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

Good afternoon, everyone. We'll get started with the next fireside chat. We're very happy to have Astera Labs here with us, relatively recent IPO. We have the CEO, Jitendra Mohan as well as Nick Aberle, the Treasurer and Head of Investor Relations. So gentlemen, probably Jitendra, Astera is a recent IPO, hyper growth story, AI connectivity, tons of exciting stuff, but might be new to some folks. We'll dive into product types and all of that later. So you don't have to go into that answering the first question. But just talk a little bit about the market opportunity you saw and how Astera fits into it and it addresses it.

Thank you, Ross. Thank you so much, first of all, for the opportunity, and thank you all for joining us here, sacrificing the sunshine outside for this room. So thank you. We started Astera Labs 6 years ago, a little over 6 years ago. So we've been sort of in the business for 6 years. And when we started Astera, it was with a very simple premise. What we wanted to do was to connect all of the GPUs together to solve models that were growing in complexity even back 6 years ago, that they will reach a trillion parameters was not something that we had predicted, but our basic premise was, hey, look, the model complexity is increasing, you will need to interconnect many, many GPUs, hundreds, thousands, tens of thousands, and now we see hundreds of thousands of GPUs together. We didn't want to be the company that created the GPUs or the AI accelerators, several folks were doing a bang-up job already. But we did want to be the one to connect them all. And that's really what we have done over the last 5 years. If you look at how these GPU clusters operate, they are trying to solve these immense problems. In order to do that, they need to collaborate with each other. In order to do that, they need to talk to each other at very, very high data rates. And that's the problem that Astera is trying to solve. We have our connectivity products that solve the problem of getting data in and out of the GPUs. And what you find with modern clusters is typically about 3 different bottlenecks, data bottleneck, getting data in and out, network bottleneck to kind of connect these big large clusters together and memory bottleneck. So these are the problems that we are trying to solve. And the culmination of all of these is the GPUs are -- the clusters are only utilized about 50% at best, sometimes even less than that. And that's a huge problem for the industry that we are trying to help our customers solve, deploy these complex clusters into their data centers. Over the last 5 years, we've released 3 different product families as well as our COSMOS software suite. And together, we call this our intelligent connectivity platform. So we have our Aries Retimers. That's probably the one that's most famous for PCI Express that solves the data connectivity bottleneck. We have our Taurus Smart Cable Module. It does for ethernet what Aries did for PCI Express, that solves the network bottleneck. And then we have our Leo CXL memory expansion products and that address the memory bottleneck when it comes to general-purpose compute as well as AI. And of course, we are not sitting idle. We are working on many new products. We introduced our Aries Smart Cable Module recently that expands the reach of PCI Express to 7 meters and working on new product families as well. And actually, maybe if I make just one more comment there. What has happened over the last 5 years is this is very unique. I've been in this industry for 30 years. It's very unique, what we've managed to accomplish is to really become a trusted partner to our customers whether they are the AI platform providers who are innovating at super rapid pace. NVIDIA said they're going to be introducing new families every year, all the hyperscalers. They come to us now and say, look, we are planning to deploy our next generation of ASIC accelerator or a third-party GPU. Here are the connectivity solutions we need. And Astera, you guys can do chips. You guys can do hardware. You guys can do software here, build the solution for us. And that is very, very unique to us. And not only have we deployed the hardware, which we are very proud of, but also our COSMOS software. And for those of you new to Astera, what COSMOS software does, it really harnesses the power that we have built into our hardware solution, whether it is for performance, diagnostics, observability, fixing issues that might come up and enables that -- enables our hyperscaler customers to use the COSMOS software into their own operating stack. So they use this COSMOS software to customize the solution, optimize our performance, have debug features, have diagnostics features. Effectively, our chips becomes the eyes and ears of the connectivity infrastructure with the COSMOS software. So very proud of what we've been able to do. The culmination of all of the hardware and the software, the results that we've delivered. We've had 2 quarters as a public company now. Last quarter, we delivered $77 million in revenue, 78% gross margins. We delivered $0.13 EPS non-GAAP. So clearly, focusing on building new products, servicing our customers and delivering great results, and we are just getting started.

Thank you for that overview. Let's -- before we get into some of the specific product types, to the extent you were prescient enough to see some of these connectivity needs in the AI market, at least directionally true, where do you think we are in the AI cycle overall? The sustainability of it, the need for a digestion period or not seems to be a very active debate. From your perspective, where are we?

I can say, definitely, we are early in the AI cycle. So we are not seeing any signs of slowdown. However, I will say that with any new technology, especially when that's evolving as rapidly as AI is, whether it's third-party GPUs, merchant GPUs or the ASIC accelerators, there will be ups and downs. There will be periods of very rapid growth, and there will be some periods of maybe not so rapid growth. My crystal ball is not much different from yours. I can't really predict when these happen. But what we are focused on is what we can do. So we are working heads down to make sure that our customers can deploy all of this technology in the fastest possible manner. And whether the market is going up or the market is going down, our job is really to make sure we deliver results in outsized results in whatever the market conditions will be.

One other evolution amongst many in the AI market is the type of processor to accomplish that. You mentioned about kind of the ASIC side of things. Does Astera care if it's a GPU or an ASIC?

No, no, not to the first order. Our content varies from platform to platform, whether it's a GPU by 1 AI provider versus GPU by another Air platform provider or ASIC, the content will vary. But we are in the business of connecting all different types of GPUs and accelerators. And by the way, NIC devices, storage devices, what have you, whatever it takes to build a complex AI server and deploy it, we are in need to connect them. The difference sometimes comes in which part of the system we play in. Typically, for some of the GPUs, let's take NVIDIA as an example, they use NVLink to cluster the GPUs together. There's a close standard, a proprietary standard. We don't play there. If you look at the ecosystem outside of NVIDIA, companies would typically use either PCI Express or Ethernet or derivatives of these protocols to cluster their AI accelerators, ASICs or other GPUs together, and that's an additional opportunity for us. And typically, our content is even higher in those platforms than it is in some of the other ones. So at the high level, it doesn't matter. But based on which one will ramp when, our content may vary.

So we'll get into more aspects of the question I'm about to ask when we get into Retimers themselves. But how much customization do you see of these platforms where the delta in your content matters more by what the CSPs choose to do and the topologies that architectures they choose rather than is it ASIC, is it GPU, is it NVIDIA, is it AMD, et cetera?

Yes, great question. So I think first thing I would say is no 2 CSPs that are alike. They're all trying to solve the same problem of enabling -- deploying AI and making it available to their customers. But everybody does it differently. And each one is big enough that they can demand a custom solution for -- to suit their needs, whether it is power delivery, cooling, how they make the -- how they monetize the solution that they are about to deploy, what software infrastructure they have put in terms of the smart mix that they are using. So what we have learned is no 2 clouds are the same, and they're all big enough to require customization of the solution. What it means to us though is we can't be the ones producing a specific solution for each cloud vendor. That's just not very practical. So the approach that we have taken is to build our products on the ground up with a software-first architecture where we have a lot of micro controllers that run all of the protocol features and make our products really what it is. And by using these -- the software that runs on these micro controllers, what we call COSMOS, we are able to customize our solution to the requirements of each and every hyperscaler. And so you will see that the solution we give to CSP 1 or CSP 2 is based on the same product, but customized to their application.

So let's talk about the Retimer business, your Aries product line. It's still 85%, 90% of your revenue, I think, this year, growing a couple of hundred percent. So very, very powerful growth and impressive growth. For this audience, please explain what does a Retimer do? And what's the sort of content that people should think about in a typical AI server?

So for the record, I've been building Retimers for more than a decade. So sometimes they get too much into the nitty-gritty. So let me try to give an analogy that I used once earlier in the year. What the Retimer is able to do -- what the Retimer's purpose is to enable communication between 2 entities. It could be a CPU or it could be a GPU, it could be a CPU to a NIC or GPU to NIC or what have you. The Retimer is kind of like -- maybe the analogy is to use a relay race. So if you're going to be running a sprint, you can run very fast for 100 meters and kind of finish that in record time. But if somebody asks you to run a sprint over a marathon distance, that's not possible. So what you do are one of the things that you can do is look at the analogy of a relay race. You run, you sprint 100 meters, pass the baton to somebody else, run another 100 meters, then you can complete the whole marathon that way. And this is really what a Retimer does. When you launch a signal at a very high rate, it gets corrupted, it gets kind of tired much sooner. And so you put a Retimer, what the Retimer will do is to recover their signal, understand what the message is and retransmit a completely fresh copy of the message. And now you can go another distance and once the message gets corrupted, you can put another Retimer there. And so we might even see multiple Retimers in a link. So Aries family does this with PCI Express. And PCI Express, for those of you who may not know, is really the nervous system of an AI server. All of the different peripherals have evolved to connect to each other over PCI Express. So we see this running around everywhere, and that's where our Retimers get used. And our Taurus family does the same thing for Ethernet for connectivity between accelerators or sometimes from a server to the top of the rack switch, et cetera.

So people have talked about the need for a Retimer potentially lessening as some of the AI architectures are going more systems approach. The GB family from NVIDIA, AMD just bought ZT Systems, so they're going to try to go down the systems route. Talk about the role of Retimers in these new kind of more system approach from not your direct customers, but part of the supply chain.

Yes, very good question. The need for Retimer is actually driven by they need to connect multiple GPUs and accelerators together. So long as you have a cluster that is hundreds, thousands of these accelerators connecting together at increasing speeds, you will need the Retimers. Where they go will keep changing based upon the unique architecture. So you mentioned Grace Blackwell. In the Grace Blackwell architecture, the CPU and the GPU are right next to each other. You don't need any Retimer to go in between, the signals are easily able to make their way from the CPU to the GPU. But what about the NIC? Who is going to drive data into this combination of Grace and Blackwell? The data has to come over the network. That is connected through a NIC. So now you need to go from this combination of race backward to the NIC. That is our PCI Express. Well, you may argue that, hey, I can bring the NIC close to this combination as well. Now you don't need any connectivity, any Retimers between the NIC and the Grace Blackwell. Well, then the NIC has to go to the top of the rack switch. So it's like a balloon. You push in one side, it's just going to pop in the other side. So long as we have new platforms coming up which are driving the data rates higher and higher, just like Blackwell platform does from PCI Express Gen 5 to Gen 6. Ethernet is going from 100 gigabit per second per lane to 200 gigabit per second per lane, the need for the Retimer type products will continue. Now exactly where they go, who's buying them? Is it the AI platform provider or the CSP? All of that changes, but the need is there.

So investors seem to try to make it a black-and-white discussion in either or, but yours is the mix will be different. The use case will be different, what you're connecting will be different, but the need for it will persist.

Absolutely.

And so how do you differentiate in Retimers versus the competition? I know you talked a little bit about the COSMOS software side of things. But just talk about it's a standard PCIe, how do you or Astera differentiate versus what Marvell reports after the close tonight, you have the guys from Broadcom coming at that market too, what makes Astera special?

You say something, let me answer that first, and then I will answer the competitive question as well, which is investors like to see it black and white. So maybe if I may kind of help understand the situation, what the mix of products will be, how exactly Grace Blackwell or Blackwell gets deployed or for that matter, other products get deployed, that's not really our call to make. NVIDIA is going to do what they will do, CSPs will choose to deploy it one way or the other. But where we do have visibility is into what designs we have won, what backlog we have. And this one I can say with confidence that with the Blackwell family, we have a backlog in place where we are going to be selling our different product families, including the Grace Blackwell, where our content on a per system basis is going up from kind of hundreds of dollars. If you look at a hopper generation at GX Board to a rack-level solution where we are in tens of thousands of dollars of Astera content. So Blackwell is great for us. We love it. The visibility we have is not based on a guessing game, but based on the backlog and preproduction shipments that we have done. And the same thing is actually applicable to an even greater extent with ASIC platforms, where as we discussed earlier, we have even greater content. Now to come to the competitive question. Contrary to what you guys might have heard, we were not the first ones out of the gate with the Retimer solution. When we first got our Retimer solution out was July of 2019, there were already 3 or 4 other companies that were trying to also execute on a Retimer solution. Fortunately, for us, based -- because of the architecture that we have, we survived, all those other companies have gone. And the reason for that is rooted in the architecture and our understanding of the problem. Most of the companies are building the Retimer to solve the retiming problem that we just described. Signal doesn't go from here to here. And so you need a Retimer in between. We did that. Other companies did that as well. But what we also realized is the deployment complexity is going to be rather large. These clusters are very complex. And the customers will need -- will have a lot of need for deep diagnostics, ability to understand what is going on with their clusters. And so we've built a lot of these diagnostics features in and, as I mentioned, enabled it over COSMOS. And so what we're able to do with COSMOS is when a problem arose in the field and there were many, by the way, not just for us, but for our competitors as well, we were able to find out what the problem is in a matter of days and solve it in another few days. So given a week, we figure out one problem, and we are on to solving the next one. So we were able to iterate very, very quickly, which our competitors were not able to do. So that is what got us into this leadership position. Now you may say, look, the new competitors know what you're doing, and we have certainly talked about it openly. And they can copy that architecture, sometimes imitation is the best form of flattery. So even if they do that, though, over the last 3, 4 years, we have learned a lot by being in the trenches. By understanding in the PCI Express standard, what works, what does not work, thousands of pages of spec, everybody implemented just slightly differently. With our COSMOS software, we can solve these problems that arise in the field, understand what those are, what are the workarounds needed, put that back into COSMOS. So over time, COSMOS has become very rich. And any new entrant will have to spend that soak time. I mean there is no shortcuts to experience. They'll have to spend this soak time to get not only their hardware up to par but also their software up to par, and then convince a CSP that they should deploy yet another software standard into their operating stacks. So over time, it will happen. Meanwhile, if you look at the reality of the situation, PCI Express Gen 5 is already done, right? The designs are already [ won ], et cetera. PCI Express Gen 6 is very close to getting there. NVIDIA has mentioned that they are going to start shipping in Q4, Q1 and so on. And so if somebody is not ready with the PCI Express Gen 6 solutions, they will again find themselves kind of going after a much smaller piece of the pie.

So going from Gen 5 to Gen 6, in and of itself doesn't really change the core value that Astera can bring relative to the competition.

And not only does it not change the core value, it actually strengthens the proposition that we have, where COSMOS is already deployed. For our CSPs, the easiest upgrade path is to simply go from Aries Gen 5 to Aries Gen 6. It is fully backwards compatible. They know the device. They know how to use it. The software is already there, everything just works. So in many ways, I would say that when we go from PCI Express Gen 5 to Gen 6, and our customers have said this, it is our business to lose. So long as we keep providing good parts, we keep supporting our customers, this is our business.

And is the Retimer itself come up as an area that they're trying to squeeze that cost out of the equation. You mentioned earlier that the GPU utilization can be as low as 50% at times. It seems like enhancing that would be a much better TCO move than focusing on getting rid of Retimers, but am I correct in that?

Absolutely. First of all, in any pricing negotiation, nobody should walk away happy. Everybody should have some level of unhappiness in the pricing discussion. But when a design-in actually happens for these Retimer type products, pricing is not the first thing they talk about. It's not the second thing they talked about, probably the third or the fourth thing. First, decision point is always, are you able to make my infrastructure work which is to say, do you have the right performance? Can you make the signals run point A to point B at the right speeds that I want? That's one. Because if you don't have it, then you don't have a product. Second one, which is absolutely important, is how robust is that link. And that's where we have spent all of the last 3, 4, 5 years in improving the robustness of these links through the hardware that we've developed and the COSMOS software. So it is absolutely critical that these links remain up at the highest speed all the time. What you don't want is I have a GPO that's running at PCI Express Gen 5, but every so often is dropping packets. Like we discussed earlier, the whole GPU cluster will run at the speed of the slowest GPU. So the hyperscalers, data center operators pay very close attention to the health of their system, and that's what we really enable with the products that we have. So robustness is #2. After that, you get into other kind of ancillary things, including pricing.

So last question on the Retimer side before moving on to the other 2 sides of the business. Some of those potential competitors, aspiring competitors also have an ASIC business. They also have big networking businesses. So to the extent you have to understand what the processors are going to do, understand what the networking is going to do and kind of the way you outlined it before. Do they have a distinct advantage in that, that they would be able to bundle and have a better shot at getting into the Retimer market?

Not for the type of customers that we are talking about. The hyperscalers that we are talking about are big enough, they have kind of enough muscle to kind of not fall prey to some of these bundling techniques and so on. If you look at technically, is there an advantage because they are building the ASIC versus maybe they're also able to build Retimers and such, I would say there is not that much because just the very nature of Retimers is that it needs to talk to one device on one side and a different device on the other side. So the fact that you're able to talk to perhaps your own device on one side is only part of the equation. You need to be able to work, interoperate with 10 different types of processor class devices, maybe 50 different types of endpoint type devices. And that's what we have done over the last 3, 4 years is to build a very healthy interoperability lab where we do this testing, figure out what's working, what's not working and fold that back into our COSMOS software. That's what the ASIC guys will have to do as well. And I would argue that we are better prepared for it because of our history than some of the ASIC guys.

So let's pivot over to Taurus, the AEC product. How are you guys at Astera differentiating versus some of the competition, the Credo's and Marvell's that are either in or soon to be entering that market as well.

Yes. So for -- Taurus products are a little bit different from Aries. They are doing the same thing in that they are both Retimers. Aries is for PCI Express, Taurus is for Ethernet. But the form factor in which we are selling is different. And that's also what differentiates us from the competition. Aries is sold mostly in the chip form factor, although most recently, we also introduced the Aries Smart Cable Modules. Taurus on the other end is sold as Taurus Smart Cable Modules which get assembled as part of an active electrical cable. So the difference between what we do versus let's say, somebody like a Credo does, is Credo says in order to solve this problem of reach, I'm going to give my customer the whole cable end-to-end. And there are some certain advantages to this, one of which is speed, right? They were able to come to the market sooner with their solution. But if you look at it from a customer standpoint, what they have to do now is they get to qualify one cable from, say, a company like Credo. But then in order to make their supply chain diversity work, they need to qualify another solution from somebody else. And then maybe another solution. That's what they are used to, 2 or 3 different cable suppliers. So we chose to do it differently partly because we are not experts in cables and I would argue maybe even Credo is not experts in cable. The cable expertise resides with companies like Molex and TE and Amphenol and so on. So we work very collaboratively with them. We design the electronics, the paddle card as it is called, where the Retimer chip resides and a few other ancillary components reside. And we sell that to the cable vendor. So now Molex or TE or Amphenol, what have you, they can build the active cable and put all of the R&D dollars that they spend on cable development, inventory management, put all of that to work and supply a full active electrical cable to the end customer, which might be a hyperscaler. Now if you look at it from a hyperscaler perspective, they get to qualify this only once because we work directly with the hyperscaler for every component that goes on this paddle card. Every piece of software is to the specification of the hyperscaler and customized for the hyperscalers, by the way, security requirements, software requirement, firmware requirement, link robustness requirements, all of them are per the hyperscaler specification. So they get to qualify this once. And as and when there are problems in the field, they know who to go to, which is, in this case, would be Astera because we are the ones who are providing the electronics. So fundamentally different business model, we certainly believe in our business model, but time will tell who is more successful.

And how should investors think about the growth trajectory of this business? It's relatively new to you, very, very small as a percentage of sales, but starting to go pretty rapidly. So just talk about the targets and the slope of the growth that you would expect.

Yes, maybe I'll start, and then Nick in -- can add. So like you said correctly, the business is just starting to bloom. We have some of the Taurus products selling in lower volumes for the last couple of quarters for 200 gig node, which is a fairly niche application. Later this year, we'll have 400 gig that is going to ramp. And I would argue that's also an niche application for some of these AI clusters where the cable thickness needs to be small and so on. We are addressing those with our Taurus-2 product family. Later this year in 2025 as data rates shift from 400 gig per cable to 800 gig per cable and 100 gig per lane, that is where we see a lot more deployment of active electrical cables. We will see this will go from being a niche application to a broad base where most hyperscalers are going to add active electrical cable solutions.

Yes, I mean, so the only thing I would add is when we talk about our guidance into Q3, we guided to roughly $20 million of incremental revenue quarter-over-quarter. A meaningful piece of that is driven by Taurus. And the exciting piece of that is it's not just one program going into one application. It's actually programs that are ramping on the general purpose side, programs ramping on the AI side. And even on the AI side, it's going to be connecting and scaling out both internal ASIC platforms and third-party GPU platform. So expecting a pretty nice growth rate. We had some partial quarter in Q3. So there will be some ramp in Q4. But these things take multiple quarters to kind of get to volume production and full mature run rate. So I think it's going to be a pretty nice driver for us.

And does the kind of accelerator/GPU versus CPU crowding out effect, AI servers versus general-purpose servers -- does that matter much within this business for you? Because it seems like it will target both sides of that equation.

Okay. So yes, I mean, there are opportunities on both sides. I think the good thing for us is we play on both sides of the equation. We approach the market with various applications, whether it's going to be a straight cable or white cable or a crossover x-cable. So we're really there to support whatever the customer is trying to accomplish. And again, kind of just going back to the kind of core premise as speeds continue to increase, these distances continue to get harder to traverse with good fidelity with the signal. So those attach rates will go up. We're seeing that happen at 400 gig, and then we'll really start to see it happen on 800 gig.

And as those speeds go up, is there the risk that there's even more intelligence added to the ends higher level of processing that would be beyond what you're targeting when people start to do some of the PAM4 side of things?

We're already doing PAM4. The Taurus-3 product that we have does 800 gig PAM4. So any -- contrary to some belief, these are DSP chips. They have DSP in them. And a case in point is we took the DSP, we took the flexibility of our COSMOS architecture, and we drove PCI Express over optics, right? So we're doing what a traditional optical DSP would do. So these are all very smart and heavy-duty DSP chips already.

Got it. So the final of the 3 segments that you've talked about is the CXL business, the Leo business. That one is a little bit more nascent. And so -- versus the other 2, where the market's big and people are more or less arguing about the growth rate of an existing market and the share you take. This one is a little bit more nebulous as to the size of the market itself. So explain what CXL is going to mean as a market and where Astera plays in it, if you could, please?

Yes. Perfect. So CXL, for those of you who don't know, is a new standard that's written on top of PCI Express. And amongst other things, it allows a CPU or a GPU to access memory over a CXL link. So this was something that was introduced, I think, in 2019. And so -- and we've been working very, very closely even before CXL became an open standard. So we really believe in CXL. What we have been able to do over the last year is actually demonstrate the problem that we are solving with CXL using Intel's Emerald Rapid CPU and equivalent CPUs from AMD as well, where we can see application level performance speed up 50% because of the additional memory bandwidth and the additional memory capacity that's available. Same thing goes for database applications themselves. So the ROI is clearly there, and we have very clearly demonstrated. But as you say, this is still a nascent market, for 2 reasons. One is the CXL capable CPU has got delayed. So what was supposed to happen in the Sapphire Rapids generation is now happening with the Granite Rapids and Turin generation and the equivalent CPUs for ARM. As these CPUs get deployed in 2025, we will see this product family ramp as well. The other factor, frankly, is just the growth in AI. People had a finite infrastructure budget and all of that got sucked into AI. So as people start to now refresh the servers from the 2 generation or Ice Lake servers to now Granite Rapid servers and equivalent for AMD, we will start to see more money going into general-purpose compute, which helps deploy DOC XL family.

And how is the competitive landscape differ in this market versus your other 2?

This one is actually the competitive mode is even stronger than any of the other ones. And the reason for that is the following: it's a little bit more nuanced. People talk about CXL all the time. And absolutely, CXL is necessary to enable this application. You connect to the CPU or even the GPU over CXL. But I will say that the real knowledge is in the memory side. So on the other side of this product, which by the way, is CPU, CXL connects to a memory expanded device like our Leo and then there is DDR5 DIMMS on the other side. For those of you who know DIMMS know that DIMMS will fail. The memory will fail, it's just a matter of time. It's very well understood in the industry. And how people deal with memory failure is very different. Intel will have a different approach. AMD has a different approach. ARM CPUs have a different approach, hyperscalers have a religion around how they manage memory. And what we have done is we've learned over the last 18 months, what this religion is, what needs to be done for each CPU platform and each hyperscaler and again, build that into COSMOS. And that is really driving a very strong competitive differentiation compared to anybody else that's out there.

And how do you think the key killer app for the technology is going to be more on the expansion of the pooling side of things? And how does the time line differ between those?

Yes. I think CXL is very much one of the crawl, walk, run type of situations. We believe that memory expansion for general-purpose compute is where it will start. That will be the first volume deployment of CXL. And we know this because we have hyperscalers who are doing rack level deployments now to [ pipeline ] the system in preparation for the mass deployment with the new CPUs. So that's where it will start, also because we are able to do this without requiring any software changes from the application. So we can bring the goodness of CXL with 0 software changes from application. So that's where we get started. Now over time, people will be able to customize these applications and then drive additional performance gains for database applications, I think that's where it will go next for expansion. And eventually, when you have CXL 3.0 and CPUs that are capable of CXL 3.0, which is probably 2026, '27 time frame is when we will start to see CXL memory pooling applications. It will be great news for us. We can actually support pooling now. It will be additional content, not just Leo, but also the Retimer products. But in order to make that happen, you do need software changes. So I'm not really holding my breath to be honest with you.

So the key metrics there are hopefully the CPU vendors are on time with their general-purpose server CPUs, the Turin and the Granite and the Diamond after that.

Correct.

Great. In the last minute or 2, let's talk -- wrap things up with a little bit of the financial side of things. The company is already significantly profitable, which is quite amazing at this stage in your life. But ironically, Nick and Mike on the CFO side, spend a decent amount of time explaining why your gross margins have to come down from that scarily high 70s to maybe 70% over time. Just talk a little bit about how are they as high as they are and what drives them down to a still ridiculously impressive 70%?

Yes. I mean -- so I mean, clearly, the gross margin where it stands today is a testament to the value that we deliver to the customers. And as Jitendra talked about earlier, COSMOS is a big piece of that. When you're giving them the capability of managing their fleets in a proactive way and saving money on these huge investments they're making, there's a lot of value in that. So we're able to capture that. So looking out over time, as we've been saying, product mix really will be the biggest driver to the gross margin trajectory over time. Today, a vast majority of the revenue is driven by silicon only or stand-alone silicon shipments. But we do have some pretty compelling opportunities here to grow the business through hardware-based solutions as well, whether it be with the Aries SCM product portfolio or the Taurus SCM product portfolio. So those naturally carry a little bit of a lower gross margin. So as those blend into the mix over time, you'll see the margins start to come down. We have a long-term margin target of 70%. We guided to 75% for Q3. And to the extent that that mix kind of moves more towards hardware, you'll start to see it glide down towards 70% over time.

Got you. Well, guys, we are exactly on time. Thank you so much for joining us here in beautiful Dana Point and explaining the impressive Astera story to us.

Thank you so much. We're just getting started here.

Thank you.