The second thing is, no one wants to be held in a, you know, restricted in some ways of how they do their business. I mean, today, if you think about it, there are a couple of ways in which they're sort of in a jail, right? There is the vertical integration that NVIDIA is pursuing, and then you have folks like Broadcom that are trying to provide a certain type of solution that combines with their custom ASICs and networking and all that. We truly believe that there is a third way of achieving this, which is to provide an open rack, open standard that services the nuances of various different use cases.
That's clearly where we see the momentum right now, and that's what we're focusing on.
Do you believe the ASIC vendors who just happen to also have big networking divisions within their company, Broadcom, Marvell , to name two, do they have a, is it more challenging for you because they have those networking divisions that would compete with you? Do they have a bundling advantage?
It's easy to imagine that way, but if you think about it, even like, you know, there are several examples you can think of where, for example, let's say Google, you know, they use custom ASIC services for the TPUs, but they do their own networking and so on. I think it's probably a little bit of a, I guess, simplified way of thinking to imagine that somehow you can bundle these two things. Because again, what you need on the custom accelerator, the requirements of it, the feature set of it, even though it could be designed, you know, manufactured on the back end by someone, the design of that is done by hyperscalers, right? Hyperscalers also don't want to be held hostage, like you noted with some of the solutions.
What's happening is it's having the reverse effect where we're getting invited more into these parties so that they could get the best out of both worlds, right? Meaning we are developing technology that is truly differentiated, truly developed for like for AI type of workload. Then in combination with the accelerators that the hyperscalers are designing, they're able to achieve the technical performance or superiority commercially and contractually and pricing and other leverage that they want. They're able to gain that as well. I think in general, the hyperscalers are not foolish. They don't want to be held in a jail. They're reacting in a way that you would expect them to do, which is to try to get the best of both the technical world as well as from a commercial standpoint.
Why don't we pivot into some of the product lines and then some of the topics around that? Let's just start with Scorpio. I know it's not your biggest product line, but it's growing really rapidly right now, and it ties into the scale-up, out dynamic that you talked about. Why don't we start with the scale-up side of things? Just talk about the various protocols. You guys are big supporters of the UALink. Obviously, you have NVLink and then, you know, Broadcom's doing its way with the SUE side of things. Talk about the advantage of UALink and why you think that's the right way to go.
Yeah, so before I talk about UALink, let's look at where things are today, right? Scale-up networking is a greenfield use case. NVIDIA, of course, has got NVLink, which with the NVL72 now you start seeing NV switches, right? In the past, if you go back to Hopper generation and so on, with eight GPUs, you didn't need a switch unless you're trying to build a bigger cluster of some kind. In general, if you go forward, most of the deployments will require GPUs to be clustered in bigger numbers, and therefore there is a need for a switch. Today there is NVLink, which is proprietary, and most of the customers are using PCIe or PCIe-like standard for interconnecting their accelerators or GPU.
There are some that have Ethernet, for example, Intel's Gaudi Habana chip that used Ethernet as a scale-up, but that's a small number compared to the ones that are using PCIe or PCIe-like. Why are people using PCIe? Because it's a very lightweight protocol. There are years of software that's built for that. It's relatively easy from an ecosystem standpoint, like in terms of IP, in terms of, you know, birds and scopes and the test equipment that are required, cable connectors and so on. The point I'm trying to make is the ecosystem exists. It's an open standard. Now going forward, if you think about why would someone not want to continue to use PCIe, it comes down to a simple challenge where the GPUs are trying to talk to each other at a really high speed to appear as one giant GPU.
PCIe is still running at 64 gig for Gen 6. Yes, it'll get to 128 gig for Gen 7, but NVLink is already at 200 gig. The challenge that the industry has is how do I get an open standard that is lightweight, that's, you know, suitable for scale-up topologies, yet that can run faster. That is where UALink comes into play. UALink, even though it's a new standard, it's truly not a new standard from the ground-up standpoint. What they did was something we believe is very smart, which is to take the PCIe transport layer and take the physical layer from Ethernet, which already is at 200 gig, right, and combine those two things. You get the simplicity that you get from PCIe and you get the speed that's available through Ethernet standard. That is what is UALink. The consortium came together in October of last year.
By, was it March, April? The first version of the spec was already released. It took less than six months to develop that spec simply because it's not trying to do everything from the ground up. People are familiar with PCIe or PCIe-like software interface, which you can imagine is a very important consideration for deploying any new standards. We love that approach because it's solving what is needed from a technical standpoint. Memory semantics are all natively supported. Software support is much more seamless, yet you get the speed that you need from a 200 gig or 400 gig, whatever is the spec, right? Most importantly, it's an open standard. There are multiple IP vendors, there are multiple test equipment vendors, cable connector vendors, and product vendors. Eventually, if you think about the history of connectivity standards, it's the open standards that have prevailed.
Go back to USB, PCI Express, Ethernet, and so on. That's why we are putting our weight behind it and several other hyperscalers and GPU providers are jumping on it. Now, to compare it with SUE, the scale-up Ethernet that is being promoted, you've got to look at it this way, right? First of all, it is not your standard Ethernet. I think somewhere there is clever marketing going on to attach the word Ethernet to make it sound like it's open. The history there is, if you think about Ethernet, when it was designed, it's designed for transporting data across multiple data centers and around the globe, right? Robustness is the name of the game for Ethernet. When you're doing scale-up, it's too much.
There's too much overhead, there is too much latency, there are other factors that come into play, which at least in this particular case, of course, Broadcom understands it. They said, let's try to redefine the standard with smaller frame sizes, packet sizes, and have memory semantics and all that. They took it to USC, which is the Ultra Ethernet Consortium, and they quickly realized it's no longer the native Ethernet anymore or standard Ethernet. It was a proprietary way. They didn't accept it. Then it, of course, got released as a standard, as an independent standard. The point I'm trying to make is that it is proprietary in all practical purpose. There are limited vendors, which is right now just one.
It doesn't quite change the problem that hyperscalers are trying to solve, which is how do I get out of the, you know, the NVLink and NVIDIA lock that they have to something that is available in a broader sense. Long answer, but that's kind of how we see the dynamic playing. I think 2027, when UALink gets out, you'll see why some of this is going to be meaningful for Astera and for the industry. You will see that there are increasingly more hyperscalers that will come out publicly and start talking about how UALink has become their roadmap for GPU-to-GPU interconnect.
Perfect. Long answer, but a good answer. It's important to understand those protocols. If we bring it back to the Astera level and think about Scorpio, that's gone from basically zero revenues last year to, I think, low double-digit percentage of sales this year. At least in my estimate, it's going to more than double again next year. Talk a little bit about the ramp we've seen thus far in the P- Series and when the X kicks in and maybe just frame kind of the content increase relative to, say, a retimer that the company was built on originally.
Yeah, like I noted, we are in our step three where we are shipping the P- Series, the PCI switch for head node connectivity. We started in Q2 for the Blackwell-based platform. The Scorpio X- Series is something we're super excited about. It's going to completely transform the company because both from a business standpoint and also these are anchor sockets. Meaning if you are a system engineer at a hyperscaler, there are two devices you will care about the first five minutes of starting a new design. One is the GPU, the other one is the scale-up switch, right? Everything else is put around it. For us, Scorpio X- Series enables that where we are part of the conversation literally from the first meeting.
We are having those conversations right now where we're looking at not just the immediate generation, but for the next two or three generations with our hyperscaler customers. It is going to be an important device or product category to provide that anchor around which we will build our business. In terms of attached rate and dollar content, like you noted, with Aries retimers, we were, let's say, a little less than $100 per accelerator on a given design win. We have graduated now to multiple hundreds of dollars per accelerator with the P- Series switch. With the X- Series switch, we expect that to transform to multiple thousands of dollars. The path is going to be pretty, pretty significant if you think about just the attached content and the dollars that we will expect to see.
It's a big market, so we are pretty excited about what could happen.
Is the timing on that going to be limited, the X I'm talking about, by the UALink, or are there reasons that you could actually ramp that earlier?
Are we going to ramp it earlier? Keep in mind today UALink is still a 2027 story, but scale-up is happening now, and it's happening on protocols like PCI Express. I don't want to say it's native PCI Express. There are some extensions being done on that. Like we've communicated, we expect our preliminary ramp to happen end of this year with a full ramp that we expect in 2026 across multiple platforms. In 2027, we graduate to UALink, which now goes to multiple customers and multiple use cases. That opens up additional opportunity for us.
Gotcha. In the last five or six minutes left, we're going to go a little more kind of lightning round. It was very important to me, at least, to get into the scale-up side of things and the Scorpio side of things. These questions will be a little more pointed. Aries, your retimer business, what do you think the sustainable growth rate of that business is?
Speeds are increasing, which means that you run into challenges of physics, and you need to solve that by putting retimers. These are devices that are, think of them as signal amplifiers. They extend the reach of signal. As long as speeds are increasing, you will always have an opportunity for retimers, and that'll continue to grow. The nuance there is that since it's trying to solve the problem of physics, sometimes you'll need it. Sometimes you'll need, you know, two of them. Sometimes you'll not need at all. Case in point, Hopper generation, the baseboard had eight GPUs, so the board was big, right? You saw retimers all around. With the GB200, GB300, just two GPUs, the board became small. Guess what? The reach problem is not as pronounced, so you don't need retimer on NVIDIA's GB200 board. You need it on somewhere else in the system.
The trick about the retimer business is that it's not always the same, but fundamentally, as far as speeds are growing, you can keep expecting the retimer socket and use cases to continue to grow.
How do you see Astera Labs' market share going from Gen 5, Gen 6, etc., because you've been so dominant so far?
We believe it'll continue to remain the same and continue to expand in the sense that if you think about it, we are the only ones, you know, everyone was worried about what will happen in Gen 6. We announced our product, we went to production, there are several others that announced, but as far as we can see, we don't see anyone going to production or being designed with an activity. The key thing to keep in mind is it's not just about the chip, it's about all the tribal knowledge we have built to make it work. More importantly, we have this COSMOS software, which provides a certain capability for fleet management and so on. That is portable from generation to generation. That provides a stickiness for us to ensure that we continue to retain the sockets we have.
Is COSMOS beneficial across the product lines, not just within the Aries?
Absolutely. COSMOS is such an integral part because the same COSMOS API fits into all of our product families. If a customer is using our SDK or APIs for talking to Aries retimers, they can use that for Taurus, they can use that for Scorpio. Of course, they keep getting additional functionality because each device brings in something more. Once they design in the infrastructure, it's just a matter of keep reusing that across generations and across product lines.
Similar sort of question on the Taurus side of things, the AEC market. What's the sustainable growth rate of that? What's the content trajectory that you see?
Yeah, so with the AEC market, like we always say, it's a case-by-case situation. Not every customer uses AEC. It's not like you can say because customer A used it, the customer B might decide to use optical or passive cables or what have you. For 400 gig, there were one or two large sockets available in the industry, right? That was the focus for us. 800 gig, it becomes more broad-based. That's where we are expanding to right now. Most of the ramp is happening in 2026 for 800 gig Ethernet. We'll continue to expand there. The nuance that I want to keep highlighting is that we compete with Credo in that space. They do the whole cable. We do the paddle cards that go inside the cable and enable traditional cable vendors, Molex, TE, Amphenol, and others. Our business model is such that we are planning for scale.
As the volume picks up and you need the capacity and multi-vendor supply chain, we go attack there. It's not like we will be on the first six months, but as soon as the volume starts hitting, we'll start essentially gaining share and addressing that particular socket. It's a slightly different way simply because we don't believe in carrying the inventory and all the headache that comes with cable type of products. That's not our expertise. Our expertise is to develop silicon and software and all that. We're trying to stay true to what our strengths are and not trying to be a manufacturing house for cables. We'll leave it to the experts.
In the minute and a half we have left, I think that's two final questions. I'll ask them both at once. That way you can allocate whatever amount of time you have to it. Leo, yes, the CXL side of things, it seems like it's taking longer to ramp than most people thought. More of a protocol issue, not so much your fault. When do you think that ramps? The last question would be, people were very excited when Scorpio came out, that fabric switch side of things, opened up $5 billion a TAM. You guys are spending a lot on OpEx. You're hiring very rapidly. Should we expect to see additional product lines beyond these four, if I remember right, being introduced?
Yeah, I think the Leo question is straightforward. Like you said, we have the design wins. It's a matter of ramping for several reasons that are outside of our control. The most important thing is Leo goes for CPU attach. It's not a GPU story. You know, everyone wants to talk about GPUs and AI right now. That's where the dollars are. We do expect Leo to eventually ramp in the 2026 timeframe. It's not a lost cause. It's just a matter of timing. In terms of our products, clearly, you know, we have the front row seat with hyperscalers and AI platform providers. We've been working on several new products. We talk about four, but you can be assured there are many, many more we're working on. We don't believe in talking about it till it goes to production. I would say wait for it.
It will come, and it'll come in a very significant way. I think for us, the vision is to deliver the entire AI rack except for the compute nodes. That's what we're working towards, and we will continue to pursue that.
Perfect. We are exactly on time. Sanjay and Nick, thank you very much for joining us on stage.
Great. Thank you so much.