Hello, everyone. Thank you, everyone. Welcome to day two of Citi's Tech Conference. My name is Papa Sylla. I am the U.S. optical and internetworking analyst here at Citi, and I am very pleased to welcome Jitendra Mohan.
Hi.
Co-founder and Chief Executive Officer of Astera Labs, Desmond Lynch, Chief Financial Officer of Astera. Welcome, both.
Thank you. Thank you.
We will start with my questions, and toward the end, I will open it up if the audience has some questions as well. I guess to get started, Jitendra, because Astera is so involved in so many of these server designs, I thought it would be helpful to start by giving perhaps the state of AI from your perspective, but also for Astera Labs, it has been now 2.5 years , I believe, since you IPO'd. It has been quite a fast and furious journey. So if you can just talk us through the journey and what is ahead.
Yes, definitely fast and furious. You put it well. Thank you, Papa, first of all, and good to see all of you, some familiar faces, some new ones. Let's start with the AI question first. I believe that we are still early in the multi-year build-out of AI that we are seeing. I think more than my word, we should look at the CapEx expenditure from the hyperscalers, which is now expected to cross $1 trillion next year. That's a very good sign that not only are our hyperscalers investing a lot of dollars, but actually the pace of that investment is even increasing. In many ways, you can see that the tide is rising, and of course, all boats will rise with the tide.
At the same time, I'm also very excited really about the pace of innovation and change that is there in the AI industry. Every three months, every six months, there is something new that comes up. Talking about the IPO, when we went IPO, the unit of compute was basically eight GPUs in a HGX board, the eight Hopper GPUs. That's how ChatGPT moment happened. Now the unit of compute has become a rack. The reason for that is these models just keep getting bigger and bigger. IPO time, we were talking barely just getting into 1 trillion parameter models. Now we have crossed 10 trillion parameter models, and there is no sign of stopping. You look at the latest Astra model that came out was trained on more than 100,000 GPUs. Inference is driving a lot of new workloads, including longer context windows and so on.
Even for inference, now you need racks of compute. The rack has now become a unit of compute. The other trend that has come up more recently is supply. The demand for compute is so much that it is really outstripping the available supply. All of our customers are also looking to figure out how to get more out of the hardware that they have available. Fortunately for us, Astera was really founded back in late 2017. We started operations in 2018 on this premise to solve the challenges with AI connectivity, remove the bottlenecks for data, for networking, and for the memory bottlenecks. This is really what we have been doing.
Our vision is to provide connectivity infrastructure at the rack level, which includes all of the different components, whether that is switching, signal conditioning components, running over copper, running over optical, hardware, as well as software. We are well on our way to do that. That has been our journey since IPO. At the time of IPO, I recall that our revenues were about $60 million something, give or take. Last quarter, we were $392 million. We guided for $550 million at the midpoint for Q3. Similarly, our earnings per share were $0.10, and then last quarter was $0.80. We have come along quite a distance in there. In fact, I brought something to show you guys just to put this in perspective. Apologies for those on the webcast, but this is where we started. This was our Aries Retimers chip.
This is the Scorpio X that we have just recently introduced. This is the 320 lane Scorpio X. That gives you a little bit of perspective about the journey. Going back to the point I made about the tide is rising with AI, with the CapEx spend, we are growing a lot faster simply because we have managed to introduce new product families and capture a lot more content on a per XPU basis. In fact, when we started our content on a per XPU basis was sub $100, largely driven by the Aries retimer products that were attached to the Hopper GPU. Now we are already at close to $1,000 of content. If you look at the Aries family, the Taurus family, the Leo CXL family, and then of course, the most recently introduced Scorpio family.
Every generation of XPUs, we increase our content, especially with the Scorpio X. We are on track to get $1,000 or more content on a per XPU basis from the Scorpio X family alone. So we are very pleased with where we have come over the years. We will continue to work on new product families that gets us into multiple thousands of dollars of content on a per XPU basis. So really proud of the growth that we have seen, really proud of the team to have come up with so many different products and the strong trust that our customers are placing in us and inviting us to the table years in advance of an actual deployment and helping us define these products and bring them to market.
Yeah, no, that is absolutely fascinating. If I knew you would be showing Scorpio, we will do a video webcast instead so that everybody could see. You mentioned kind of Scorpio, and perhaps that is where I want to start. It seems like moving forward, this will be probably your most important kind of business. You mentioned this will be a product that will drive kind of content per accelerator beyond $1,000. I am curious if you can parse through the different types. You have the P version
Yeah
The X on kind of content, but also the growth outlook of both as we go into 2027.
Yeah, certainly. So for those of you not familiar, we have two series of Scorpio devices. Both are purpose-built for AI applications. So before Scorpio, the PCIe switches were used for general purpose compute applications to connect storage and sometimes networking devices, et cetera. So when we decided to enter this space, we wanted to build something that was purposely done for AI, that understands the AI workloads and so on. So we started with the Scorpio P series, which is meant for scale-out connectivity, connecting GPUs to NICs, and to CPUs to connect these different clusters together, typically over Ethernet. That was a product that we introduced in October of 2024 at OCP, and then it very quickly became our fastest-growing product line. Just in three quarters last year, it was 15% of our total revenue. So definitely one of our fastest growing product lines.
That growth has continued. We have added another hyperscaler customer. We are into multiple customers that are now deploying Scorpio P into their AI infrastructure, just because of some of the unique things that we built into Scorpio P. Scorpio X, especially the 320 lane device we introduced just last May, and it has just got a tremendous amount of traction from our customers. So just to explain what Scorpio X is, it sits right in the center of our scale-up network. As I said earlier, it is connecting all the GPUs together so that it looks like one GPU. So from a software perspective, it is very easy to just address, let us say the 72 GPUs as just one big GPU. So a lot of that intelligence goes into this chip.
On top of that, we've added more features like in-network compute and Hypercast that really understand the traffic patterns that go for both training as well as inference workloads and essentially make the GPU work better. So a GPU that's connected to a traditional PCIe switch will produce a certain amount of tokens at the end of the day, whereas the same GPU connected to a Scorpio X will produce more tokens per second. Just think about that. That allows our customers to have less infrastructure for a given workload, or for the same amount of infrastructure, service a lot more workload. They need, at the end of the day, fewer GPUs, fewer racks, fewer nuts and fewer bolts. So it's a phenomenally good device, and we are so thankful to our customers for helping us define and bring this to the market.
We are already shipping this to our lead customer. This device was introduced in Q2 and already entering volume production here in Q3, is also behind the strong guide that we provided for Q3. So we're very thankful for that. It is shipping to additional customers. A lot of customers are actively designing this, and we expect to convert many of these to full design wins to ramp in 2027. Even so, just in Q3, Scorpio will become our largest product family by revenue. That's just incredible. Within the Scorpio family, Scorpio X will overtake Scorpio P. Recall that I just said Scorpio P was our fastest growing product line. So we are very excited by this milestone. There is a lot more to do, a lot many customer applications to solve.
People are just finding different use cases for Scorpio X now that this is available to them as a toolbox.
Yeah. No, absolutely. It seems like, at least from last earnings, that kind of the expectation that Scorpio will become the main product is almost pulled forward by a quarter or so. I guess I'm curious on what's the driver of that. Is it kind of a faster conversion of qualification or design win into revenue? Is it kind of market share gains, or is it just the time overall of the internet working, ramping faster?
No, we are very happy with the progress that we had in the last 18 months or so with the Scorpio family. One of the reasons that this is ramping so fast is actually not something that happened now, but started many months, even years ago with this extreme amount of co-design that we did with the lead customer on both defining the product, simulating it, emulating it, just the team doing intense amount of planning even before the chip taped out and certainly before the chip showed up. After that, the teams on both sides, the customer as well as at Astera Labs, just did an incredible job of bringing up this complex chip and then a complex system in really record time.
I do want to give a shout-out to the team for just the immense lift that they provided in getting this complex system up and running. Now it is getting deployed, and we are starting to see the fruits of their labor. In general, we do try to be conservative when it comes to the amount of time it might take to qualify a system in production. We control certainly our chip. We control a few things, but we do not control everything. There is certainly some amount of, call it just old-fashioned luck, that everything landed in the right way, and we were able to ramp this as quickly as we have been able to.
Right. Quickly, one for you, Jitendra, and I have one for Desmond as well. For both P and X, I am curious if you can frame the competition there. I think in P, you are more of a kind of coming into a market. You found a player there, but it seems like you are doing really well, while X seems to be kind of its own domain, although there are other kind of proprietary scale-up. I am just curious if you can discuss a little bit the competitive dynamics of both Scorpio products.
Certainly. Happy to do that. Before we got on the scene, the dominant switch supplier was actually Broadcom with their PCIe Gen 5 switch products. What we did when we entered this space is we decided to focus on AI. We did not have any baggage. We did not have other customers to support, other applications to support. So we purpose-built the Scorpio family, both P as well as X, for AI. Now, folks have realized since then that this is a very big TAM, both for Scorpio P was already always there, but also for Scorpio X. So it is understandable that the competition will come, and we welcome that. There is a lot of demand. It is not a zero-sum game.
Having said that, the fact that these chips were purpose-built for AI have all these different feature set that is built into the chips, that is really this help us stand apart. If you look at Scorpio P first, we are shipping this in the highest volume for PCI Express Gen 6. We are in a very good position. I feel very confident where we are with our Scorpio P series. The Scorpio X is just getting started. At the root of all of this is the trust that our customers have placed in us, in helping us understand what the applications are, both for Scorpio P and for Scorpio X, and therefore giving us the direction on what features to build, how to build it, and so on. First of all, very thankful to our customers for that.
Secondly, as I've mentioned now three times, these are just custom built for AI. Just think about it one more time that a GPU attached to a Scorpio X just does better tokens per second than a GPU attached to a non-Scorpio X device. That's just phenomenal. In today's supply-constrained world, that is gold. If you can get more tokens per second from your infrastructure. I think that's fantastic. We are already shipping this to multiple customers, just a strong amount of interest that we have from our customers to design this in. Really looking forward to deploying this at scale, not only this year, but then ramping into next year. We are just getting started with a multi-quarter deployment for Scorpio X. Multiple applications within the lead customer, then many more customers.
Right. No, that makes sense. I'm curious on when we think about gross margin, and you gave out the longer-term kind of guide of eventually around 70%, but the company is hovering around low 70s, low to mid-70s. I guess when we think of Scorpio X coming in, it seems to be a very kind of unique product, but also I assume as part of it will probably early on as the ramp starts, the gross margin for the product will continue to ramp as well. Just what would be the impact of kind of these products coming in in late 2026, 2027. Is the risk to the 70% gross margin on the upside or perhaps on the downside?
Yeah. We've been very pleased with our performance on the gross margin side, where we've been operating around the mid-70% sort of range. What we have indicated to investors is that over time, we do see the margins trending down towards 70%, which is in line with our long-term target. I would say the couple of factors here would be on the product mix side. Firstly, would be the mix of silicon versus module content, and secondly, it relates to the Scorpio X sort of series contributions there. If we drill into the sort of Scorpio X, what we're really seeing is a wide range of sort of margins. We have a broad sort of portfolio today. The margin profile will really be dictated by a use case as well as lane sort of configuration.
But overall, I would say that our Scorpio margins are in line with our corporate averages. In any given quarter, depending upon what shipping, the margins can sort of move around from there. But it is important to remember that we are going after large opportunities, and what you do see is a significant fall-through to gross margin dollars, as well as operating sort of leverage from there. But I would say overall, I would say that our margins will trend down towards the 70% range, which is very healthy for a product semiconductor business, and it is in line with our overall expectations.
Got it. No, that is very helpful. Coming back to you, maybe Jitendra, on slash maybe moving to a different topic, kind of UALink slash perhaps one of the key kind of server that you would be potentially kind of getting more content, the Helios server as well. To my understanding, at least the first Helios server would be kind of UALink on top of Ethernet. So the fabric is really Ethernet. But is your understanding as we move kind of generation to generation, it will end up being fully UALink, or do you see kind of two paths, kind of UALink on one side and some type of a kind of Ethernet version of it?
Yeah. First of all, I think we should not talk on behalf of AMD. I will kind of reiterate what they have stated publicly. First and foremost, you are correct that for the MI450 series, they are going to go with UALink running over Ethernet. That has really got everything to do with the fact that UALink switches are not yet available. So in order to deploy MI450 series, they have to use UALink that is tunneled over Ethernet. As we go into future generations, what AMD has stated publicly is they believe that UALink is the highest performance protocol on the planet. The reason UALink is the highest performance is because it is a memory semantics-based protocol, which gives you the low latency and the ease of access of multiple GPUs, but also gives you the fast speeds of Ethernet.
Now they've also stated that they will continue to offer both options, running UALink over Ethernet as well as running UALink natively, depending upon where the customers are. I think that's probably a fair way to say where AMD is. We do think that both of these will continue to coexist.
Right. No, that makes sense. In terms of the content you have with that particular customer, I'm sure it doesn't go from zero to a lot of content when we get to UALink. You already have some content prior to Helios at this early version of Helios and eventually with UALink. Can you perhaps just kind of contrast for us how content can potentially grow from prior to Helios to ultimately a UALink version of it?
Yeah. Again, I can't comment specifically on the Helios platform, that's AMD specific. At the high level, what you can look at is, what is our attach rate when the fabric is something that we support, whether it's PCI Express or UALink, versus its fabric where we do not support switching content. In the case where we don't support a switching content, our content is typically focused on retimers, whether they are PCI Express retimers on the scale outside, Scorpio P series, again, on the scale outside, as well as Ethernet Taurus retimers or active electrical cables on the scale outside. We are more limited to scale out, and that's the content or the TAM that is available to us.
Now, when you have a scale-up protocol that we do support, whether it's PCI Express or UALink or something in between, then our content really does explode a lot, because scale up is a very rich area for connectivity solutions. As and when we have versions of Helios or other systems that deploy UALink natively or, let's say, a PCI Express-based protocol natively where we can have a Scorpio X solution, then our content grows very meaningfully.
Wow. On UALink, I think you were one of the companies or the company that led the discussion there. I am just curious on engagement there as well. Recently, we heard many companies also aiming to do some type of UALink switch or jumping into the UALink bandwagon. So I am just curious on, for investors and sell side analysts, what should we look out for to try to determine who will lead this race moving forward?
Yeah, I think maybe I will answer two ways. The first is to look at UALink itself. UALink, we believe, is going to be an evolution of other memory semantic-based protocols like PCI Express. While there is a lot of focus on UALink, as there should be, there is a lot more to it than the protocol itself. The things we talked about earlier, in-network compute, Hypercast, these are all of the feature sets that are really important for GPUs to deliver high utilization. We will carry forward all of this knowledge from our current generation of Scorpio X to the one that we do for UALink. Diagnostics, telemetry, all of these things are extremely important, and we offer all of that through COSMOS.
To somebody who is already using a Scorpio X and therefore using our COSMOS software, it just becomes very easy for them to move from Scorpio X to a UALink-based Scorpio X. However, if I level this conversation up a little bit, if you look at the overall scale-up TAM, we have proprietary protocols like NVIDIA's NVLink, we have memory semantic-based protocols like PCI Express and UALink, and then we have Ethernet, which is a messaging protocol. Our vision, our goal over time is to address a full scale-up TAM. We are already now active with the NVLink Fusion, so I think that is going very well. We are active with PCI Express. You are seeing the results of that now. We are working on UALink as well, and over time, as our customers demand it, we will address the rest of the TAM as well.
Because again, protocol is one thing, but understanding how a scale-up network works, deploying this at scale, taking those learnings and applying them into new generations, will really determine who gets what share of this overall big market.
Got it. And maybe switching to a different aspect of your business that I think a lot of investors are very excited about is the optical side of the aiXscale acquisition. To my understanding, the early sales or the early products are more the fiber type of products, but you will also have, or you have the optical engine as well. I am curious on how you see that optical business in terms of both growth and how large you see it being. What is your strategic approach to that business? Is it mostly to support, let us say, your Scorpio portfolio, or do you aim for that business to be a standalone business on its own?
Yeah, great question, Papa. The optical TAM is actually very large, tens of billions of dollars in TAM that we are able to go after. The strategy that we have laid out for our optical is really straightforward. We will start our optical journey, we have actually already started the optical journey two years ago when we put a team together to address this. But from a revenue perspective, that will happen in 2027, starting with component level sales. These component level sales will come in the form of the connector, which we get from the acquisition of a company called aiXscale that we acquired in December of last year. So they build this connector that is really very good for CPU applications in particular, and some of our customers are trying to qualify that component for use in their own CPU applications for scale-out applications.
That might very well end up becoming the first revenue that we see in 2027. That will then lead into NPO. At the 200G data rate, we believe NPO is a very good solution that does not stress the supply chain, but also gives you the distance and the density that you need for scale-up networks. So we are fully ready to participate in the NPO side of the house with optical for the 200 gig per lane data rates. As we go to 400 gig per lane data rates and even higher densities, that is where CPO comes in.
This is where we will have an optical engine that uses an electrical IC that is done by Astera Labs, a photonics IC or PIC that is also done by Astera Labs, as well as the packaging technology to attach the fiber to the silicon photonics also offered by Astera Labs. So we will have a full optical engine which will reside not only on the Scorpio, so it becomes an optically enabled Scorpio SKU, if you will, but also on the XPU side. So on a link where previously there was no such component, you get two components. So that is a very big opportunity and will drive our content on a per XPU basis into multiple thousands of dollars.
The good news is we are already well on our way towards delivering this full solution, starting with component level sales, then going into NPO, and then eventually to CPO. The last thing that I will mention here is that while we talk about optics usually in isolation as a media choice for carrying this data, having a switching platform is really key for customers to decide which optical solution, or in general, which signal conditioning solution they will use. So for us, having this switching platform with Scorpio X and then expanding into multiple different protocols is going to be really key to also influence a full end-to-end link, both for copper as well as for optical.
That is very helpful. I have a couple more questions, but I want to pause for a second and maybe open it up if the public has some questions. Sorry, let me just give the mic.
Thank you so much for sharing. Can I quickly ask about customer concentration? As you grow your Scorpio line very quickly, and obviously some customers will ramp up faster than others. So what is the medium-term plan in terms of your breakout into different customers, or is your goal more concentrated than others? Thank you.
We continue to be pleased with our customer engagements. What I would say is that across all of our portfolio today, we are shipping to all of the U.S.-based hyperscalers. But given the end markets that we are in of data center and AI, customer concentration is to be expected from here. But what we see is a continued diversification efforts across the portfolio. Jitendra touched upon Scorpio P engagements. Again, that is led by our lead customer. But what we are seeing is hyperscaler customers coming into the mix, as well as AI infrastructure providers there. So we are seeing the diversification efforts take place there, and that will continue into 2027 and beyond. As we look at Scorpio X series as well, what we have talked about initially was having 10 customer engagements.
That number now today is up into the high teens, and we expect to convert several of these opportunities into design wins by the end of the year. Going into 2027, we will see that continued revenue from there. Also on Taurus and our Leo product families, we see diversification as well. We are really pleased with how the customer diversification is taking place and the engagements that we have with customers. We will see that contributing into the model going into 2027. It is important to note that our lead customer will also continue to grow as well. Overall, we are very pleased across the entire portfolio of the diversification efforts that we continue to see. Thanks.
Hey, quick question. In terms of scale-up domain going past one rack to multiple, how do you view that in terms of growing opportunity on a per XPU basis or just as a total?
Yeah, that is a great question. It is just exactly what I was talking about the optical side earlier. We do expect that optical and copper will coexist, and in particular, what that means is at the rack level, within the rack, the connections will predominantly remain copper. They are copper already today at 200 gigabit per second, but even at 400, we believe that there is a way to keep those connections copper. When you go from one rack to another rack, as cluster sizes increase and your people are no longer satisfied with 72 GPUs in a rack and you want to have 144, 288, even 576 GPUs in a cluster, you have to go to multiple racks. When you go to multiple racks, as physics gets in the way and copper will not cut it at these increased data rates.
Then we go into optical solutions, starting with NPO, and our content goes up very significantly for those links because you still have the switching content that we had before. That does not change. Now we introduce two NPO devices, one on either end of the link. This is actually a very exciting development and we will start to see initial revenues in the back half of 2027 from our NPO efforts. Then full year ahead in 2028.
Switching goes up.
Yeah, of course. Switching goes up as well. Switching goes up as well because now the cluster size has gone up, and so the radix of the switch goes up. There is more data rates are going up. The number of GPUs connected to each switch is going up. These advanced features that we talked about, they are there. Roughly speaking, that translates into higher ASP on a per lane basis as the radix of the switch increases. Contrary to popular belief that price per lane goes down, it actually goes up with radix.
Maybe I will kind of pursue with some of additional question. I think you mentioned also the custom solutions part of your business, let's say like NVLink Fusion. I am curious on that part, how large do you see that opportunity? Also, I guess as you get perhaps more NVLink Fusion coming in, does that take away, I guess, from other parts of your business for, let's say, the Scorpio X or UALink?
Yeah, no, not at all. First of all, custom is a very big market opportunity. It is tremendous. Just look at the amount of revenues that Broadcom has and Marvell has and others. We are not going after all of it. We are very selective in the parts of custom that we are going after, but we are very excited with the opportunities that we have. We have talked about two of those. One is a custom, the most recent one is actually a custom opportunity for Leo CXL to get deployed in a KV cache offload application in AI. So very excited about that. That is in qualification now and will start to ramp in 2027. That hopefully will open many more doors for similar applications at other customers as well. Then the second one is NVLink Fusion.
With NVLink Fusion, we are able to help our customers take their XPUs and attach them to NVIDIA's NVL72 or MGX platform. So that is incremental TAM for us. There are some cases where the alternate, for example, in case of, let us say, AWS, as they have said, they will have a native solution with a UALink switch as well as an NVLink Fusion solution. With our NVLink Fusion solution, the total revenue opportunity on a per rack basis is about the same given the differences in attach rate and ASP.
Maybe it is a wash. But if you look at other opportunities that we are finding customers wanting to use the NVIDIA infrastructure to deploy their XPU, that becomes net positive incremental TAM that is available to us. More importantly, it is a great sign of NVIDIA trusting us to deploy the NVLink infrastructure with other customers. I think when you have that level of trust with the biggest customers, certainly good things will happen. Like I said earlier, our goal is to play in all of these different ecosystems, NVIDIA ecosystem, PCIe UALink ecosystem, and others over time.
Got it. No, that is pretty very clear. I am curious, maybe this is a high-level question. In terms of agentic, when we think of agentic AI, and when we think of CPU racks and combining CPU and GPU racks, I guess earlier on there was questions around the retimer content ratio. But now with this agentic AI, it seems like that ratio could even go up in those type of development. I am just curious on agentic AI, can you tie agentic AI, I guess, to your opportunities in the retimer and other products?
Yeah, absolutely. Not just the retimer, by the way. Of course, it is an opportunity for retimer, but actually it is bigger than that. If you look at an agentic system where people are trying to have, or customers are trying to have different ratios of CPUs to GPUs and disaggregate GPU and CPU or create a composable system, PCI Express just naturally emerges as a protocol of choice because all of these components, CPUs, GPUs, storage, NICs, they all support PCIe natively, and it is designed for exactly this purpose. We have the largest portfolio of PCI Express devices, starting with switches and signal conditioning like retimers, active electrical cables, active optical cables. We can run PCI Express over optics, gearboxes. We are a one-stop shop. Oh, Leo. I forgot about the CXL. When you want to attach more memory, you attach it over Leo.
We have a full portfolio for our customers to optimize for agentic, and it creates a large opportunity for all of our devices, which we are happy about. In particular, of course, retimers because as these systems expand multiple racks, the reach requirement goes up, data rates are going up, and that gives us additional opportunity for retimers as well. It is an overall broad portfolio story.
Got it. I know we are getting close to time, but just two more questions, one for Des and one to finish up with you, Jitendra. I guess with the balance sheet, it seems like now the cash lockup expiration, we are getting past that period. Just curious on how you think about cash allocations moving forward. Is it mostly reinvesting in the business, R&D, or perhaps could we think of investor kind of returning cash to investor there?
Yeah. We are very fortunate to have a robust balance sheet with continued strong cash generation, and our approach to capital allocation really centers around organic investments as well as inorganic investments. Organically, we have done a really nice job of expanding the portfolio, as Jitendra has outlined today, growing from retimer sort of business all the way through to the switching sort of portfolio, and that has increased our overall market opportunity there. We will continue to invest at the right rate in the business to go after these long-term opportunities, and we are making investments in optical and custom solutions, which will continue to grow from there. Inorganically, we have made some smaller acquisitions, and that is something we will continue to look at going forward.
aiXscale Photonics was a great example of that, of where we were able to bring in some of the glass coupler technology in-house, and that strategically is going to be important for our optical strategy from there. We have also done some smaller acquisitions on acquihires, which has brought skilled design teams in-house to the company, and that is something we will continue looking at going forward. I would say our same playbook would play out. We do expect to continue to invest both organically and inorganically in the business, and that will be our focus from a capital allocation perspective.
Got it. Very helpful. Jitendra, maybe to finish it off, any part of this Astera story that you believe is still underappreciated you wanted to address, or any question I didn't ask that you want to touch on?
Yeah, just briefly. I understand that we are over time, but just briefly. I think the part that's probably not very well understood is this is a multi-generational opportunity that we have to not only benefit from the rising tide of AI but to just gather more and more content on a per XPU basis as we come up with these different product lines. Already our content is now up $1,000 per XPU. With the new Scorpio X devices, we get to $1,000 per XPU on Scorpio X alone. Then to add to that, the optical multiple tens of billions of dollars of TAM with optical. So we are looking at a significant growth opportunity that we have with our products. That's one. The second part is just to reiterate what I said earlier.
Once you have a switching platform, that gives you significant influence over what happens in the rest of the system. So once you have a switching platform, really it's yours to lose for when it comes to signal conditioning, whether it is running over copper or over optical. Over time, we are very pleased to be providing solutions that scale from chips to hardware to software to allow our customers to really get the best out of their AI infrastructure.
Right. Thank you so much, Jitendra.
Thank you.
Thank you, Des. Thank you for joining us.
Thank you, Papa.
Very good.