Good morning, everybody. For our next session, we have Sanjay Gajendra, the President and COO, and Nick Aberle, the Vice President of Finance and Investor Relations with Astera Labs. I should mention I've got plenty of questions here, but if anybody else has any questions, there are QR codes on the table, and if you just scan that and enter your question, it'll send it on up to me, and we'll ask if there's available time. OK, so with that, let's get started. Sanjay, can you just give us a brief history or overview of Astera Labs to kick us off?
Yeah, absolutely. First of all, thank you for taking the time, and good to see all of you guys. So Astera Labs was started in 2017. We are essentially the company was started on two basic premises or theses. One was based on the fact that the AI infrastructure that needs to be developed and deployed was transitioning from what used to be a traditional CPU-based architecture to more of a heterogeneous compute with different kinds of CPUs, GPUs that are interconnected. So we saw that as an opportunity for developing connectivity infrastructure, or the nervous system, if you will, that interconnects the CPU, GPUs in a very different way than the way it was done in the past. And this is obviously a necessity for deploying AI at scale. So that was one technical opportunity or reason that we thought we should tap into.
The second basic premise was based on the fact that the hyperscalers were starting to become more vertically integrated, doing their own chips, which obviously you can see a whole bunch of announcements happening. We're following all the re:Invent stuff going on with AWS. So based on those two requirements, one technical, one more supply chain, and how the hyperscalers were behaving, and the general premise that AI is going to become mainstream is why the company was started. So today we are a little over six years since we started operations. We went public earlier this year, and I'm very glad and happy to share that most of the AI deployment that's happening worldwide with different hyperscalers uses technology that Astera has developed for interconnecting GPUs, CPU, and so on for protocols like PCI Express, Ethernet, CXL, and so on. So that's sort of the context of the company.
We are headquartered in California, but obviously have offices worldwide.
Great. So you guys are obviously levered to the AI mega trend as a whole, but specifically, what would you say are your primary growth drivers?
Yeah, so for us, again, we are a pure-play AI company. More than 80% of our business comes through AI deployment, whether it's based on third-party GPU platforms like NVIDIA or AMD, and also for internally developed ASICs that various different hyperscalers are developing. So for us, we are benefiting from the overall macro trend in terms of the investment that's going into the AI infrastructure. From various reports, you can see that's probably about $300 billion of investment that's happening next year, twice of what it was last year. So in general, we are benefiting from the overall growth and deployment of AI infrastructure. Specifically for us, what we're also benefiting is the products that we have developed, which are all grounds for supporting the AI infrastructure, are first-to-market, innovative first products that were addressing some fundamental bottlenecks for data, networking, and memory, which all are very critical bottlenecks.
Because if you think about GPUs today, they're only about 50% utilized, meaning even though people are writing big checks to companies like NVIDIA, half the time they're just sitting there collecting dust because they are waiting for data or memory. So we are solving that problem. So there are several different opportunities for us to grow our business. And the fact that we are deployed in all the major AI infrastructure worldwide, we do have that front row seat. So we are expanding our product line based on what we see as requirement. We have the trust from the customers, so we are rapidly getting more business, which is reflected in some of the numbers that we have shared as part of our earnings call.
We do see a significant path ahead with more opportunities as we are developing products like our fabric devices that are becoming more central to the AI deployment. These are high-dollar content opportunities, and per GPU, our dollar content is growing significantly generation over generation.
What about GPUs versus ASICs? Are you agnostic between a GPU-based and an ASIC-based deployment, or are you happier to see one or the other from a content point of view?
Yeah, so I mean, in general, we are agnostic because they all follow industry spec. But maybe one caveat to that is the NVIDIA ecosystem versus the non-NVIDIA ecosystem, in the sense that today, obviously, we have significant business on the NVIDIA ecosystem. But most of that comes in what is called as a front-end network, where the GPU is talking to the CPU, the storage, the networking. The back end, where the GPUs talk to each other, NVIDIA uses a protocol called NVLink, which is proprietary to NVIDIA, so we don't get to play in that space. But if you think about the ASIC platforms or internal accelerators, there what happens is that we are servicing sockets that also exist on the back end because they're relying on standards like PCIe or Ethernet and other standards. So we get to play in both the areas. That's number one.
Number two is that when you're playing in the back end, you're connecting each GPU to several other GPUs, depending on the size of the cluster. You could have eight GPUs or 32 GPUs, so every GPU has to connect to like 31 other GPUs. So there are a lot of I/Os that go, and I like to call it like a fertile land, meaning there's a lot of opportunities for us, for folks that are following the AWS announcements this week when they announced the Trainium 2-based platforms. And if you look at those things, there are every half a rack, full rack has got like almost 128 of our PCIe-based AECs, two retimers per, as an example. So 256 links on the back end. Compared to front end, it tends to be much less.
So from a dollar content standpoint, opportunity, both because we get to play in front end and back end, and there are many, many links on the back end that we can service. So that tends to be a bigger opportunity for us. And the fact that the hyperscalers are continuing to double up on their own ASICs and accelerators that they're building, so for us, that's sort of good news because we continue to see a significant growth in opportunities there. If you look forward, the fact that we announced our Scorpio X- Series, which is for GPU to GPU interconnects, that also is a driver because these are much larger devices, higher value, higher ASP. So that also makes the custom ASIC platforms much more attractive from a business growth standpoint.
But I say all of this stuff, it doesn't mean that we are continuing to win designs on the NVIDIA-based ecosystem. We continue to do that. But just in terms of dollar content-wise, we do see a higher dollar content per GPU on the ASIC-based platform compared to an NVIDIA-based GPU platform.
Right. So next I want to ask about something that's pretty similar to that. So internally on our team, we're always debating the right attach rate of the networking equipment to the compute equipment. And I think your stat about the 50% utilization of the compute resources suggests maybe networking is not well developed enough already. But do you think that the networking attach rate is going up over time? And as inference grows relative to training, is that something that should help as well?
Yeah, absolutely. And I think companies like NVIDIA have done a great job building these fantastic compute engines, machines. But these machines have to be fed with data. They need memory. That's really where the bottleneck is. I mean, like you referenced, it's only about 50%, 51% utilized. So for the most part, that's where the opportunity is to interconnect this, whether it's from a scale-up standpoint or scale-out standpoint. There is also the nuance that the AI clusters are built in a mesh topology. Compared to general compute, you have this tree structure or hierarchical topology. Here what happens is with the mesh topology, like you noted, you have every chip, every accelerator, or GPU talking to everything else so that together, all of these GPUs appear as one giant GPU.
So the emphasis on robustness, bandwidth, and the fact that you have many of these links essentially implies that the number of sockets that you have tends to be higher compared to a compute type of socket.
Yeah. So you mentioned Scorpio already, and I'd love to dig in a little bit more on that. So what exactly is Scorpio? What are the pieces of that, and what opportunity does it address, and what are its unique selling points versus a competing product?
Yeah, absolutely. So first of all, again, for us, the vision or the way we approach our product development is to look at what it takes to solve at a rack level. Because I think AI infrastructure, you can't just look at one piece of the puzzle or one server or one piece of the server as a solution space to tackle. You've got to look at it at a rack level or cluster level and start saying, OK, what are the data, network, and memory bottlenecks to solve? The way Astera has approached the market is to focus on four protocols. There is the UALink, which is emerging for the GPU to GPU interconnect or scale-up topology. We're focusing on Ethernet for scale-out to build the larger clusters.
And then we are focusing on CXL for supporting the memory expansion and PCI Express for interconnecting the storage and networking and things like that. So the key point here being is that we're focusing at a rack level what is required, whether it's from a semiconductor product, from a hardware product, or software. So Astera is sort of unique that way that although our key IPs in semiconductors, we take a look at it holistically with software, which we call Cosmos, along with hardware. So that's one piece of the answer to you, just so that we are in the same context. Now, addressing the Scorpio device itself, so Scorpio is the industry's first fabric device that's developed for AI interconnects, whether it is to interconnect the accelerators to each other, which we call the X- Series.
And the second series in that is what we call P- Series, which uses the PCI Express standard to connect the GPU cluster to the CPU storage and networking. So they all talk a language called PCI Express. So the fabric device is used to route the signal from the GPUs to the CPU storage and so on. So those are the two product lines we have. The X- Series, obviously, is at the heart of an AI server because that is the equivalent of, let's say, an NVSwitch that NVIDIA builds as part of their ecosystem. So it's a greenfield use case in the sense that as the non-NVIDIA ecosystem continues to expand, the Scorpio X- Series is designed to address that space, which is multiple billions of dollars of opportunity for Astera.
We are essentially implementing that by being on the board and driving the technology of a standard called UALink that's now starting to become more of a common way of interconnecting GPUs for non-NVIDIA ecosystem.
Have you guys sized how big of an opportunity Scorpio is for you next year in dollar terms or not?
Yeah, so total TAM, as we've sized it by 2028, it's right around $5 billion. The way that we would dissect that is really between the P-S eries and the X-S eries opportunities. So if you look at the market as it stands today, it's really more what the P- series would address in terms of connecting GPUs, CPUs, networking, and storage. So that's a real market today that's primarily addressed by a competitor. But as Sanjay mentioned, we're the first to market with PCI Express Gen 6. So as these sockets start to turn over and transition to PCI Express Gen 6, we would expect to prosecute a lot of those sockets. That market is growing, and it will continue to grow going forward.
The exciting piece of that $5 billion is really a greenfield opportunity that the X-S eries product line is going to address, and that's using switches to scale up on the back end, and then that's effectively zero or close to zero market today, but we see many opportunities across multiple use cases and across multiple customers that are going to present themselves as a big opportunity to utilize a switching topology to help scale up on the back end, so we see that growing from effectively zero today to multiple billions of dollars over the next four or five years.
Great. So I want to move on through the constellations now from Scorpio to Taurus. So can you talk a little bit about the Taurus product line, what opportunity it addresses, and any differences in approach between what you do and what some of your competitors like Credo might do?
Yeah, absolutely. So again, for folks that are not familiar with the constellations, we have Ares, which is a PCIe re-timer. Taurus is for Ethernet AECs. Leo is for CXL, memory expansion, and Scorpio is for fabric switching and all that stuff. So Taurus, in particular, to answer your question, these are products that are designed to go for interconnecting 400 gig, 800 gig Ethernet connections, whether it is for general compute with a server connecting to top of the rack switch or for a scale-out topology in an AI infrastructure. So what these products do is we offer them in a module form factor, meaning it's not just the chip. There's firmware and other components that come together. The modules are designed in a way that they fit inside the connectors that various different cable vendors make, like Molex, Amphenol, and so on.
What they are doing is essentially enhancing the signal. The reason for that is as speeds have gone up from 25 gigabit to 50 gigabit to 100 gigabit per second on Ethernet, the signal obviously is doubling in speed, but the reach requirements haven't changed. Just like if you try to run faster, you get tired sooner, same thing happens with high-speed signals. As the speed increases, they get tired, and they don't make the journey from point A to point B, and they die off. What our chips do is that they ensure that the signal is amplified and the noise is removed so that they can make the journey. That's what the Taurus product does as a chip.
But in terms of what is the approach that's different is Credo, obviously, did a good job, recognizes the market, and their approach was to build a cable, a complete cable. We believe that the cable vendors like Molex and Amphenol, who have invested hundreds of millions of dollars and billions of dollars into R&D and manufacturing, are better suited to do the cable. So we focus on providing the electronics, the chip, the module. And this module is designed in a way that it can fit into multiple cable vendors' connectors. So it's scalable. And hyperscalers generally like that because when they deploy or buy their cables and connectors today, they like to have multiple suppliers, and they obviously divide the share between them. So we are essentially coming up with a form factor that does follow the existing supply chain mechanism versus having a complete cable.
And honestly, we also don't want to get into inventory issues where you build a cable for one meter, but the demand changes to 1.5 meters or something else. And what do you do with all of that? So with our module form factor, what happens is it's completely transparent in the sense that it could be used for any cable length, any cable vendor. So it provides us much more sort of flexibility. The AUP or ASP tends to be different. So you'll see that reflected in what Credo announces and what we announce. But generally, our business is more profitable. So for that standpoint, we are taking slightly different approaches. And then you have someone like Marvell that just provides a chip, which is their approach. We believe the cable guys are not equipped to do the electronics that are needed.
So to that standpoint, I think our approach is something we believe is much more scalable and portable.
Yeah. OK, so moving on to Leo, can you talk a little bit about CXL and when you expect that to be more broadly adopted and as that grows, how big that opportunity can get?
Great. No, definitely. So for folks that are not familiar with CXL, so CXL is a standard that allows for memory expansion. The reason that is important is as the model sizes or databases have gotten bigger, the memory that's directly attached to the CPU or GPU through DDR or HBM channel is not sufficient. So that's, again, a bottleneck that exists today. CXL solves for that. CXL, to be honest with you, it's a new technology, so it's gone through its phases of adoption. The first place where we see CXL being deployed is on the general compute side for large database applications where you need more memory in order to reduce the latency and all that stuff. So I want to say at this point, CXL is sort of transitioning from that crawl stage to the walk stage. We are not running yet.
We have at least the four major hyperscalers in the U.S. all developing CXL-based platforms right now. Some of them are in pre-production racks in the data center, meaning there's a good chance that some of the workloads that you guys are accessing use CXL. We expect CXL to be more broadly deployed in production volume starting second half of next year. We also are seeing a lot of inference use cases where CXL benefits. So we have done some white papers with a lot of the hyperscalers that you can access from our website. But we do expect that starting second half of 2025 through 2026 and later is when the CXL technology will continue to get deployed from a production volume standpoint.
Got it. Can you talk about how you participate in the UALink Consortium? I think your involvement with them is not new, but you also had a recent change with your board membership. So can you talk a little bit about the work that consortium does and how you see it?
Yeah, absolutely. So again, UALink is a standard that is similar to NVLink, but again, for non-NVIDIA ecosystem to interconnect accelerators. So we replaced Broadcom as part of the board as an interconnect technology vendor. And the rest of the board members are companies like Google, Meta, Amazon, which generally doesn't join boards. This is probably the second time that they're joining a standards body like this. That kind of gives you an idea of how critical UALink is for the deployment of GPU to GPU interconnect. So we expect the first version of the standard to be released in Q1 of next year with actual ecosystems starting to form in 2026. So Astera, obviously, will be taking a very active role that we have already been doing right now.
We do believe that UALink will end up becoming the alternate for GPU to GPU interconnect that will be broadly deployed. There are several benefits of UALink compared to things like Ethernet from a latency, cache coherency, memory semantic standpoint. We are promoting that and working with the partners to ensure that it gets broadly deployed.
Great. So lastly, not named after a constellation, but all of them together is the Cosmos software, which you referenced earlier. So can you talk about what value that brings and what kind of competitive differentiation you get from that software layer?
Yeah, so if you think about the challenges that hyperscalers have today, obviously, they're running at a million miles an hour right now, and then they are also, even though you can call them with one word like hyperscalers, each one is different, meaning the workloads that Meta is running are very different than what Amazon is trying to do. Because Meta is running mostly internal workloads, Amazon is running external workloads, so their infrastructure really is designed differently, so what we have innovated and done is our chips are software-defined, meaning 60%, 70% of our chip is actually implemented in software. The benefit of a software-defined architecture is that it is customizable. You can upgrade it, just like your cell phone. You get an upgrade, and suddenly it's got new features, and bugs are being fixed, meaning it's the same hardware.
So it's a very, very powerful architecture in terms of supporting the capabilities of the chip. The third important thing is these systems have become very complicated, and you need to monitor them in terms of telemetry, diagnostics, fleet management, predictive failure. So what we have done with our chips is to essentially implement a lot of sensors and algorithms in our chip that's monitoring the health of the nervous system, the connectivity nervous system. In some ways, I give a very crude example where if you think of your own nervous system, wouldn't it be great if you're able to detect a clot that's forming or some other issue that's going on? So we are able to detect that as part of the software-defined architecture and sensors that we have built in our connectivity products. That functionality is captured in the software interface that you call Cosmos.
In some ways, if you think of CUDA as the API that NVIDIA provides for the compute application, what we are providing is Cosmos APIs for the connectivity infrastructure in the AI server. So a lot of our hyperscalers essentially use APIs, a program on our Cosmos API, to either monitor their infrastructure and their high-speed links, detect failures before they happen, customize the functionality of the infrastructure, and optimize it for things like performance and so on. So it's a very powerful architecture and API that provides the control and visibility and observability that's needed for this complex infrastructure that's being rolled out.
The nice thing about Astera is that the Cosmos API is the same, whether it is the Ares product line, Taurus product line, Leo, or Scorpio, meaning once customers integrate our API into their operating stacks, we are able to leverage that to add more functionality for a given device, meaning the device keeps getting better over time because of the upgraded software, or we're able to sell them more products because they can very seamlessly add more products within the API that's already integrated into their platforms, so that's what Cosmos does, and that's the architecture that's given us the tremendous sort of design win activity that we see today, and for us, really, like I noted early on, we envision our solution space to be at a rack level, meaning it's not just about chips. It's about hardware. It's about software.
We are trying to provide a holistic solution that allows our customers to get to market sooner, which is very important in the AI race that's going on right now. It simplifies the qualification and deployment for our customers because of the software-based architecture that we have. Most importantly, it allows them to invest less of their resources or look at us as their extended engineering team, which is something that has allowed us to continue to expand our engagement with these customers.
So we only have a few minutes left, so I want to zoom way out and say, as you look at your entire product roadmap, what are you guys working hard on, and what are you really excited for that's coming up next?
Yeah, so I think there's many things going on. Like I noted, because of the success we've had and the trust that we've gained from our customers by focusing on two main things: innovation, execution. That's all we think about. We are not trying to over-optimize or do things that don't necessarily solve a customer need. Because of that mindset and having proven ourselves where we have released four product lines in little less than six years of operation, we are developing features that are fundamental to deploying AI at scale. So now customers are coming to us and saying, hey, you have four product lines. Can you do a fifth or a sixth or a seventh? So we are hard at work addressing those new product lines and new features. And we will expect that to continue to help us grow the business.
Again, we are approaching it in multiple different form factors. But what I will say is that, again, we don't talk about unreleased products. That's not who we are. These are complex systems. We talk when customers qualify and start placing POs. So to that standpoint, we are not revealing everything we're doing right now. But trust me, on a daily basis, I think it's amazing how much energy there is, how much traction and engagement we have as we grow our product lines.
Great. Well, I think I will have to ask you the same question again next year and see what's new in a year's time if you're not going to let us look behind the curtain in advance. Sanjay, Nick, thank you so much for your time today and for all your thoughts.
Thank you so much. Thank you, guys.