Great. Morning. Thank you. Welcome to the 2nd day of JP Morgan's 54th annual Technology and Media Communications Conference. My name is Harlan Sur. I'm the semiconductor analyst here at the firm. Very pleased to have the team from Astera Labs. We've got Jitendra Mohan, Chief Executive Officer and Co-founder, and Desmond Lynch, Chief Financial Officer here with us today. Company, obviously, as most of you know, leader in accelerated compute connectivity, networking, memory and storage controller solutions. Their silicon optical software solutions are integrated into 90% of the world's AI compute servers and clusters. It's been a busy earning season. I've asked the teams to start us off with brief overview of the March quarter results, June quarter outlook, we'll go ahead and kick off the Q&A. Gentlemen, thank you for joining us today.
Jitendra, let me turn it over to you.
Yeah, thanks, Harlan. There was some high-speed communication there. We've been at it now for two years, actually. We went IPO about two years ago, March 2024. At that time, our revenues were $65 million in change, and our EPS was about $0.10. 8 quarters later, we closed our last quarter at $308 million in revenue and what was it, 60?
$0.61.
$0.61.
In EPS. About 5x increase in revenue, and more than that for EPS. Very happy with where we are. Last quarter was $308 million. We grew sequentially about 14% and about 93% year-on-year. The growth there was broad-based. All of our products grew. Certainly the highlight was our Scorpio X-Series family, which started ramping into volume production in Q1 of last year. Overall, you know, very pleased with where our PCIe Gen 6 portfolio is. It is more than 1/3 of our revenue for the quarter, that's a great milestone to have. Just kind of we're getting started on PCI Express Gen 6.
Today, there is really only one Gen 6 GPU out there, which is the Blackwell GPU from NVIDIA.
Yeah.
As more come, we will see, you know, continued strength in this portfolio. For the next quarter, we've guided up 17% at the midpoint of the range or $360 million in revenue. Again, that we anticipate a broad-based growth. All of our products, Aries, Taurus, Scorpio will grow. Certainly the highlight again would be the Scorpio family, where we do expect by the end of the year, it should be our largest product family.
Mm-hmm. Mm-hmm.
Very, very excited about that and continue to work on new product families and announcing new constellations throughout the year.
No, that was a great overview. Thank you for that. Continuing on to that, I mean, it's been a very strong growth profile, as you articulated, right? calendar 2025, you drove your revenues 115%, earnings 140% over the past year. Your calendar 2026 revenue and EPS outlook has been revised higher by 70% and 55% respectively. Total revenue growth of 80% this year, three-year revenue and earnings growth CAGR of 40%-50%, right? This is probably conservative, right? Help us break down the drivers of upside over the past year. More importantly, what are the strong growth drivers for the team looking out over the next sort of two to three years?
Yeah. The last year has been very good. As I mentioned, all of our product families grew very well. Aries continued to grow. There were, you know, a lot of questions asked about, "Hey, you know, what is the longevity of the Aries family?" Happy to report that we grew from 2024 to 2025. We are growing from 2025 to 2026. We are also on track to grow from 2026 into the outer years as well. That product family continues to grow very well as we deploy it, not only for scale out applications.
For scale-up applications and also different form factors, starting with the chip applications where you take the chip and solder it on the board, to cabled applications where we supply Smart Cable Modules that eventually go into active electrical cables. That has gone very well. Just continuing with the theme of PCI Express, we introduced our Scorpio P-Series family last year, that ramped very nicely. Even though it was a partial year for Scorpio P-Series, we ended the year with, you know, 15% of our overall revenues from the Scorpio P-Series family. That traction will continue this year.
On top of that, this year in 2026 will be a lot of growth coming from Scorpio X-Series family, and we anticipate that together combined Scorpio becomes the largest product family exiting the year. Within the Scorpio family, Scorpio X-Series will actually overtake Scorpio P-Series. That just goes to show the amount of TAM that's available for Scorpio X-Series, which is very large, about $10 billion as we anticipate. It's a completely greenfield opportunity. Very excited about the growth potential there. Our Taurus continues to grow. It was the largest growing product family in Q4.
We will see more growth with our lead customer this year. We are anticipating further growth as we diversify from one customer at 400 gig to many customers at 800 gig, which would also happen towards the end of the year. We're finding new opportunities for Leo. Leo is our CXL memory controller device that.
Started out for general purpose compute applications at Microsoft Azure in particular. We are finding new use cases for KV Cache, so that's, you know, progressing as well. We are also very excited about some of the new product families that we are working on, namely our custom line of business.
Yeah
Where we talked about NVLink Fusion as a new opportunity for us to participate in the NVLink ecosystem. That's fantastic. We also talked about a custom version of Leo being designed in at a hyperscaler for KV Cache offload. Last but not the least, optics. Today, most of the deployments are all focused on copper, and we benefit immensely from our content in copper. You know, generation on generation, we've increased our content from starting like sub $100 when we went IPO to crossing $1,000 in some of the content rich applications with our existing product families. You add to that optics with NPO, CPO out in the future, you know, we are set to increase that content significantly beyond $1,000.
All in all, very, very happy with where we are, and we are investing. We are not shy about investing to make sure that we secure our place in the future.
If we step back for a second, and we think about your customer's customers, and we think about the AI workloads, right? A big inflection happened in the second half of last year. Inferencing workloads overtook training workloads, and ever since then, inferencing workloads have continued to grow at an exponential rate as measured by things like token generation, and so on, right? This created new silicon opportunities, right? Inferencing has created new silicon opportunities. It's created new storage and memory tiers. It's created more demand for high-performance CPUs. Obviously, storage and CPUs communicate via PCIe right in the sweet spot, right, of your technology and Astera Labs' product leadership. That's one example, right? Your CXL solutions, as you mentioned, targeted at KV cache applications may be another example.
Help us understand, like, how the transition from more inferencing-based workloads, especially agentic-based workloads, has potentially helped to create new opportunities for the team, maybe even potentially expand your SAM opportunity.
Yeah, certainly. We've always maintained that our solutions are applicable for both training and inference, and that rack level becomes sort of the single unit of compute that is going to be used for both training and inference. That continues. With the advancements in agentic, I would say our strategy doesn't need to change. Our strategy is still the same, which is to address connectivity-
Bottlenecks, for these training and inference-based systems. The applications have become more broad. In particular, maybe I'll point out three things, right? If you look at the inference workloads in particular. For inferencing, especially as we go into these long context windows and, you know, all of us are probably kind of guilty of uploading big documents and saying, "Hey, here is a big document. Tell me, you know, what we should do about it, or a large piece of code and write me some more code." As the context windows go up, the bottleneck shifts towards memory. Utilizing memory, giving more memory bandwidth has become more important, and this is where we see additional, you know, applications for our Leo products.
In particular, this customer opportunity we talked about where the hyperscaler is looking to deploy KV Cache offload application over CXL attached memory. Those type of applications are very good, and they're additive to the applications we already have for general purpose compute. That's one thing I'll mention. The second one is just the different type of inference workloads themselves. For example, if you look at Mixture of Experts workload for inferencing applications, they have a certain traffic pattern which likes all GPUs to talk to all other GPUs, because you don't know where the expert, which GPU the expert resides in. What you need to do is in your switching portfolio, you need to deliver full peer-to-peer bandwidth.
Our Scorpio family was designed bottom up to deliver this exceedingly high bandwidth across all of these GPUs. Not only that, with the addition of features like Hypercast and in-network compute, we can really offload the GPU. We can free up a lot of, you know, IO bandwidth from the GPU by moving that feature into the switch. We can also free up some of the compute bandwidth by moving that feature also into the switch. That becomes, you know, a very, very important part of the overall architecture that our customers are deploying. Last but not the least, with agentic, in general, the sort of this requirement from the frontier labs to lay their hands on compute, whichever way they can find.
You find a lot of diverse configurations coming into the mix and customers trying to deploy different configurations. As you said, Harlan, PCIe is really the nervous system of a server.
Right.
We are, you know, in a leadership position with our PCIe products. We find additional opportunities where people want to connect different combinations of NICs and SSDs and GPUs and whatnot over PCIe. That's, you know, a fantastic opportunity for us.
Yeah. Another dynamic that we've seen changing in the marketplace is, you know, this adoption by custom, your hyperscale and cloud customers and frontier model builders to wanna develop their own custom silicon XPU solutions. We call those ASIC XPU solutions, Google TPU, Amazon Trainium, Meta's MTIA, and so on. Our view is that by calendar 2028, unit-wise, the market will be 50% merchant off-the-shelf GPU, 50% custom ASIC implementation. Our view is that the move to custom ASIC implementations is a net positive for the Astera team. Because these custom guys are gonna have more reliance on standard-based protocols for their scale-up, scale-out connectivity, networking requirements. Wanted to get the team's view on Astera dollar content capture opportunity, sort of merchant versus custom ASIC XPU.
Yeah. We take pride in calling ourselves kind of the Switzerland of connectivity. We definitely address both merchant solutions, GPUs, as well as ASICs.
Having said that, there are two factors that really determine what our content is. One is the number of XPUs, ASICs or GPUs, that you are going to use to deploy for a particular workload. The workload isn't changing. We are still gonna ask the AI to do what we are gonna ask it to do. In the back end, how many of ASICs or GPUs you deploy really impacts our content. In the case of ASICs, typically they are not as performant as the leading NVIDIA GPU.
They end up deploying more ASICs to solve the same problem, and therefore we get more content.
Yeah
With the ASIC platforms. The second thing is what is used for scale up. You know, if somebody uses PCI Express or PCI Express-like protocols, that provides additional opportunity, much more, actually multiplicative opportunity for us to participate. That's the other place where we benefit from in ASIC deployment. Having said that, and our Scorpio X-Series family is, you know, a prime example of that. Having said that, we are participating in GPU ecosystems as well, both when a customer takes a GPU and deploys them in their customized racks. We participate there in scale out. That is actually what led to our growth in 2025 with Scorpio P-Series and Aries Gen 6 deployments that happened.
As well as additional opportunities we are participating in now as part of NVLink Fusion, for example.
Yes.
Over time, what we would like to do is to address all of the AI platforms that our customers are using, both the proprietary ones where we can participate like NVLink and NVLink Fusion, PCI Express, and UALink, you know, as customers evolve from going from one protocol to the other, and over time, Ethernet and other protocols as well.
You know, we spent some time talking about your customers' custom silicon sort of initiatives, but, you know, as you mentioned in your prepared remarks, I mean, Astera themselves has a multi-billion dollar opportunity in front of them doing your own custom chip solutions for some of your customers, right? The team is engaged in a number of custom chip programs. You have your NVLink Fusion partnership with NVIDIA that's gonna enable heterogeneous, you know, GPU, XPU, rack scale solutions. You also mentioned, you know, you've got two custom CXL memory controller solutions for inferencing-based KV Cache offload applications, right? How many custom programs is the team working on? What is the dollar content opportunity per program? When do you see these custom programs, like, ramping into production?
Yeah, this, you know, custom business is very exciting for us. The way this got started is really because of the trust that we developed with our customers. We've been working with our customers now as part of Astera Labs for about eight years. This time we've earned the trust from our customers where they share with us what their, you know, platform is going to look like, what our solutions need to look like to participate in those platforms. Lately, what has happened is, you know, these platforms are becoming so kind of co-designed with software, hardware, you know, the whole rack infrastructure, that our customers want customized solution to get just that much more performance out of their AI infrastructure. They're coming to us to develop these custom silicon.
It's an exciting opportunity, and frankly, we just don't have enough, you know, people or enough resources in the, in the company to address all of these applications. The two that we have talked about so far, one is NVLink Fusion, which now gives us the opportunity to participate for the first time in the NVLink ecosystem.
We are collaborating with NVIDIA and the hyperscaler customer to allow the hyperscaler customer's compute to be deployed over NVLink ecosystem. That's very exciting. The team is fully engaged. We expect the ramp for this product to happen in 2027 alongside the ASIC ramp as has been announced by our customer. That's one. The second example was the custom Leo. This is a little bit different. We did not do a new chip design-
For this application. As I mentioned, our chips are software heavy. We like to architect our chips to use as much software as possible. We were able to use some of those features in the chip to create a custom version that was more.
suitable for this application. That's another example. These are the two that we have talked about publicly. As time goes on, we'll give other indications of, you know, in these custom programs. One thing I do want to mention is that these custom programs are not like ASIC opportunities.
Yeah.
The typical ASIC opportunity where the customer has done the front-end design and the ASIC vendor is responsible for the back end and manufacturing. These are from concept to manufacturing.
Type designs, where a customer comes in and says, "I want this product." We do the full, you know, front-end design, the back-end design, and take it all the way through manufacturing. That allows us to capture more value than you would be able to capture in a, in a just kind of a pure play ASIC model.
Before we go into some of the specific merchant customer families, just wanted to see if there are any questions in the audience. If you do have a question, please wait for one of our microphones to get to you. Any questions from the audience? Okay. Let's continue. Let's move to your Scorpio.
There is one.
Oh, there is one.
We do have one.
Okay, go ahead.
Yeah, just curious about the Cerebras, that's sort of like a bigger, like a huge wafer, like a pizza type of where I mean, where you have like more connectivity directly-
On semi, like a system on chip, offering. If that does kind of ramp up, how does that change the kind of the PCIe memory link that needs to be there, you know?
Yeah, I mean, first of all, you know, what Cerebras has done is fantastic. Congratulations to their team on a very successful IPO. It's a very different model of deploying compute, and as they become more successful, that might evolve to be another platform. At the end of the day, you still need to get data in and out of that system, and that is where we play. The traditional scale-up where you connect many of the GPUs together, may not be applicable because all of these, you know, XPU or ALUs or whatever you call it, are there already on the wafer.
Right.
In order to get data in and out of that, you still need a framework, and that'll be an opportunity for us.
Any other questions? It feels like also, like for some of these more, Cerebras-like opportunities, at some point, they will be needing to leverage, like, big storage repositories, right? That would be a big opportunity for you guys as well.
Yep.
Okay, so on the Scorpio family of scale-up and scale-out fabric switches, two families, as you mentioned, Scorpio P-Series PCIe-based switches, Scorpio X-Series scale-up switching family. You started ramping these products 2025. By the end of the year, with Scorpio driving like 15% of your total sales, right? This year, Scorpio, like you said, is gonna cross over to be one of your largest product segments. Scorpio P-Series, let's start off with that. In addition to NVIDIA, you've been working with two other tier 1 hyperscalers outside of this lead customer. Is the team on track to ramp Scorpio P-Series to these additional customers second half of this year? Has the overall, just overall customer engagement, continued to expand?
Yeah, we've been very pleased with our sort of progress on Scorpio P-Series, which is targeted towards scale-out applications.
We launched the product, say, last year, and it became quickly our fastest growing sort of product line within the company, representing about 15%.
Yeah
Total company revenue last year. Much of this growth was really driven, by our lead customer.
Yes
Across multiple programs from there. What we've continued to do is invest in the product line. Just recently, we announced a 320-lane scale-out, application sort of switch, which we continue to receive sort of strong feedback.
Yeah
From our customers. As we look to the second half of the year, we are on track to be shipping to our two hyperscaler customers, and that will ramp in the second half and continue into 2027 and beyond. We are attracting a sort of high market here, $4 billion sort of market opportunity.
On the scale-out switching applications by 2030. Given our sort of strong product position as well as the customer traction, we expect to continue to grow nicely on the P-Series products.
Perfect. On your X-Series, Scorpio X-Series products? This is your scale-up switching solution. Started pre-production in Q4 of last year. Now ramping initial production shipments with high volume in the second half with your lead customer for their flagship custom XPU solution. In addition to your lead customer engagement, the team has been talking about engagement with 10+ customers. When do these other sort of customer programs start to fire for the Astera team?
Yeah. We're very excited about the Scorpio X-Series product family and the growth opportunities ahead of us. This is certainly a greenfield TAM opportunity for us. We have been shipping our lower radix solutions. As you mentioned, our higher radix solutions will start shipments here in Q2 and continue into the sort of back half of the year. We have talked about having 10 customer engagements, which continues to sort of grow.
Yes.
This is spread across hyperscalers plus AI platform providers. We do expect to turn a couple of these into design wins by the end of the year, which will then lead to sort of revenue opportunities in the 2027 timeframe. Very excited about the sort of growth opportunities that Scorpio X-Series offers us.
No, that's perfect. If we think about the next generation, thinking about the scale-up roadmap architectures, then we think about next generation, then we start to think about standards-based scale-up opportunities, right? UALink standards-based is the next big opportunity to unfold for Astera. It's primarily a 2027 opportunity, but you need to have your solutions to customers kind of second half of this year. Your flagship customers here, we all know, are AWS and AMD Trainium 4, MI500, I believe are the target platforms. Can you guys just give us an update? I mean, are you on track to ship eval products second half of this year? More importantly, outside of your two lead customers, again, has the overall customer engagement pipeline expanded around UALink and UALink adoption?
UALink continues to grow very nicely, actually. There's a lot of traction, you know, for two reasons. One is UALink was designed from the bottom up to do scale-up. It is a really efficient, you know, high throughput, high bandwidth, low latency protocol meant for scale-up. Also because it's an open protocol, so it encourages a vibrant open ecosystem. We have seen continued progress in there. The UALink consortium has put out specifications. You know, several vendors have adopted it, so there is IP available, there is verification IP available. In general, the ecosystem is doing quite well. Harlan, as you correctly pointed out, the two companies that have publicly stated their support for UALink is AWS and AMD.
Both of these, GPU and ASIC are going to ramp in 2027. Our product development is on track to intercept these deployments, you know, as they ramp their ASIC and GPU platforms respectively. Now, on top of that, there's actually a lot of traction from other customers. Some have already decided to use UALink outside of these two. Unfortunately, they have not yet publicly stated their position, so we cannot do that until they do so. There is a lot of traction. The market for UALink and PCI Express combined, we think is about $10 billion in TAM, and we are very well positioned to address that.
Not only because of the, you know, the trust that we have with our customers, but look at what we are doing with the Scorpio X family now.
Yes.
We are, you know, with the 320-lane devices, we will learn a whole lot in the, in the coming months on what it takes to build a good scale-up switch. As you know, it is beyond just the speeds and feeds that you need, right? You know, 200 gig and the number of lanes, et cetera. Features like In-Network Compute and Hypercast are very important to optimize how a scale-up network operates. These are the things that we will learn, and we will transition these into our UALink switch. From a customer standpoint, these feature sets and how they use it, how they deploy it in their software is seamlessly ported from the current generation of Scorpio X devices to what we do for UALink using our COSMOS software.
Yeah. That's a good point. On the new product front, you know, my sense in covering the team before and through the IPO and through now is that it feels like at any given point in time, you're ramping a product and you probably have two or three next generation products in the design win pipeline, right? The introduction of on the last earnings call of your high radix 320-lane Scorpio X-Series, Scorpio P-Series family products is sort of a reflection of that, right? Higher lane count simplifies large scale topologies, improves performance with single hop latency. You mentioned it a couple of times here, right? You've integrated features like In-Network Computing and Hypercast technology. Help us understand, like, how these new features actually boost AI inferencing performance for your customers.
If I just turn and talk about the new introduction of the Scorpio X-Series 320 Lane, there are, you know, a few things that happen with this introduction. One is it's a high radix, high lane count switch. What that means is you can connect multiple GPUs together with a single hop, and you get the advantage of having really, really low latency that natively comes from PCI Express and PCI Express-like protocols, and, you know, the lossless networking. Meaning when a GPU says, "I want to get data from this other GPU," it is guaranteed to arrive in a fixed amount of time.
That's the benefit of, you know, having a large radix and large lane count and, you know, high, high throughput, peer-to-peer bandwidth, et cetera, that is part and parcel of what we have with the Scorpio X family. The second thing is Hypercast. When you go for Mixture of Experts type of inferencing workloads, the so-called experts reside in multiple GPUs. For any given inference query, you may be lighting up different experts in your system.
Mm-hmm. Mm-hmm.
That creates a lot of, you know, 1-to-many and all-to-all type of traffic. What Hypercast allows you to do is instead of 1 GPU having to individually send messages to four other GPUs, it just sends a message to the switch and says, "Take this message and send it to four other GPUs.
The switch now does the work of sending the message to other GPUs. Therefore, offloading the IO bandwidth from the GPU and they can use that for other purposes. Similarly, when you go for a training workload, you have the ability or you have the requirement of multiple GPUs to solve part of the problem and then combine the results together. Instead of each GPU going to the other GPU and getting their piece of the puzzle, they just send one message to the switch. The switch does the combination and sends it right back to all the GPUs. Now you are able to offload communication overhead, you are able to offload compute overhead from these devices. Last but not the least is the COSMOS framework that I talked about.
Yeah.
In the past, we were using COSMOS to customize our solution, optimize it, you know, provide a whole bunch of diagnostics and telemetry type information to make these systems running smoothly. With the introduction of Hypercast and In-Network Computing, both are hardware features, but they get exposed through COSMOS. COSMOS software also becomes much better and much more embedded into our customer software, and that will have a lot of, you know, benefits to us in the future as well.
Perfect. We've got a question here. Can we get the mic? Yeah. Thank you. Just right up here.
Thank you very much for your time. Do you have a major optical design with ramping, I believe, in 2027 with your new fiber coupler?
Thank you very much for your time. Do you have a major optical design with ramping in 2027 with your new fiber coupler?
It seems that the broad volume timeline for CPO optics seems pushed to post 2028. I was wondering, does this mean the upcoming generation of ultra accelerator platforms are still fundamentally designed around the copper first architecture?
Yeah. Worldwide happens on copper.
Yeah.
The reason for that is, you know, copper is reliable. It's just, you know, the lowest cost of ownership, the lowest power when it comes to scale-up applications. We think that the overall copper and optical will continue to coexist for a very long period of time. The connections at the rack level will continue to be copper, and certainly us and others in the industry will continue to support that because that's what our customers are asking us to do. Having said that, as these cluster sizes expand from one rack to multiple racks, you know, three, four, five, six racks, copper will run out of steam. It's just physics. You cannot run from, you know, one rack, four racks over on copper at 200 gig and 400 gig.
That's where optics will come first. Different customers have different approaches on how to deploy it. Some customers want to go straight to CPO, and I think those type of applications, you will see CPO for scale out happen as early as 2027, and then CPO for scale up happen in 2028 and beyond. We will certainly participate in that with AI scale glass coupler technology. It is very good for higher radix implementations of scale up and scale out. Some of our customers are wanting us to just sell them this component. We will have those component level sales. Other customers want to go to NPO first because NPO does not stress the supply chain of the XPU or the switch.
The XPU and the switch remain completely electrical, but you have an NPO component that allows you to go longer distances when such reach is required. We are 100% engaged with our customers on the NPO roadmap, and we should start to see the first revenues for NPO happen in 2027, for again, from rack to rack type of links. For both of these, for both CPO and NPO, you need an optical engine, and I'm happy to say that we started our development of, you know, towards getting to optical engines nearly two years ago. We only put a team together to work on analog mixed signal technology for custom PHYs that are needed to drive silicon photonics. We've been working on that.
We are in a very good place with our electrical IC component of our optical engine. Electrical IC will then drive a photonics IC or a PIC. We have a full team of PIC designers working on a CPO solution as well. However, we are also open to our customers choosing a particular optics or a PIC solution because many customers are pretty religious about what modulation technology, what fab they want to design. We have designed our EIC and our glass coupler to be photonics agnostic to the first level. We can work with, you know, TSMC COUPE, or we can work with GlobalFoundries or TowerJazz or whatever the choice might be. Of course, there is a glass coupler technology that we acquired through AI scale acquisition.
The three of these things put together make up an optical engine that we can deploy with the switch, as well as on the XPU, which is a fantastic multi-billion dollar market opportunity.
Yeah, the Xscale acquisition was actually quite interesting, right? 'Cause when you talk to some of your partners and others out there that are working on these optical-based solutions, fiber coupling is the most challenging and critical barrier, right? EIC, PIC, I think that's a relatively more mature technology, but when you listen to the CPO and NPO guides, it's all about fiber coupling, reliability, scalability, and so on. I think the Astera team was smart in acquiring the fiber coupling technology first. I guess the question there is, what did Xscale bring? What is the differentiation that they bring in terms of this very critical fiber coupling capability?
I think maybe two things I'll mention. First of all, as you said, this fiber coupling is a mechanical aspect, right? The electrical IC, yeah, you can produce them by the bushel. TSMC is very good in their process nodes. Similarly, silicon photonics at the end of the day is a wafer level technology. Whereas the glass coupler is an individual piece that needs to be aligned and glued in place and so on.
AI scale team has developed some very cool technology on how they do this alignment, how they do the mechanical attachment, how they can make this a detachable technology so that if something were to fail, then you don't have to throw away.
Yeah, right.
You know, the kitchen sink. You just replace the part that fails. I think that's one benefit AI scale brought. The second one is, as I mentioned, this high radix capability. When you want to connect, let's say, you know, 72 and in the future 144, 288 GPUs together.
Mm-hmm, mm-hmm.
You want to be able to escape from the switch in a way that you can bifurcate these links into the number of GPUs. That, that's a very different technology, that is needed to build very high density, glass coupler, which is AI scale does.
Perfect. Then, on the financials, long-term gross margin profile, 70% gross margins, 40% operating margins. You'll be driving close to that exiting this year by our estimates. That's almost two years ahead of our views back at the time of IPO. As you layer on new initiatives like your custom silicon programs, optical scale-up initiatives, as you roll out your standards-based PCIe, UALink-based products, puts and takes around the sustainability or even outperformance of your long-term margin profile.
Yeah. as you mentioned, our long-term operating model is 70% gross margin leading to 40% operating income remains sort of unchanged. We expect to continue to grow faster than market given the investments that we've made, we'll continue to see the new products layering into the model as well as the customer diversification as we talked about earlier. In terms of the sort of gross margin, we are targeting 70%. We have been performing slightly above that here-
That's right.
In the past couple of quarters, we do see that the margins will trend down towards our targeted model driven by the warrant impact.
As well as the mix impacts. We'll see the modules versus silicon.
Yes.
We'll also see, you know, the additional impact of the additional switching use cases that will come into the model from there and the diversification of the portfolio. 70% gross margin's very rich for a semiconductor sort of business and really is reflective of the value that our customers place on our solutions. We'll continue to invest at the right level-
In the business to make sure we can continue to grow the top line.
Yeah.
That will lead to the.