Astera Labs, Inc. (ALAB) Earnings Call Transcript & Summary

Okay, we'll go ahead and get started. Welcome everybody, my name is Quinn Bolton, I'm a semiconductor analyst for Needham & Company. Thank you for joining us here on the second day of Needham's 27th Annual Growth Conference. It's my pleasure to host this fireside chat with the Astera Labs. The company was founded in 2017 and is a global leader in purpose-built connectivity solutions and software for AI and cloud infrastructure. Astera went public last March, and has seen tremendous growth in both its operating results and stock price since the IPO. Joining me from the company are Jitendra Mohan, CEO and Co-Founder; Nick Aberle, VP, Treasurer and -- VP of Treasurer and Investor Relations, and Leslie Green handles Investor Relations, is in the audience. So Nick, Jitendra, thank you for joining us. I guess maybe as a starter for folks that may not be as familiar with the company, can you give us a quick overview? Walk us through the company's 4 product lines. And as you do, can you highlight for us what are the company's core competencies that link these products together.

Yes, happy to. So first of all, welcome, everybody. There are some familiar faces. You'll have to bear with the storytelling for a few minutes here. And for those of you who are new, not as familiar with Astera Labs, let me give you some context. So we started the company in 2017 with a singular focus on connecting thousands of GPUs together. So we believe that the AI workloads would require a new architecture where we would have many GPUs connect and this workload will only be possible to run in the cloud. So as Quinn mentioned, we founded the company to deliver AI infrastructure to run in the cloud. And that's what we have done pretty singularly over the last 5 or 6 years. Now what we did not imagine is that these models will grow into the trillions of parameters and the thousands of GPUs will become hundreds of thousands of GPUs. And because of that connectivity has become a huge bottleneck, whether it's data bottleneck, network bottleneck or memory bottleneck. And we have been solving these problems to make sure that the utilization of these AI clusters is as high as possible. Today, even the best-run AI cluster is around 50% efficiency. You give billions of dollars to the GPU providers and build your ASICs and only utilizing that 50%. And the way we address this problem to bring up the utilization is by attacking these data bottlenecks and networking bottlenecks and memory bottlenecks, which gets us into the product lines that we have. We started with our Aries product family, which is PCI Express retimers that address the data bottleneck. These devices are in full production. We are many times known for our retimers. We have a majority market share, ship millions and millions of these devices. Following our Aries retimers, we came out with our Taurus smart cable modules. These are Ethernet retimers that eventually get assembled into active electrical cables, and they are going to start ramping -- they actually started ramping in Q3 and Q4, and we continue to see those ramp in the rest of the year. When we came out with our Leo family of CXL Memory Controllers and they address the memory bottleneck by delivering more capacity and more memory capacity and more memory bandwidth to CPUs and GPUs, and we should start to see them contribute meaningfully to our revenue this year. And last but not the least, last October, we introduced our Scorpio family of products, which were like our other families, purpose built for AI applications. And as Scorpio comes in 2 families. One is the P-Series. The P-Series is used for scale-out applications. And I'm sure during the rest of the talks, we'll get into what that is and the Scorpio X-Series, which is responsible more for scale-up connectivity, which is a great opportunity for us. To your point, Quinn, on how are we different from others. There are many ways we are different. There are speeds and feeds and power and latency and so on. But I would say the 1 thing to point out is just the deep relationship that we have managed to establish with our customers over the last 5 or 6 years. Of course, it probably started out with them not really believing what we told them. And then what we told them came true, and we did not disappoint. And so over the years, it has become so close that we get involved in the design decisions of their future platforms as they are developing their products. Oftentimes, we tape out our chips at the same time as our customers are taping out their SoCs. And so that gives us just incredible amount of visibility into where the customers are going, what type of products we should build, what features to add, what features not to add, what timelines to target and so on. And I think that's a distinct advantage that we have over our competition outside of software and chip architecture and things like that.

Yes. That's where I was going to go sort of next was one of the other differentiating factors for Astera is your architecture. You've used sort of more of a software-defined approach where you have a number of microcontroller sensors embedded in your designs. Talk about that design approach because many of your connectivity competitors are more hardwired designs, DSP-based.

Yes. So I've been doing semiconductor design myself for like last 30 years. And traditionally, the way these chips were designed was to figure out what the problem is and put a solution in hardware -- state machines and dedicated hardware and so on. And that was good enough. Those chips actually worked pretty well, solved the problem very efficiently and life was okay. But that was also the era where the solutions lived for many years. You developed a chip and it continue to work in that protocol for that application for several years. That is absolutely no longer true in this kind of AI world, where the design cycles are now shrinking to 1-year cadence. And so those traditional architectures are simply not able to keep up with sort of the evolving use cases, applications that our customers want. And so we were lucky enough or fortunate enough to realize this and build our chip from the ground up with a software-first approach. So we do whatever we can possibly do in software. And when the software is not fast enough, then we will use hardware to supplement that. And what that does is result in a chip that is extremely flexible. One that we can customize for different applications. We can customize it for different hyperscalers, we can tune the chip to deliver optimizations that our customers want. And on top of that, when you start to put hundreds of thousands of these chips into your AI infrastructure, the diagnostics and telemetry capability has become really, really important. And so we added those on top of the performance determining parts of the chip. So we have very performing ships that are very flexible and give you incredible amount of diagnostics that lets you manage your infrastructure. Over time, really, we have become the eyes and ears of our customers' connectivity infrastructure. And the way we make all of this available to our customers is through a software we call COSMOS. So there are parts of COSMOS that run on the chip that are responsible for the optimizations and the customizations and the diagnostics features that I talked about. And there are other parts of COSMOS that are running in our customers' cloud stack. So when they are running their -- managing their infrastructure, they pull all of this information out of our chips and then act on it in an intelligent manner. And this provides us with a lot of stickiness. It's just not possible anymore to just rip and replace, right? If you take our chip and we white label it and give it to somebody else, they still will not be able to displace this because of this strong software component that we have. And this COSMOS software is applicable to all of our products to support all the products, it's very easy for anybody to upgrade from 1 generation to the next and it's very -- it's much easier for us, I would say, not very easy, but much easier to introduce a new product under the same COSMOS umbrella.

Perfect. I wanted to start now moving to some of the product lines, maybe starting with the Aries and the PCI retimers. The business really sort of took off about 6 quarters ago as you launched or NVIDIA launched this Hopper platform with pretty significant Aries retimer content on it. As we look forward now to the launch of Blackwell, talk about sort of the puts and takes for Aries content on the new Blackwell-based systems.

Yes. So this will launch of Blackwell did create a lot of confusion in the investment community. We have been aware of what Blackwell means 2 years ago, since 2 years ago because that's when we started working together with the GPU provider as well as the eventual hyperscaler customer on what it would take to deploy Blackwell at scale. But to be crystal clear, if we look at the Blackwell platform as a whole, our content, Astera's content goes up relative to the Hopper generation. However, it goes up because of the additional Scorpio content, which is the PCIe Switch Fabric and not just because of the Aries. In fact, Aries content goes down on the whole with the Blackwell platform. Within that, if you look at each individual form factor, whether it's [ GX ] Board or an [ MGX ] or what have you, the content differs. Sometimes, it's the same as Hopper, sometimes it's less than Hopper, but in the cases where it's much more, and those are the cases where a hyperscaler is customizing the Blackwell platform for their own infrastructure needs. We have an amazing amount of content that kind of overshadows any lots in retimer content. And I'll make 1 other comment that if you kind of zoom out of that and look at the retimers as a whole across GPUs and ASICs then the retimer content goes up as well.

I wanted to obviously hyperscalers with their custom-designed ASICs is a growing part of the business. Late last year, AWS hosted its Reinvent show, and I think they demonstrated some of their new servers, so they're training 2 server has a pretty high attach rate for PCI retimers, the training to Ultra has both your retimers as well as the PCI AECs. And so can you talk just about some of those platforms that have been shown in public about your content? And specifically, I think there is a GB200-based platform that may have had Scorpio content. So talk about your content there and sort of some of the Amazon platforms.

So clearly, we'll be very restricted in talking about the platform themselves, let on our content in it, but I think we can definitely talk about the application. And there are certain things that are easy to see, for instance, our PCI AEC based on Aries retimers, you can see them. The important part about this launch that Amazon had in this architecture is we now are able to play in the scale-up network. So going back to the Blackwell family, NVLink is a scale-up fabric for the Blackwell family, so we don't get to play in that. We only get to play in scale-out. But with these other architectures where people are using PCI Express like fabrics for scale-up, we have a huge opportunity. You mentioned PCI Express. So that's something that has been ramping now for 1.5 quarters, and we continue to see strong trends because scale-up requires a large amount of interconnects. You are connecting a lot of GPUs all-to-all connectivity and that just drives up a lot of volume. There's also a lot of application for the Scorpio X family in these architectures. Again, because you're trying to connect these GPUs together and each GPU must talk to the other GPU. And the last part was the Grace Blackwell. Actually, I was there at Reinvent, I gave a talk at Reinvent. But they had the Grace Blackwell-based server on display, which is actually unique for AWS. Typically, they don't display this. It was a thing of beauty, honestly. It had 2 Grace Blackwell boards that were coming from NVIDIA with basically 2 Grace CPUs and 4 Blackwell GPUs and 9 Nitro NICs. That's all you could see on the board. So that's all I can officially talk about. But that kind of gives you the idea of the application where you are trying to customize the -- and deliver the power of an NVIDIA GPU or NVIDIA-Grace Blackwell combination into an infrastructure into a form factor that is suitable for your infrastructure. Now in order to connect 8 or 9 NICs to the 4 Blackwells, you need a solution that is like a Scorpio switch for scale-out connectivity, where you can take GPU on 1 port and NICs and storage and other things on the other ports. So both were very, very impressive demonstrations from an AWS and very good for us.

Perfect. Competition in PCI Gen 6 devices is sort of heating up. You've had Marvell, Broadcom, Credo all announced devices, Marvell and Credo I think talking about a better reach of their systems. I think Marvell highlighting perhaps lower power. But how do you expect to stay ahead? Or what do you do to stay ahead in the PCI retimer market?

Yes. Many, a lot of people are talking about these different power specs and latency spec. We are talking about design wins. So I think that really is the difference. We introduced our product. We saw it -- we showed it working at GTC last year. Some of these announcements came around then. I think Broadcom announced at the same time, and some of them have followed. But if you look at or at least to the extent that we see -- we don't see anybody else with a viable Gen 6 product yet, it will change. I'm sure it will change. They will get it all working and so on. Now we were able to deliver our Gen 6 product first because we are building on the legacy of Gen 5. We learned so much in the last 3, 4 years when we have been shipping Gen 5 of what it takes beyond what's written in the standard to make these products work. And all of that knowledge through our COSMOS platform goes from Gen 5 to Gen 6 and so we are able to ramp that much faster. The other thing to -- the important thing to realize is this is the time when our customers are ramping their Gen 6 based solutions. If you talk -- we've heard about Grace Blackwell ramps and so on, and if you are ready with your solution now, great, such as we are. But if you're not ready, then I'm afraid sort of the train is leaving the station. So at this point, we are a little bit more cautious back in March when we introduced this. There is a lot more confidence now that we will be able to keep our position in the Gen 6 generation, just like we had for Gen 5.

Perfect. Let's turn to the Taurus AEC product line. You talked about the ramp of the 400-gig products here in the second half of 2024. Talk about sort of how do you see that business in 2025? When do you see sort of broader adoption and the move to 800 gig on the AEC front?

Yes. No. Actually, Taurus doesn't get talked about often enough, but it's very exciting for us because it's actually going to be the second product line that ramps in volume. And we have seen that ramp happen already in Q3 of last year than a full quarter in Q4. We announced the results during our earnings call. We're so far living up to our expectations. And there is also a lot of headroom to grow at 400 gig. The thing to realize is 400 gig is still a niche application. So not everybody is going to adopt active electrical cable solutions at 400 gig. Our approach at 400 gig or our opportunity at 400 gig is differentiated as well as broad within that 1 account. We have multiple applications, both AI as well as general-purpose compute. We have multiple form factors, X cables, Y cables, straight cables, and we enable multiple cable vendors at 400. So there is some diversity within the main customer -- but nonetheless, it's still fairly concentrated. Now as we go to 800, that will change. We are seeing a little bit of the same fragmentation now happening at 800 as well, whereas I would say previously we figured everybody would use active electrical solutions for 800 gig. What we are finding is some people are able to rearchitect their systems so that passive cable is good enough. And if passive is good enough, nobody will use active. And also finding people who are able to -- who are not able to contain the connectivity to 1 rack, they need to go to multiple racks, in which case, optical becomes a solution. So all of these will coexist. In general, the 800-gig customer base will be broader than 400. That is true, and we'll see volume ramps pushing into 2026.

It looks today that you've got AEC certainly in the NIC-to-Taurus, sort of the first stage of scale-out network. I think Amazon and perhaps Google may also be starting to use AECs and their scale-up networks. Can you talk about the use cases? Is it -- as you look forward, do you see it in both scale-out, scale-up? Does it tend to lean 1 direction or the other?

It really depends upon the hyperscaler and their choice for scale-up. So if you -- first, if you look at scale-out, that is very consistent across all opportunities. Scale-out is typically with Ethernet. So you'll have a PCI Express based NIC. You may have a retimer connecting the NIC to the GP or the CPU or not. But when you leave the NIC, you have an ethernet cable that goes to top-of-the-rack switch. So that application always exists across the board. Scale-up depends upon the hyperscalers. So where the hyperscaler is using a protocol that's similar to either Ethernet such as, as you mentioned, Google, or PCI Express as some of the hyperscalers do. We absolutely have a play, and we are seeing the benefits of that in both our product families, both the Aries SCM as well as the Taurus SCM.

Perfect. Maybe move next to Scorpio fabric switches and UALink. Can you give us a background on how the company was able to sort of replace Broadcom on the Board of the UALink consortium? And what advantages to being on the board of UALink, does that bring to the company?

Well, absolutely. First of all, I can't tell you why Broadcom left the consortium. We certainly talk about why we got added and the net result was a replacement. So let's step back a little bit and talk about what UAL is and what UAL mean. So NVIDIA has NVLink and they've been very successful using NVLink and building larger and larger clusters with NVLink. Other people are trying to do their own way today, whether it's Google with ICI or Amazon with their neuron fabric. Some people are using straight up Ethernet, but there is a lot of diversity. So I think there is -- and there are Pros and Cons of this approach. So UAL, what it tries to do is sort of consolidate this all at the industry-wide level, where you want to deliver the simplicity of PCI Express, right? So PCI Express is a very simple protocol has been used for a lot of time. It's very easy to have 1 GPU talk to the other GPU. So they start with PCI Express protocol, simplify it even more. But then they take the line rate, the SerDes or the [indiscernible] from Ethernet and marry it up with the simplified UAL protocol. And the 2 together form UAL and it is -- could be the industry's kind of answer, if you will, to NVLink. Now so there's a lot of promise on paper. And in addition to us, we are the promoter members. Part of the reason I believe that we were invited to join is because we are already playing in the scale-up networks through our Scorpio X family of switches as well as Aries SCM. And this is also the benefit that both we get and we can provide to the consortium of delivering an overall connectivity fabric that includes switches, potentially includes retimers over copper or other different media over time. So it's a very exciting development. I would also caution that it's relatively new. So we should wait and see which of the hyperscalers start to adopt UAL and then the ecosystem will flourish around that.

When I look at scale-up networks to date, they've all been copper-based usually limited to just 1 or 2 racks. As you look to the future, I think the intent is to increase the size of those scale-up networks to get larger virtual GPUs or XPUs. Do you -- what's your sort of vision on those scale-up fabrics? Do they remain copper-based? Do they start to introduce optical solutions. And I know you haven't announced optical products yet, but to the extent you see optics in the scale-up fabrics, could that be a future opportunity that the company would look to address?

So short answer is yes. But again, let's look at kind of what the lay of the land is today. Most of the customers that we talk to are still limiting their scale-up clusters to 1 or 2 racks. And there are more reasons to do that, that go beyond copper versus optics. So that's what we see today, and we are laser focused on enabling this connectivity through 1 or 2 racks and some of them have a little bit more as well with copper. And the reason we want to do is just not because we are good at copper. It's because that's what the customers want. Customers like copper because it is more reliable. It is lower power, and it is less expensive. So my belief is that if somebody can run their scale-up networks over copper, they absolutely will. But as you correctly pointed out, as the sizes of these clusters go up and potentially data rays go beyond 200 gig, which, by the way, only NVIDIA is in that camp right now. Everybody else is just trying to get to 200 first. At some point, optical will become an important technology, and we are working very, very closely with our customers to understand what that time frame is and what solutions they need from us and that allows us time to build those solutions. So optical is definitely a great area of exploration. We are watching it very carefully. And at the right time, we will introduce solutions that makes sense. I mean, you have to understand, Marvell and Broadcom have got a strong position in the optical space, and they are great companies. So no point trying to be another 1/3 and try to take share from established players. We need to figure out where the unique opportunity is, and I do believe it comes with the scale of networks and people adopting new protocols like UAL and so on.

For you, UALink or just the fabric switch opportunity. Can you give us a sense what is that ultimate TAM or what's your kind of content opportunity as you look to build up these larger scale-up networks?

Yes. So on Scorpio overall, we've kind of outlined a $5 billion TAM in total, of which about half is for scale-out and about half is for scale-up. If you look at the piece that scale-out, that's a real market today, Broadcom shipping on or about $1 billion into that space today. But we do believe it continues to grow as accelerator growth happens going forward. But it's really that scale-up piece that gets really exciting. That's effectively a 0 market today that's poised to grow very nicely. There will be opportunities to grow across a multitude of different protocols, whether it be PCI Express, Ethernet or UALink, as Jitendra outlined. So we are very steadfast right now in terms of engaging with all those different parts of the market, engaging with the customers and building that up. But all these accelerators are going to -- going forward are going to start to support scale-up in a big way. and we are hoping to enable that with various forms of Scorpio.