Okay, guys, we're going to go ahead and get started here. This morning, we are very fortunate to have Nick Aberle with us. He's the Senior VP of Finance at Astera Labs. That ticker is ALAB. It's one of the most interesting and probably, I think, in the last 20 years, one of the fastest-growing companies that I've come across in the world of semiconductors. With that, I'll ask you the first question, Nick. Thanks for coming here. Thank you for being here. Your growth has been explosive since, I think, 2021, 25x in four years or something like that, revenue growth, pretty stunning numbers. How do you think about growth going forward and the dynamics of AI in which you play in?
Yeah. Just first of all, thanks, Harsh, and thanks to BMO and everybody for joining and being interested in the Astera Labs story. Maybe just for a touch of context right off the bat, for you guys that don't know us, we develop and deliver silicon hardware and software to hyperscalers and AI platform providers for AI, rack-scale AI infrastructure. When the company was initially founded by the founders, I think it was about 8.5 years ago, the initial mission was to address data bandwidth and memory bottlenecks within AI infrastructure applications. That mission really hasn't changed as we fast-forward to today, and you could probably argue that it's more critical today than ever, especially given some of the tight resources and shortages that we have on compute specifically and memory within the market.
For us, growing has always been really kind of a two-pronged story. Obviously, we're 100% exposed to AI infrastructure, so we've ridden the wave of just the secular trend of AI populating across the GPU providers like NVIDIA and AMD, and then more recently, some of the hyperscalers building their own captive XPU solutions as well. But the real kicker for Astera is our product expansion and our silicon dollar content per XPU story that's layered on top of that kind of underlying XPU benchmark that we're growing along with. If you look at the company just maybe four years ago, we could grab approximately $50 of content on a per-XPU basis in an AI server. AI servers used to be 4U, 6U rack mount servers, and fast-forward today, you have 62 or 72 GPUs within a rack.
Today, we can command upwards of $1,000 of dollar content per XPU. What we just outlined on our most recent earnings call is that we see a path to doubling that dollar content generation on generation for each subsequent generation of XPUs that's launched over the next, call it, five years. It's been a very fun ride and a dynamic growth story to date, as you mentioned some of those metrics a minute ago. But I truly believe there's a huge opportunity in front of us to continue to grow in excess of the secular demand that we're exposed to in AI, and that's going to be through a combination of new products, new customers, new designs, and us taking advantage of that increasing silicon dollar content per XPU.
Thanks, Nick. About three to four quarters ago, you launched this new product line called Scorpio, and you have two versions of it, the X and the P. Maybe you can talk about the difference between the X and the P, who is suitable for which line. Also, I think it is doing really well. The line is doing really well. Maybe you could hit upon the traction you are seeing as well.
Yeah. The launch of Scorpio, which was just around like two years ago, just a little over two years ago, was transformative for Astera Labs, and it really kind of put us into a new phase of our growth story as we moved from kind of smaller point solutions in the signal conditioning space that would basically look at signals that were traveling between two endpoints, to switching solutions and AI fabric solutions that were much more instrumental to the connectivity performance within a system. Especially as cluster sizes are increasing and domain sizes are growing, the fabric and the switching capabilities that you can embed within your system are going to be very influential on the ultimate kind of performance that you have within that system.
As Harsh mentioned, we have two sub-segments of the Scorpio product line. The P-Series is for what we would call kind of a scale-out or head node connectivity type of application, where that P-Series switch would sit down on a compute board and interface between CPUs, GPUs, storage, memory, and networking. That is all done over the native language within an AI server, which is PCI Express. The solution was specifically built with AI in mind and was well-positioned to grab sockets as we moved to the current generation of PCI Express Gen 6, and helped to displace competitors that were also targeting that market that were bringing kind of more general-purpose solutions to that space.
The X-Series, as you mentioned, has been kind of the more kind of explosive part of the growth story, and that started to unfold here over the course of the last couple of quarters. If you go back to just Q1 of this year, Scorpio X-Series was effectively just slightly above $0 in revenue. Exiting this year, we will be on a well above a billion-dollar annualized run rate for just that product alone. The reason why it has been able to grow so fast is because it carries tremendous value within an AI system, so it is specifically tailored for scale-up connectivity.
Scale-up connectivity references the interconnect between accelerators, XPUs, or GPUs within a system. Hyperscalers are ultimately trying to deliver a system that has the lowest latency and highest performance when you are speaking and talking between these XPUs, and those XPUs are supposed to act as one large computer or one large XPU. You need a robust AI fabric, such as Scorpio X-Series, to enable that connectivity.
What we see in that space is the ability to capture significant value, significant ASPs. These are large, very complex solutions. That has allowed us to bring, as I was referencing earlier, silicon dollar content in the $50-$100 range up into the $1,000+ range, as we started to embed those solutions into these products.
How many customers? You have talked about some wins for the X-Series and the P-Series. Whatever latest numbers you have, we would appreciate it.
Yeah. On the P-Series side, it is very ubiquitous that PCI Express is used within these systems to interface. It is the common language between GPUs, CPUs, networking, storage, memory. Really, the market there, while the ASPs and value is probably not quite as high as the scale-up piece of the portfolio, there is a very broad customer base because every single piece of AI infrastructure out there is going to leverage PCI Express and a switching solution similar to this to enable that head node connectivity. We have been shipping to upwards of five customers over the course of the last six quarters. Our lead customer, AWS, has ramped that in high volume on customized NVIDIA-based GB200, GB300, and then forthcoming Vera Rubin racks.
Then we have also ramped a second hyperscaler in the Q3 timeframe as well. That will kind of fold into the story, into the revenue line here over the next couple of quarters. Then there are several customers beyond that as well. That will continue to get more diversified as we move into 2027 and 2028. On the X-Series side, it is a bit of a more narrow market from the perspective of Scorpio X is specifically focused on enabling scale-up over PCI Express or protocols that are based on PCI Express or customized versions of PCI Express. We estimate, if you take out proprietary protocols like NVLink out of the market, that basically the market will split roughly half-half between PCI Express and Ethernet-based scale-up. We can address roughly half of the, what we would call merchant AI fabric market.
X-Series started to ramp at a lead hyperscaler customer earlier this year, and that's driven a large share of the big revenue growth that you've seen over the last couple of quarters. We started shipping to a second customer in Q3, which is an international hyperscaler customer. Then we've said that we'll ship to at least two or three more X-Series customers by the end of this year. We should be exiting the year around four or five customers. Then we have a pipeline of customers that's kind of in the range of kind of the teens, I would say, that we would look to prosecute, convert, and start to ship product to over the course of 2027.
Great. I want to shift gears a little bit to your old main product, Aries, the retimer line. What kind of share do you have in the market? How do you get there? Also, how do you see the growth of the Aries line now that you have this big share of the market?
Yeah. Aries, yeah, it's old, but not that old. The Gen 6 Aries solutions are a couple years old at this point, and we just announced our new Gen 7 portfolio at OCP last month. Plenty of legs left to the Aries story. I think the important thing about Aries as well is if what you look at what we've done and have been able to do with Aries over the last couple of years, and I think you're alluding to this a little bit, is gain a very substantial share in the market. It's probably upwards of 90%, 95% today. It was that in Gen 5. It'll likely be that again in Gen 6.
Having that breadth and that kind of wide deployment, kind of incumbency allows us to gather quite a bit of tribal knowledge and understanding of different corner cases and how to address interoperability between a lot of eclectic endpoints that we connect between in the market. That's put us in a pretty good position from just a stickiness standpoint to kind of retain a decent amount of share here going forward. It's also allowed us to introduce all of our customers to our software suite COSMOS.
COSMOS historically has been used to provide predictive analytics, telemetry, diagnostics back to fleet managers that are presiding over tens of billions of dollars worth of infrastructure within these data centers, and provide valuable insights as to what's working, what's not working, what needs to be driving higher utilization, where you might have heat and temperature issues, degrading links. It's been a very important tool for hyperscalers and AI platform providers to proactively manage their compute and their fleets and get in front of big issues to basically optimize utilization.
What you see now as a result of that is you have COSMOS effectively saturated across the market. It's been embedded in most, if not all, operating stacks by fleet managers across the entire ecosystem. We haven't sat still on building out that portfolio, so to speak. We launched a new product line at OCP called Orion, which takes basically all of our hardware products and builds them into a full rack. That rack is then managed and enabled by COSMOS, which is now not just managing one link connecting two endpoints, but it's managing an entire rack, all the XPUs within the rack, all the switches within the rack, all the connectivity. It's now a full-fledged rack management tool.
Customers already have the experience of leveraging COSMOS at a smaller scale, but now they're able to deploy that in a more broad way. Aries has been very instrumental for our foothold into the business. I would say that going forward, we would expect it to continue to grow very nicely. It'll obviously be exposed to continued growth in XPU shipments, but the move to PCI Express Gen 6 and then ultimately Gen 7 will drive higher ASPs. Usually generation over generation, we see around a 20% uptick in ASPs.
As speeds go up, the distances that you can move signals shrinks, so we would expect to see higher attach rates overall for that line of business as well. We're pretty optimistic that Aries continues to grow pretty nicely here. Sadly, it's growing very nicely and coming down as a percentage of revenue because it's just a much smaller ASP than Scorpio, but it's still driving good growth.
It's a good problem to have, right? You want those kind of problems. The other problems you don't want. Let me I always try to make it interactive and just at least give you guys the opportunity, if anybody in the audience has a question, I'm happy to repeat it so it's webcasted, but please raise your hand and shoot away if you do.
Okay, great. You talked about COSMOS. You have a couple of other lines. We never talked about Leo CXL. I'm going to ask you two questions, then you can answer it in any which way. Let's talk about the Leo CXL controller, what it does, and why it's important, and what the prospects are for it. Then the idea of Orion, did that come from a customer? Because we're seeing situations where customers are coming to companies and saying, "Hey, give us a rack." And then that gives you guys an opportunity to put a whole lot of content into the rack and spiff it out, if you will, before you ship it. Did that come from you guys or did Orion come from some customer suggestion?
Yeah. Starting with Leo, we have been in the market with Leo for the last 3.5 years, and kind of one of the forefront players in the CXL controller market. What a CXL controller effectively can do for a system is allow a CPU or a GPU to access more memory, kind of per endpoint basis. Because these units are pin limited, they can only access a certain amount of memory. By using Leo memory controller next to it, you can access twice as much memory as you would be otherwise. It can also be used to share and pool memory between resources.
The same GPU, pool of GPUs or the same pool of CPUs can access effectively a tray of memory dynamically on the fly, and whichever endpoints are needing the most memory at that time will be allocated that memory, and the others will be idle. It helps to kind of deal with those dynamics, which is particularly important as we sit here today, given memory dynamics from a commercial standpoint. You have seen increasing prices and supply shortages across the industry, so you have a lot of customers now coming back and looking at CXL very aggressively, saying, "How can CXL help me increase utilization of the memory I can get my hands on?" Which is pretty tricky to do.
CXL as a business for us has been relatively small up until now. We have talked about activity and engagement upticking pretty nicely here over the course of the last six months. We have design wins across two hyperscaler customers at this point, both for general purpose and AI-based applications, and we would expect those to start to ramp in the middle to back half of next year. So, pretty excited about that line of business. It has been a long time coming, but hopefully we can see some contribution.
Can I push you on what those applications are? Like, what is the work that is being done? What is being connected to what in those?
Yeah. On the AI side, it effectively will be for a disaggregated AI inferencing solution that would be looking to do one piece of the AI inferencing workload. And so that would be used in those boxes to help offload KV cache. On the general purpose side, what we've seen is customers looking to, again, given the supply shortages and the tightness in the industry, reuse DDR4 within their systems. So taking DDR4 out of older legacy retired servers and repopulating that memory into newer servers because they can't get their hands on DDR5. You use the memory controller to basically allow newer endpoints that only have DDR5 controllers to be able to access DDR4 and leverage that pool of unused memory. Those are the big two applications we see today.
We got like two minutes. I want to ask about Taurus. Because we left that out so far. Just what is it? What does it do between the AEC and the Ethernet? What part? And then I also want to talk about the culture of your firm, because when I and again, two-parter. When I look at your focus as a company, it's probably the most intense that I've seen in a decade. You guys are hyper-focused on quality. We don't see shares of 90% upward in things. In this. Let's be real about it. So, where does that come from, the culture? Maybe you can answer that first and then talk about Taurus.
Yeah. From a culture standpoint, if Jitendra was here, and he often repeats this, the culture is a very customer-obsessed culture and one that's very focused on execution. I would say that those are the big two pieces. And those things have lent us, or have blessed us with some nice, a nice seat at the table with large customers, large AI platform providers like NVIDIA and AMD, and all the hyperscalers as well, to really kind of see what's coming up in the next couple of generations.
As you've seen every generation go by, customers are coming to us saying, "Hey, you did a great job on this one. We have X, Y, and Z problems coming up for the next one. We're going to need you guys to do more." The relationships that we've built on the back of the track record and support over the last couple of generations has really afforded us some nice momentum.
You're not taking the credit for it. You're giving it to your CEO.
Yeah. The team is amazing. Obviously, these guys are studs. Taurus, similar story to PCI Express and Aries. As we move to faster speeds, the need for more hardened and more capable signal conditioning solutions is going to be needed. The big transition for Ethernet is going from 400 Gb port switching speeds to 800. As we see that transition, you will start to see platforms adopt more AECs, and we would expect to participate in that roll-out as we see it.
Great. We are out of time. Thanks, Nick. Thank you for your time. Thanks, everybody, for being here.
All right. Sounds good. Thank you.
Really appreciate it. Thank you.