Okay, great! Why don't we go ahead and get started? Welcome to the second day of J.P. Morgan's fifty-second Annual Technology Media and Communications Conference. My name is Harlan Sur. I'm the semiconductor and semiconductor capital equipment analyst for the firm. Very pleased to have the team from Astera Labs, Jitendra Mohan, Chief Executive Officer and Co-founder; Mike Tate, Chief Financial Officer, here with us today. As you know, we helped bring Astera to the public markets in March. Company is a leader in accelerated compute connectivity, networking, storage solutions. Their silicon and software solutions are integrated and powering 90% plus of the world's AI compute servers and clusters. I've asked Jitendra to start us off with an overview of Astera Labs, and then we can go ahead and kick off the Q&A.
Gentlemen, thank you for joining us, and, Jitendra, let me turn it over to you.
Yeah. Thank you, Harlan. Yeah, you asked me to keep it brief, but if you know, we started the company a few years ago, so I can talk forever. So please give me a sign if, you know, I'm going too late. So what is our mission? Our mission is to really innovate, design, deliver semiconductor-based connectivity solutions that are purpose-built to unleash the true potential of cloud and AI infrastructure. Now, at dinnertime, that stops any conversation dead in the track. But I think, for this audience, I will unpack this a little bit. Let's talk about the problem that's facing us today in the AI side, the problem that we are trying to solve.
I think all of you are probably well aware that the model complexity keeps increasing exponentially, doubling every six months. Recently I heard of a model that was almost 2 trillion parameters. So how do we handle this complexity? Well, what you do is you start with the fastest, most capable compute element that you can find, which is typically a GPU, and then you connect thousands, hundreds of thousands of these GPUs together to tackle the complex AI models. Now, what happens is these GPUs are so good that it creates... it's very difficult to keep them fed, and it creates other bottlenecks in the system. Namely, you get a data bottleneck, you get networking bottleneck, and you get a memory bottleneck.
The upshot of all of this is that even with the fastest possible connectivity, these GPUs are underutilized 50% of the time, and that's a huge problem that's facing the industry, and that's what we at Astera Labs are trying to solve. We are not alone. Our friends at the AI platform providers, the AMDs, NVIDIAs of the world, they're coming out with faster and faster GPUs. At the same time, our hyperscaler customers are trying to deploy these AI platforms in their own unique way that addresses the requirements that they have. The good news is, over time, we have built a very good, very trusted relationship with our partners, and as well as our customers, and we have incredible visibility into what they're trying to do.
Oftentimes, they will come to us and tell us, "Hey, go build me this solution because I know you guys can do chips, you can do hardware, you can do software." And that is really what has fed our product definitions and our roadmaps. What we've been able to do over the last five years is to build out our intelligent connectivity platform, which comprises of three product families, as well as our Cosmos software suite. So the products are Aries, which is our-
Mm
... PCI Express smart DSP retimer, widely deployed today. As Harlan mentioned, 90%+ market share, gold standard in the industry, solves the data connectivity bottleneck. Then we have our Taurus smart Ethernet cable modules, which address the networking bottleneck, and we will get meaningful revenue from those later on this year. And last but not the least is our Leo smart CXL memory controllers, which are leading in the industry right now, and we expect to see them get deployed in the 25 timeframe, delivering higher memory bandwidth and memory capacity to the modern-day compute elements.
Recently, we also introduced our Aries smart cable modules that enable us to connect GPUs together in a cluster, as well as our Aries Gen 6 retimer family that builds upon the leadership that we have with Aries 5. As you can imagine, we are not sitting idle. We are working very busy on new product families that we will certainly announce at a future date. On top of that, we have our Cosmos software suite, which really unifies and underlies all of these product families. Our, our chips are built with a software-defined architecture, which makes them extremely flexible, which allows us to to collect a lot of telemetry information, provide performance optimizations, and all of these different things that our customers require through the Cosmos software that runs on our chips.
And then our customers actually also run Cosmos APIs on their platforms to pull out all of this telemetry information and use, really, our devices as eyes and ears of their connectivity infrastructure, that they can use to fine-tune their AI infrastructure and make sure it's always operating at peak performance. So Cosmos has been great for us. It's become very robust, very rich over time, and makes it very easy for our customers to do generational upgrades of our current devices, as well as adopt new devices. So what does all of this mean for us? You know, that's a record quarter we had last quarter. So for the first quarter of 2024, we delivered $65 million in revenue, which is about 29%-30% increase over the previous quarter.
We came in at 70% gross margins, a little north of what we had the previous quarter, though Mike will tell you not to get used to those, those numbers. We delivered 24% non-GAAP operating margin, about $0.10 EPS. As you can imagine, we are just getting started. There is a lot more to come. Our customers trust us. They are telling us what new products to build, and we are building them at a feverish pace, so lots more to come.
... Oh, that was a great overview. Thank you for that. I'm gonna start off with your Aries PCIe Smart Retimer family of products, as they are the bulk of your business today. You know, just over the past few months, it seems like activity and deployments on AI infrastructure has accelerated and broadened out significantly, right? NVIDIA H200, Blackwell, all the ASIC guys bringing their custom chips to the market on a very rapid cadence, right? Google TPU Next Gen, Amazon Trainium Inferentia, Microsoft Maia, Meta MTIA, AMD MI300X, right? Are you seeing this pull forward in time to market and deployments? What's driving that? You mentioned this 2 trillion parameter model. My guess is that the model evolution continues to move at a-
Mm
... very quick pace, right? And it could just be a heightened level of competition, but are you seeing this pull forward in time to market on these deployments, and is this partially what drove the strong results for the March quarter and June quarter guidance?
Yeah, absolutely. You're absolutely right, Harlan. We see a lot of pull from the industry. And if you think about, you know, why people are deploying these very, very complex AI platforms, it is because they have to.
Mm.
The people like you and I, you know, we are demanding additional functionality from these AI models to make our lives easier, to make our lives more productive, and the only way that the industry knows how to do this today is to increase the complexity of the models. And once you increase the complexity of the models, you need more AI platforms, and hyperscalers are on a mad rush to really deploy them in their own unique ways. And we have ringside seats to see these deployments happen. We understand what the challenges are. We can incorporate the solutions for those challenges into our upcoming products as we have. So it's all great, and if you look at the announcements that the leading four hyperscalers have made, they have said their CapEx is increasing 45% year-over-year.
Mm, mm, mm.
So all of this does provide us a lot of good tailwinds for the type of business that we are in.
You know, I articulated all of the different XPU programs coming to market. All of them have different connectivity architectures. All of them have different server-
Mm
... cluster configurations, right? This is where I feel like, and you mentioned Cosmos, but this is where I feel like your Cosmos software and your Cloud-Scale Interop kind of help to differentiate, right? Because you're probably getting inundated with all these requests for interoperability, qualification, and so on, but maybe you can help to frame, you know, how does Cosmos and the Interop Lab really help customers in trying to drive this sort of very fast time to market?
Yeah, no, that's exactly right. So there are two problems that are basically facing us. One is the complexity of these AI platforms, which are increasing in speed, and they are increasing in complexity-
Mm
... that needs to be managed. And then there is the diversity on how the hyperscalers are deploying these AI platforms in their unique ways within their unique, requirements, and constraints. And our Cloud-Scale Interop Lab and Cosmos address both of these problems. So let's start with Cosmos first. As I mentioned, we build our chips with a software-defined architecture. What does that mean? It makes our chips extremely flexible. So when a hyperscaler wants to compose a system using the AI platform to their own unique requirements, we enable them to do that. We also enable them to have deep diagnostics information about the data that's running through our chips. We enable them to make performance optimizations that are uniquely suited to their, data center deployments.
So all of that is done using the software-defined architecture that we have on the chip, based on the Cosmos software. Then we also have the Cosmos software APIs that are running at the cloud operators'-
Mm
... stacks. So that gives them just incredible amount of visibility into the telemetry information that's coming out. Is the link to their GPU healthy or not? Is it about to go down? Do we need to deploy a software solution, or do we need to send a person to maybe tighten up a loose connection, et cetera?
Yeah.
So that's a huge advantage, and I would say it's, it's a must-have for deploying anything at scale, in the AI infrastructure that, that is being deployed in the cloud. So that's one. The other one is a Cloud-Scale Interop Lab. That's extremely useful.
Yeah.
Now, for those of you who don't know what Cloud-Scale Interop Lab is, this is where we mimic our customer systems in our lab, and we run all of the tests that they would typically run in the data center, but in a lab setting. Given the fact that we are such a leadership position, and we have such a leadership position in PCI Express, what has happened is now our customers and ecosystem partners will send us their unreleased products to our lab. So these are, like, SSDs-
Yeah
... or GPUs or NIC cards-
Mm
... that are not released, and we test them first and provide a passing interop report. Now, here is the, the, you know, not often talked about thing. Sometimes they don't work. They don't work out of the gate.
Yeah.
What we do is, we figure out which part of the spec got, you know, interpreted incorrectly, and figure out the solution to that, a workaround, and we put that into our Cosmos system. So when the end customer deploys the whole system, including our retimers, including the new endpoint and whatever root complex they have, this thing just works. But if you try to do the same thing with a different, let's say, a retimer provider, you will run into issues because they haven't had the cycle of learning from the Interop Lab feeding back into Cosmos. So this combination of Cosmos and Interop Lab has really become very, very important for us and a true differentiator for our products.
Yeah, and I think that, you know, as a part of our due diligence ahead of the IPO, talked to many of your customers, right? Number one on the list was first to market, performance of the silicon... but a close number two was the Cosmos platform and the Cloud-Scale Interop Lab, right? And a lot of it had to do with time to market, debugging, flexibility, you know, driving fast time to market on Interop. But then the other very, very sticky part of it, as you mentioned, was that they're integrating your Cosmos software stack into their data center software management stacks-
Yeah.
Right? And so this is a very, I think, sticky part of your solution. You mentioned some of them, but things like telemetry, the audience may not know what sort of terminology like telemetry is, but what are some of these capabilities that are built into your silicon and software, that the data center managers feel that it's so important to monitor these metrics on a real-time basis that, you know, that they have to have your software enabled to do that?
Yes, so first of all, Harlan, you are right. It was not easy to get here. In fact, you know, somebody that we talked to from the investment community said that getting qualified at a hyperscaler is like getting an organ transplant. The body is actively trying to reject the foreign entity that you are trying to insert. But unfortunately, and by the way, it didn't happen in the reverse order. It wasn't that people said: "Oh, your Cosmos software is so good, I'm gonna design your retimer.
Right.
It didn't go like that. They needed a retimer solution because signal integrity was a problem. The CPU could not talk to the SSD, and so you needed a retimer in the middle, and we had the right device, for them, which not only solved the signal integrity problem, but provided all of this telemetry that I will touch upon in a second.
Yeah.
So, once we were in, that's when we exposed all of the, the benefits of Cosmos software and telemetry, et cetera, that they could utilize. So to break it down for you, there are many things that the chip does, many, many types of information that we are able to collect, based on our architecture of distributed microcontrollers and lot of sensors. The easiest, first and foremost, easy to understand, is physical-level sensors. So they will tell you what the temperature is around-
Mm
... in your surroundings.
Mm-hmm.
What the voltage conditions are. Was there a spike in something that can get recorded? Then comes the electrical sensors. So since all of this data is running through our devices, we are in a very unique position to see what exactly is impacting that data. Is there a loose connector? This, you know, AC coupling capacitor come off-
Yeah
... or something like that? Are you really operating close to a cliff's edge? So our electrical sensors are able to collect this information. Then we have protocol sensors that look at it at a protocol level and see what's happening in the protocol. Are we entering into retries too often?
Mm-hmm. Mm-hmm.
These are the things that both sides of the link may not be aware of, but we know uniquely what's going on.
Right.
And then we make this information available to the cloud operators to do fleet management. As I mentioned before, they really look at this information to ascertain how close to peak performance their AI infrastructure is operating, and when there are problems, how they can deal with that on a proactive basis.
Back in March, you introduced your next generation Gen 6 smart retimer solution. Doubles the per-channel data speeds, more complicated PAM-4 signaling modulation. You developed a solution that was actually more power efficient versus your competitors, but even more power efficient than your current Gen 5 solution, right?
Yeah.
You've been sampling, qualifying. You already have, like I said, 90% market share on Gen 5. All of your AI customers are familiar with Cosmos software, the Interop lab. How is the Gen 6 eval and qual progression relative to the same period back when you were qualifying your Gen 5 solution?
Yeah, that's actually a very good observation, and in fact, the other day, I was telling somebody in the office that it feels very much like the Gen 5 days in 2019. We would pack four people in my minivan, and we would drive to one of the CPU vendor's lab because that's the only place where you could find Gen 5 for testing. So we are in a similar place. You know, back in 2019, we learned a lot by being first. We again understood it was a new spec, it were new chips.
Right.
We understood where people interpreted the spec, like, you know, we thought they would, and also where they did not, and what were some of the nuances and the warts with the different chipsets. We implemented those solutions into our Cosmos software, which is, you know, one of the reasons why we are so successful with Gen 5. The same thing is happening now with Gen 6, but there are two important differences. One, I would say, is the cloud service providers are running even faster than they were-
Mm
... back in 2019. There is just a mad rush, you know, to do the work of-
Mm-hmm
... 10 people. Hyperscalers probably have three. So, so they don't have time to kind of muck around with, you know, this thing and that thing.
Yeah.
They just want the proven solution, trusted solution, and then deploy it and move on. The second advantage that we have is our current market position. So at the time, we were a nobody.
Right.
We were just getting started.
Right.
Now, as you mentioned, we have a dominant market share, and what that has done is it has translated into a lot of this, you know, tribal knowledge that we've built over the last 4-5 years, that's now part of our Cosmos operating system, and we can bring this to bear on all of our Gen 6 designs. And just understanding what the customer wants, what's important, what's not, is very obvious in the announcement or the demonstration that we made for Aries Gen 6 device, where we showed almost 30% 20% lower power consumption compared to our competition that made a press release. To the best of my knowledge, we are still the only working solution at PCIe Gen 6, and the evaluations are taking place now.
If somebody does not have a working device now, then I'm afraid they might be too late. I think customers will likely make design decisions in the balance of this year.
Yeah.
So we feel very good about where we are in the market, and we are happy to just service our customers, make their GPUs run faster.
Although AMD and Intel won't have PCIe Gen 6 support until their CPU introductions in the second half of next year, we will see AI servers or will we see AI servers sort of future-proofing their designs and pre-integrating Gen 6 retimers well ahead of the CPU launches, right? I'm just trying to get a sense for sort of the ramp profile of Aries Gen 6 as we move through sort of 2025.
You know, that's a good question. Like you said, we are, we are kind of where Gen 5 was back in 2019. So what we are seeing now is AI, AI, you know, platform providers-
Mm-hmm
- evaluating this technology now. We expect that their deployments will start to happen in 2025, and potentially maybe a Gen 5 to Gen 6 close crossover may happen in 2026. Having said that, I think there are two things that are really driving this growth-
Yeah
Is just these GPUs or AI accelerators keep getting faster and faster.
Mm-hmm.
And in order to get them fed, you have no choice but to increase the speed of the interfaces. PCI Express is going to a Gen 6. NVIDIA has made that announcement, others will follow. And we are working very, very closely with them to enable this on the AI infrastructure. They really need this drive. Now, by the time we get to 2026, CPUs will also start to come about-
Yes
... that have Gen 6 in them, whether it's the Diamond Rapids from Intel or the Venice platform from AMD. So that will provide further momentum for Gen 6 deployment in that time frame.
Fast-growing markets are always gonna attract competitors, right? And we've seen competitors coming in at Gen 5, now at Gen 6, right? Investors are concerned that customers want diversification of suppliers. But I'm wondering, especially in AI compute, where the cadence of new product introduction is so aggressive, performance and reliability is such a high priority, if supplier diversification is no longer a high priority... In other words, as long as my current supplier has a strong record of execution, delivering products to me at the time that I need them, from my perspective, like, there's just too much risk to bring on new suppliers with unproven software. I just wanted to get your views on that.
Yeah, I mean, otherwise, just to summarize, there is what people want to do-
Mm-hmm
... versus what the time to actually do.
Right.
So maybe I'll give three things. First and foremost, it is really the onus on us to make sure that customers don't actively want a second source solution.
Right.
We need to give them no reason to worry about our solution, and which is exactly what we have done. We've supported them incredibly well when it comes to technical support. I mean, these are complex systems that they are deploying. We've given them no reason to worry about when it comes to the manufacturing supply chain. We have a very diversified manufacturing supply chain that is designed to manage the sort of the, kind of the crazy ramps that hyperscaler customers have. And we are investing a lot into our roadmap devices to address the challenges that will come. So that's one point I wanted to make. The second thing is, the connectivity business is interesting in that there is a huge first-mover advantage.
Once a hyperscaler has figured out their connectivity infrastructure, they don't go back to tweak it and optimize it, not for a couple of dollars. That's not the driving factor. What is important is robustness. So the first movers typically end up winning a lion's share of the market. We have seen that happening before. We are in a good position to kind of see that repeated, as you mentioned. And the last one that I will say is we have to stay paranoid. We can never take our competition for granted. You know, we have a big bull's-eye on our backs. There are people who are coming after us, and we just need to stay ahead. And, you know, the announcement that we made at GTC is a good example, where our most recent product is more suited, Gen six-
Mm-hmm
... the retimer is more suited for these applications, better in power consumptions, and we are adding a lot more of the fleet management features that we talked about based on what we have heard from the customers. You know, competition will take some time to get the same feedback and incorporate it.
You introduced your Aries PCIe AEC smart cable solution, right, at the beginning of this year. And it's pretty amazing because you've already secured design wins that are gonna start shipping volumes to your customers in the back half of the year, right? The smart cable opportunity opens up more content capture in an AI cluster because it's not just GPU to CPU connectivity or GPU to SSD or NIC card. The smart cable solution now enables GPU to GPU connectivity across the racks. I believe today, the current, sort of the way we think about it, is retimer to GPU attaches somewhere between 1-2. If you capture the Aries smart cable for back-end GPU to GPU connectivity, how much does it increase the retimer ratio, you know, the retimer GPU to attach ratio?
Let me make a couple of comments, then I'll-
Yeah
... answer your question. First of all, very happy with the trajectory that we are seeing with the Aries smart cable modules. You know, great testament to a couple of points that I made earlier, where our customers trust us, and in this case, they came to us and say: "Hey, look, in our architecture, this is the solution that we want to deploy. We know you guys can do chips. We know you guys got..." You know, they are already using our Aries retimers. We have our Taurus smart cable module family, so we know we can do the hardware.
Right.
And we have the software to make it all work for our customers. And also a great testament to the team, who are really able to, you know, being very nimble, executing like crazy, and able to deliver this solution record time, to the point that we, you know, are looking to get some good revenue from this-
Mm-hmm
... towards the end of this year. I think that's all very good. Now, if you look at it in terms of the attach rate, so, you know, unfortunately, we won't comment on specifically-
Mm, mm-hmm
... what happens, you know, in this particular case. But the comment that I'll make is, the GPU-to-GPU connectivity or the back-end scale-up cluster connectivity is a little bit different from the front-end connectivity, in that it's very dense. Every GPU needs to connect to every other GPU.
Yeah, that's right.
So you end up with a lot more links, and typically, these are running at the fastest possible data rates. So from that standpoint, the attach rate goes up pretty significantly, but then you have to average it over the whole portfolio or all of the AI systems that are getting deployed, wherein some places we do both backend and front end, other places we might only do the front end.
But I think-
On the whole-
Yeah
... we do see the Retimer content go up, so very excited about this.
Yeah, 'cause that's what I was gonna say, right? Because in the earnings call, you said it several times, right, that over the coming generations, you just see that your retimer content is gonna continue to move higher, right?
Yeah.
Whether it's Aries or whether it's the smart cable, I mean, I think that's what you're sort of referring to.
Yeah, there is actually incredible interest from the customers for all of our products right now. I would say it's a record high-
Mm
... in terms of just all the design activity that is going on. But if you talk about what is driving this activity, is the secular trends of increasing speeds, right?
Yeah.
These GPUs and AI accelerators are getting faster, increasing complexity. Then the third one, which is also interesting, is not technical, but the pace of deployment is increasing.
Yeah.
The way we have built the company, the way we are executing, how nimble we are to our customers' request, makes us the perfect partner for our customers to come to us for these solutions.
A few of your GPU and AI ASIC chip partners have developed sort of proprietary protocols, right, for their GPU-to-GPU connectivity. Some of these protocols, like AMD's Infinity Fabric is the one I can think of, right, use actually, it's a proprietary protocol, but it looks very much like a PCIe-based protocol, right? Is the team developing custom-tuned smart cables and retimers that can run these PCIe-like protocols?
Yeah, so, so I would say, first of all, that customization is really central to what we do. This is, this is what the Cosmos software enables.
Mm.
It allows us to customize the solution to the hyperscaler based upon their unique requirements, and the performance optimizations that they want to do. So we've been doing that already, and, to the extent our customers want it, we will continue to do that. Now, you're also correct that, some of the protocols that people use for GPU-to-GPU connectivity are based on PCI Express-
That's right
... and on Ethernet.
Mm-hmm.
Now that we have the Aries smart cable modules for PCI Express, and we have the Taurus smart cable modules for Ethernet, we can address both of these requirements for our customers. At the end of the day, we won't just stop there, right? We are just maniacally focused on our customers. If they want a particular solution from us, and again, you know, if the returns justify it-
Yes
... and so on, we will absolutely enable it.
Yeah.
You know, likely, and, and so far it's been mostly software, but if it requires us to build a new product-
Yeah
... we will absolutely do that.
Absolutely. Before I move on, are there any questions in the room? If you do, raise your hand and we'll get a mic over to you. Oh, we've got one right up here.
Where do you see PCIe kind of evolving from sort of within the node to kind of sort of across nodes in rack cluster? Do you see that coming anytime soon, or are we still gonna be stuck with Ethernet and InfiniBand?
No, absolutely. It's a great question. So PCI Express is basically the nervous system of any AI server. If you look at within the AI server, PCI Express runs everywhere. And what we are starting to see now, and as evidenced by the Aries smart cable modules, is our customers saying, "Hey, PCI Express natively gives me the ability to address the memory of other GPUs," which is what you need to run these AI workloads. So people are looking at PCI Express as that connectivity, and that's what our Aries smart cable module enables, because we can build seven-meter-long cables that allow you to connect two racks, GPUs in two racks together. Now, at some point in time, I think customers will want us to connect three and four and other racks, and we'll have to look at the appropriate solution then.
But we absolutely see the benefits of PCI Express in the GPU-to-GPU connectivity. The other thing is a little bit nuanced, is if you think about it, if you had PCI Express, just because it's, like, so central in AI servers, you can attach other things to it. You're not locked into a particular, you know, protocol. So I think that's another advantage. But having said that, some people will go PCI Express, we'll support them. Some people will go Ethernet, and we are equally happy to support them. And there are others, like, for example, NVLink. We don't-
Mm
... we don't. It's a proprietary standard. We don't play in that market. We haven't in the Hopper generation, and we don't in the Blackwell generation. So that's likely to continue.
I think... Oh, right here.
I wanna go to the, the topic you said, just all these secular growth drivers, increased complexity, speed, you know, history rhymes, it doesn't repeat. You know, how do you have confidence in five years that there's not an overbuild of AI data center capacity, and this isn't a repeat of the telco fiber build out of the late 1990s?
First of all, I don't think my crystal ball goes out five years, so it'll be hard for us to kind of predict what happens. But what I will say is, we are still in the early innings of AI, so the way the AI model complexity is evolving in response to what we are asking it to do, I think that's a decades-long phenomenon. That, I don't think there is any sign of, you know, that slowing down over time. But as is true with any new technology, you will have, you know, highs and lows. There will be a fast growth rate at some point in time, and then maybe a digestion phase. When this will happen, I can't predict. What we are seeing in our business is continued ramp.
Even, you know, we talk about Gen 6 a lot, but even Gen 5 products are just ramping. Some of them are just getting started. So, you know, we will not be immune to whatever happens in the market, but what we can control, what we do, is to make sure that we grow faster than where the market is growing, whatever its growth rate might be. And we do that by focusing on our products. That's what we are doing. We are, you know, creating new generation of our devices. We are working on new products that are going to be very, very important to our customers. And these are, by the way, the customers that can really write the big checks. So I think we have our wagon hitched to the right customers.
We are in the fastest growing, you know, market that industry has seen in a while. So, and so we will continue to chug away. And what happens in five years, I guess we'll find out. There was a question back there.
Yeah.
Can you talk about competition and anything you're seeing from Broadcom, and to the extent that your buyers are looking to dual source?
Yeah, so I think as we talked about earlier in the conversation, competition will come. It's a big market. People have seen how successful we have been, and so they will definitely try to come and take share from it. The two things that will make it difficult is, one, the stickiness of our solution, right? Our customers know our products, they know how they work. The software that we provide, Cosmos software, is already integrated into their operating stacks, so they will need to have, you know, time and energy and resources to adopt yet another solution. And that's gonna be hard given how fast the hyperscalers are running, so that's one. The second thing that I'll say is, you know, we have to make the assumption that our competition has equally smart engineers as we do, right?
Eventually, they will get the right product, but they do need to get the right product. And once they get the product working correctly, then they need to have some soak time to understand all of these nuances of the protocol that we talked about and implement them in their software, whatever software they provide. And then they go to the customer and say, "Hey, here is my chip that—look, it works very well. Here is my software, every bit as good as Astera's Cosmos." And then you have to question, you know, what motivation does the customer have to integrate yet another piece of software? So you know, having said that, we have to stay paranoid. Like I said, so far so good.
We see us as the only working solution at PCI Express Gen 6, despite the other announcements that have happened, and we'll continue to work hard to stay ahead of the competition.
We're running out of time here, but I did wanna touch on some of the other product categories, especially Leo CXL, 'cause the opportunity here, looking out over the next few years, is actually quite large. From a customer perspective, given the varying complexity of workloads being supported per servers, right? There definitely appears to be sort of building demand for server architectures to be able to flex up and flex down memory usage. You combine this with the launch of Granite Rapids and Turin x86 CPUs in second half of this year, that finally have production-ready CXL support. Has the activity level on CXL picked up? How does your customer design win pipeline look, and confidence on customer deployments of your CXL solution next year?
Yeah, I think what I will say is the, at least to us, the ROI for CXL is very clear. It allows us to break, or allows the customer to break the fixed ratio of, of compute to memory. That's really the benefit that CXL offers. And we have demonstrated 50% performance improvements at an application level with no change in software. So that's really the promise of CXL.
Mm-hmm.
I think the ROI is very clear. What's not clear is the I of the ROI-
Yes
... where is the investment? And just as, you know, customers were looking to deploy CXL, the AI wave happened. We are very thankful for it. But yet it did suck away the dollars from-
Right
... general purpose compute into AI. So as we see those dollars come back to general purpose compute, as we see the new CPUs come about, we are very optimistic that CXL will become a kind of a mainstay in the data centers. And all of the CPUs will support CXL, and all the hyperscalers will deploy it.
And then back to Taurus, your Ethernet networking connectivity solution. You're ramping your Taurus AEC 200 gig applications, some 400 gig in the second half of this year. The move to 800 gig in second half of next year should drive a more material revenue inflection, right? How is your design win pipeline on 800 gig? Do you expect an expansion of customers for your AEC solutions for 800 gig?
Yeah, so like I said, I think the design pipeline for all of our products remains exceedingly strong. What happens with 800 gig in particular in the 2025 timeframe is it becomes a more broad-based deployment.
Mm.
What we see now at 200 gig and 400 are more kind of niche applications to do with AI servers and then the particular constraints that the hyperscalers have. But at 800, which runs at 100 gigabit per lane-
Yeah
... it becomes a physics problem. You cannot shuttle that data from the bottom of the rack to the top of the rack. And so we'll start seeing more applications, you know, more hyperscalers deploy it, and get a bigger base for our 800 gig products.
And then, Mike, for you, gross margin guidance, 77%, that's tracking 700 basis points above the long-term model of 70%. I understand you will be driving a higher mix of hardware products over the next several years, which could put pressure on gross margins, but at the same time, you've got new product refreshes like Aries Gen 6, that also carry higher ASP. So help us understand how to think about the puts and takes in your gross margins, kind of mid to longer term.
Sure. Our standalone gross standalone ICs garner very good gross margins. You can see historically, we've been mid- to high 70s%. We see that continuing in our markets we serve. But when we do hardware modules, those generally carry a lower gross margin, so we have guided folks to expect our margins to trend towards a longer-term model of 70% as the hardware ramps as a mix. Now, to the extent we introduce new products that are more standalone ICs, those will be favorable as well. So we feel comfortable we should be able to maintain at least 70%.
Great. Well, we're just about out of time. Jitendra, Mike, thank you very much for the participation.
Thank you.
Great insights.
Thank you.
Thank you.
Thanks, Harlan.
Yeah, thank you.
All right. Thank you, guys.