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

All right. Good afternoon, and welcome to JPMorgan's 53rd Annual Technology, Media and Communications Conference. My name is Harlan Sur. I'm the semiconductor and semiconductor capital equipment analyst for the firm. Very pleased to have the team from Astera Labs here with us today. Jitendra Mohan, Chief Executive Officer and Co-Founder; Mike Tate, Chief Financial Officer, here with us today as well. The company, leader in accelerated compute connectivity, networking, memory and storage controller solutions, their silicon and software are integrated into 90% of the world's AI compute servers and clusters. It's been a busy earnings season.

So I've asked the team to maybe kick us off with a brief overview of the March quarter results, June quarter outlook, and then we'll go ahead and kick off the Q&A. So gentlemen, thank you for joining us this afternoon.

Thanks, Harlan. Good afternoon, everyone. Last week, we reported our Q1 earnings. We reported $159 million of revenue, which was up 13% sequentially and 144% from the prior year. And we did 74.9% gross margins and we beat our earnings by $0.05 with $0.33 of EPS on a non-GAAP basis. We continue to see very strong growth from our Aries and Taurus product lines. In particular, right now, we're seeing those product lines do very well with internally developed AI accelerator platforms for scale-up and scale-out connectivity. We guided Q2 to be up $170 million to $175 million of revenue, which is 7% to 10% sequentially. We continue to expect good growth from Aries and Taurus, but we're also excited to be shipping volume for the first time for Scorpio, which a lot of the incremental growth in the Q2 was from our Scorpio product line.

Great. Perfect. Let's start off at a high level sort of big picture. It's great to look back over the past year post the team coming to the public markets in March of 2024. So if we use this year as a proxy for the team's performance, right, you're currently targeted to drive about $700 million in revenues this year. A year ago, calendar '25 revenue estimates were $400 million, right? So you're tracking 70% better relative to 12 months ago. Earnings power this year, $1.35 per share versus a year ago where we were estimating calendar '25 of $0.50 per share, tracking 2.5x better relative to what we thought a year ago. So it looks like the team will be driving at or close to $1 billion a year before our expectations relative to what we had a year ago. So significant upside, significant outperformance, both revenues, earnings powers over this period of time. The team has brought a plethora of products to the market over the past year. And as we look forward, right, from calendar '24, we still anticipate the team driving a 35%, 40% revenue CAGR, 45%, 50% EPS CAGR. So Jitendra, what are the strong drivers you think for the team over the next kind of, call it, 2 to 3 years?

Yes. Thanks, Harlan. First of all, as Mike would say, we only guide 1 quarter out. So please keep your excitement in check. But we do have very exciting things going on. If you think about what led to majority of our revenues in 2024, it was really the Aries Retimer family as it was being rolled out with the Hopper platform from NVIDIA. So that enjoyed pretty significant growth in 2024 and even before that. But towards the second half of '24, we started ramping our Aries products and Taurus products for scale-up and scale-out opportunity and also in the form factor of an active electrical -- actually smart cable module, which then gets assembled into an active electrical cable. So that was our first foray, if you will, into scale-up networking and it's going to be a significant growth factor even in the outer years. Looking at 2025, the first half was driven by ASIC platforms where we participate in scale-up as well as some amount of scale-out as well and the growth being driven by Aries and Taurus smart cable modules. As we look into the second half of 2025, we see actually strong growth from our Scorpio family for scale-out and GPU connectivity, specifically with some of the Blackwell-based platforms as they get deployed in a customized form factor in our lead hyperscaler customer. So that should drive good growth for the rest of 2025. And then as we look at 2026, our Scorpio X family becomes even more important. So as a reminder, Scorpio X family is what used to have scale up to connect GPUs to each other in a scale-up network. And over time, we expect that to become even a larger part of our overall portfolio because that is the largest TAM that we are targeting. And currently, the amount of revenues there are just really starting from scratch. So Scorpio P for the rest of this year and then Scorpio X layering in, in 2026. Outside of the AI, we also have general purpose compute, where our Leo platform will start to show some meaningful revenue in 2026. We are in advanced stages of qualification with our lead hyperscaler to implement memory expansion for a large database application. So that should also sets us up nicely for 2026.

Perfect. Let's talk about the overall environment, right? So on the overall AI and data center spending environment, there have been some concerns on the CapEx spending momentum potentially peaking this year, maybe some AI compute digestion, right? You've had some -- a lot of noise around efficiencies from model innovations like DeepSeek. We also saw more AI bands to China recently and then obviously, tariff and trade concerns, right? But on the flip side, strong ramp of your merchant GPU customers on the next-gen AI platform, strong new AI XPU ramps, new entrants into the ASIC XPU market, right, since -- and so since the beginning of this year, rewind to the beginning of this year, has anything changed meaningfully positive or negative on customer programs or the demand outlook for this year? And more importantly, your confidence level on continued strong growth into next year?

So from a customer ramp standpoint, nothing fundamentally has changed from our expectations that we had laid out. If anything, we are kind of pleasantly surprised at the amount of traction that we are getting for some of our new products in the scale-up side. You will hear the word scale up many times today, somebody keep count. So definitely, that has changed. I think what has also meaningfully changed is our own confidence in execution. So we just started the qualification process with our Aries Gen 6 Retimer as well as the Scorpio P-Series for the Blackwell platform. That is going very well. We are a large way through that qualification process, and we expect to see some meaningful revenues even earlier than what we had anticipated. So just this quarter, in the second quarter, we'll see some revenues from Scorpio P that continues on for the later part of the year as we discussed. Your other question upon CapEx spending this year versus next, I mean, we don't have the crystal ball. But we still believe that we are in early phases of this AI rollout. There is still -- I mean, scaling laws are pretty much intact, especially with the new innovations like DeepSeek and so on. If anything, there will be more performance gains to be had by adding more hardware. And so we expect that our hyperscaler customers will continue to invest. Our job really is to grow faster than how the CapEx is growing. So that's what we are focused on. Regardless of which way the market is going, we grow faster than the market based on the increased content we have with our new products.

A dynamic that has continued to unfold in AI compute infrastructure is the adoption of custom ASIC AI XPU accelerators, right? Google TPU, Amazon Trainium, Meta's MTIA, Microsoft's Maya program. Additionally, we've seen new ASIC programs even just over the past kind of 6 to 9 months. OpenAI, SoftBank, Arm are good examples of that. We estimate that from a unit XPU perspective mix this year, ASIC XPUs to be 40%, 45% of the overall unit mix with 55%, 60% merchant GPU moving to 50% by calendar '27. Our view is that the move to custom ASIC implementations is actually a net positive for the Astera team, more reliance on standard-based protocols for their scale-up and scale-out connectivity and networking requirements, right? But I wanted to get the team's view on Astera's dollar content capture opportunity, merchant GPU versus sort of ASIC.

Yes. So I would say, first of all, we like to play nice with both GPUs as well as ASICs. So we like to be the kind of the Switzerland of connectivity, so to speak. And we are supporting the Blackwell rollout very well. And as the ASICs rollout, we will continue to support that. The meaningful difference in content comes based on where we are playing in these complex AI systems. If we play on scale outside so if you play only on scale out, then that's sort of the baseline of content that we get, which could be a Aries Retimer or Taurus Ethernet smart cable modules or even a Scorpio P-Series. But then we also play in the scale upside, that's where the content increases pretty significantly because in addition to having, let's say, Aries smart cable modules, we now also have Scorpio X devices. And because of the nature of scale-up connectivity, there are more links that are running faster. It's really a much more important part of the overall AI infrastructure or AI rack. And so that allows our content to go up meaningfully into hundreds of dollars on a per GPU basis, which is significantly higher than where we used to be just a year ago with effectively 1 or 2 retimers per GPU.

And would you agree -- you guys, like I said, ship into 90% of all the big AI clusters out there. Are we kind of in the ballpark? This year, kind of 40%, 45% sort of AI ASIC XPU mix, kind of 50%, 55% share merchant GPU. Are we kind of in that ballpark?

It sounds about right. You probably have a much better idea of what that mix is than us. We -- like I said, we support both of them. Over time, we do see that the hyperscalers will value their own internal developments perhaps more than depending upon external platforms. There are internal use cases for which the internally developed ASICs are plenty good. And then there are use cases where they need to offer these GPUs to outside customers where the platforms, the GPU platforms are better suited today, but that might also change over time.

Before I start talking about some of the product dynamics, does anybody have any questions? If you do have a question, feel free to raise your hand, and we'll get a mic over to you. Any questions?

Could you help us understand as inference becomes bigger and bigger portion of this industry, how does that play into your business mix? And any elaboration on your last point? You mentioned that hyperscaler value more internal development. That would suggest that they're leaning more towards pushing ASICs higher and higher. How do you sort of view that world?

So maybe answer the second question first. It's just hyperscalers have a good view of what their customers want, what workloads they want to run, whether they're internal or external. And just overall, we'll have better control of the supply chain with their internal solutions. So there's no surprise that they will value those products more. The commercially available GPUs, the third-party GPUs are better today. The products that are coming out of NVIDIA, Hopper and now Blackwell are just state-of-the-art. So it will be a while before these guys catch up. But I think at some point in time, they will catch up. Now to your first question on inference, inference is, of course, becoming a lot more important today and requiring actually much more compute than previously thought. As we have these chain of thought models or reasoning models, the amount of compute required for inference is now 10x of what it used to be just a few months ago. And with that, what we are seeing is the basic unit of compute is becoming a rack level AI infrastructure. So whether you're doing training or you're doing inference, you need a whole rack. And given that our opportunity is at the rack level and more opportunity at the rack level because of just the complexity of the system and how far these signals need to run, it is overall a good thing for us. There are some customized inference solutions that can be deployed. In fact, we released our MGX reference board, which you can use for small-scale inference. Having said that, though, I would -- I think most people will try to use the same hardware for both inference and training. And we have content in both of those. So to the first order, it doesn't matter for us.

Can you elaborate if the inferencing is going to be mainly done on GPUs? Or is there a different version, maybe CPUs can also do it? Any color there?

Yes, different opinions. Mostly people who have CPUs seem to think that they can do inference and CPU and certainly, they can. But the reality is people are already deploying GPUs. And if you already have a GPU cluster deployed, maybe you will deploy a newer cluster for doing training and use the older one for doing inference. So we believe that inference will also continue to be on GPUs, especially as you have more GPUs available as well as the inference requirements go up.

Let's turn to your product category. So on your flagship ARE smart retimer product, still the largest part of your revenue mix. We estimate about 65%, 70% of your overall revenues. All of it up until recently has been Gen5 PCIe Retimers, 90% plus market share. Strong growth here in the first half, primarily driven by continued adoption of merchant GPU as well as ASICs, right, first half of the year. You've qualified your next-generation PCIe Gen6 Retimers starting to ramp this quarter to support the ramp up, like you said, some of the customer Rack-Scale architectures of the #1 GPU supplier in the market. Additionally, you have continued step-up of new ASIC programs like Google's next generation TPU ramping in the second half, still using Gen5. How do you see the Gen5 retimer mix versus Gen6 exiting this year and in and through calendar '26. And what are the custom ASIC XPU platforms, when are the custom ASIC platform is going to start to migrate to Gen6?

Yes. So the Gen5 still has a lot of growth ahead of it. What we saw last year was a lot of the adoption for AI servers on the scale-out networking, especially third-party GPUs platforms. In Q3, we started to do the scale-up and the scale-up is a much bigger unit opportunity because of all the connections that need to be made. And also, we sell service that one as a module, so it has a higher ASP. They use active electric cables for the scale-up connectivity. And that's still building momentum in 2025. So it still has a lot of legs to it. And then within that, you still have general purpose servers adopting Gen5 retimers as well. That being said, we're the first with Gen6, and we're getting good traction out there. But the ecosystem is still evolving. So you have the Blackwell platform as the first Gen6 capable GPU, and we have designs with that. But that will build a lot of momentum in 2026. So it will layer growth on top of a growing Gen5 mix?

Perfect. And I said 90% share of the retirement market fast-growing markets will always attract competitors. And we've seen competitors introducing Gen5. We've seen competitors introducing Gen6 retimer solutions, right? Investors are concerned that customers do want diversification of suppliers. But I'm wondering, especially in AI compute where the cadence of new product introductions is so aggressive, right? And performance reliability is such a top priority, right? And I'm wondering if supplier diversification is actually not that high of a priority, right? But as long as the current supplier has a strong track record of execution in delivering the right products at the right time. From my perspective, there's just too much risk to bring on new suppliers with -- especially with unproven software. You've got, like I said, 85%, 90% share Gen5, given your design win visibility, does the team anticipate maintaining continued strong share into the Gen6 adoption curve?

Yes. So Harlan, a good point, I will agree with you on sort of how quickly this market moves and what are the main careabouts that our customers have, which are first and foremost, can you make my system work? So you need to have the right level of performance. And then can you make the system work robustly, meaning days in and days out, you are able to maintain that performance, have the right diagnostics capability and so on to understand if things go wrong, why are they going wrong, where they're going wrong so they can take corrective measures. And we make all of this possible through our COSMOS software, which is already deployed at our customers. So that does make our solutions very sticky. And that's what we need to continue to do. We have to always stay paranoid in terms of competition. Competitors will come. It's a large market. Typically, when we start out, we make the assumption that we will get our kind of equitable share of the market. But over time, as time evolves and our solution gets mature, while the competitors are still not there, we are in a very good position to kind of repeat the story that we had with PCI Express Gen5 also with PCI Express Gen6. I do want to mention 1 other dynamic that is at play, which is it's no longer about kind of piece components, my retimer versus your retimer. Not only do you need to have the software, but you need to have that portfolio. So now if you look at PCI Express Gen6, we have PCI Express Aries Retimers. We will have the active electrical cable. So just like we have for Gen5 we have the switches for Gen6. And now we also introduced the gearbox so we have a full portfolio of product. Anybody who's designing a PCI Express Gen6 based system is likely to come to us because we offer the full solution to the different chips as well as the software that goes with it.

Yes. Speaking about the portfolio being a significant differentiator, the team launched its Scorpio family of Fabric switches last year. It's a $5 billion market opportunity over the next few years, taking your total addressable market opportunity, I think, to about like something over $12 billion over the next few years. You've already had achieved design win success for both the P-Series and X-Series of Scorpio, which is expected to drive about 10% of revenues for 2025. Assuming the average ASP for these products is several hundred dollars, I mean this would imply that the team is only going to ship like 250,000, 300,000 units in the second half of this year. Assuming a one-to-one attach switch to GPU, this is only 2% to 3% of the market, right? So huge market opportunity still in front of the team, right? Could you maybe comment on the number of customer engagements and maybe the breadth of deployment of your solutions?

Yes. Happy to do that. And first of all I would say that even though we are shipping that number of units or close to that, we see that in and of itself is a huge achievement because the Scorpio family will constitute 10% of our revenue, and we just started shipping this year. So I think kudos to the team, some of you might be listening to this webcast. The other point is about the market opportunity, and it is indeed huge. One of the things that I've been very pleasantly surprised over the last 3 months is just the number of engagements that we have with our Scorpio family, both on the P but even more impressively on the X family, where Scorpio has really become this anchor socket where customers start designing their AI racks around the switch as well as the connectivity components that we have for PCI Express. And so not only do we get these conversations going for the switch and what the switch brings, we can also pull in other components to deliver an overall connectivity solution at the rack level. And while we are doing this at PCI Express Gen6 today, this dovetails nicely to the future where we might do the same thing for example, UAL.

Yes, we'll get to UAL in a second, but great to see the team getting qualification, or ramping your PCIe Gen6 portfolio, which includes Scorpio P. That being said, there are still a lot more custom AI ASIC programs coming to the market that are still going to be Gen5, right? They want to get to market quickly. They're familiar with your Retimer. They're familiar with their COSMOS software solution. Switching performance from our products is very strong. So has the team also been able to get design win traction on your PCIe Gen5 Scorpio P-Series based switching solutions as well for these new ASIC platforms? Or are they just future-proofing their platforms by adopting your Gen6 platform.

So anybody who's doing a new design is going to look at Gen6 as sort of the way to go because as Mike mentioned, only Blackwell is Gen6 capable GP out there, but it's fair to assume that all new GPUs that are going to come out will support Gen6. So definitely, that's a very strong driver to consider PCI Express Gen6 based solution from the get-go. Any solutions that are already using existing PC Express switches are unlikely to transition because just like you said earlier, these things move too fast. It's not worth it to enable a second source there. So that's not of super importance to us, and we never counted on that. But there are some opportunities kind of in between where there's a new design being done at Gen5 using Gen5 accelerators, which requires a switch. And we are very actively engaged in a few of those a -- few important ones for us. I wouldn't quite call it a design win yet because design win is a very strict definition for us, but very strong engagements, board schematics, layouts being done and so on. So stay tuned.

Let's move on to the Scorpio X. As you mentioned, that's potentially the bigger opportunity here, right? XPU to XPU, scale-up, switching and fabric solution. We've articulated all of the XPU ASIC programs coming to the market either in the market or coming to the market, right? All of these programs are Rack-Scale. Customers really need help with the Rack-Scale connectivity architecture, perfect entry point for Astera Scorpio X to enable this XPU to XPU connectivity. Given the urgency of these programs, I mean, will you have some customers that will be deploying Scorpio X in production Rack-Scale deployments second half of this year?

Yes. So I think if you look at a couple of years out, the vision that we have as Astera Labs is to be providing the connectivity infrastructure at the rack level, which includes the switch, which includes the retimers, active electrical cables or whatever components that might be needed. It's copper today, maybe it's going to be optical in the future. So that's our vision, and we will continue to support our customers towards that vision even starting now. So to specifically your question on Scorpio X, we are already in preproduction. The qualification work has started. It will continue through the rest of the year where we start to achieve production volumes for the Scorpio X family towards the end of the year. But really, the big opportunity for X is going to be in 2026 when it gets deployed at the rack level.

As you engage with customers on Scorpio X, are most of them choosing a PCIe-like protocol? Are they choosing proprietary protocol or the new industry standard-based UALink protocol? And maybe if you can talk a little bit more about the industry standard UALink initiative and the benefit it can bring versus proprietary or Ethernet-based solutions.

Yes. So today, it's a diverse ecosystem. Everybody is using something that's proprietary, like, for example, in NVLink or something that might be already existing, which is, for example, PCI Express or PCI Express like protocols, but also, in some cases, Ethernet because just Ethernet happens to be available, a switch happens to be available today for Ethernet. So it's a fragmented space today, and that's why we feel very hopeful that UAL, which is a new standard, will be sort of the one unifying standard for rack level connectivity. Now UALink or Ultra Accelerator Link is an industry-wide consortium and Astera Labs is proud to be a promoter member of that Board. But what UAL does is it builds this protocol from the ground up for AI so the whole protocol is targeted for scale-up networks connecting AI accelerators, ASICs or GPUs to each other. And so what they did is they started with PCI Express because PCI Express natively has the capability to connect GPUs together and make them look like 1 large GPU. So they started with that, the memory semantics, the loss less nature of PCI Express is all carried over into UALink. And then they throw away anything that's not necessary. So you even streamline PCI Express protocol as part of UAL. And then they couple it with the fastest SerDes available from Ethernet so you're getting the best of both worlds with UAL and just even more importantly, you're getting a very good ecosystem where there will be many vendors providing solution into this space and customers looking to adopt it in a truly open ecosystem. Now if you contrast that with UEC. UEC also will eventually do the same thing, but they are starting with an Ethernet protocol, which was not meant for scale-up. And they're borrowing many of the features that are present in UALink and in PCI Express into UEC so by the time it's all said and done, maybe it is Ethernet in name, but a lot of niches that make it not like Ethernet. So we will see. I think the biggest difference, though, is again going to be in the ecosystem. UEC is driven by 1 large company today even though it's an open ecosystem. Over time, how it evolves, it remains to be seen. But I do think that hyperscaler customers are looking for a truly diverse open ecosystem for their Rack-Scale connectivity solutions.

Any questions from the audience?

Jitendra, I just wanted to follow up on the comments you made about the move to Rack-Scale being a real positive for the company. Does that also apply to NVIDIA-based racks?

Yes. So we are very happy with the progress that we've made with the NVIDIA platform, Blackwell, in particular so as we have said before, Hopper to Blackwell is a positive factor for us, our content grows up very significantly. However, I think your question is specifically to NVIDIA reference designs that use Blackwell. So in those reference designs, we have minimal content as we move forward from the Hopper generation to Blackwell. But that is much more than offset by the customized designs that the hyperscalers are doing using the same great technology that Blackwell offers and then customizing it for deployment into their own data centers. We get content for our retimers there, we get content for our Scorpio devices. So on the whole, it's a big positive for us.

Software. Let's talk about software. Software is a key part of your silicon solution. It's a key part of actually the differentiation, right? In fact, your COSMOS software and management platform is deeply integrated into your customers' data center software management stacks, right? This is a very sticky part of this solution. Cosmos integrated in your chips enables things like real-time diagnostics, telemetry, debugging, right, ultimately enabling faster time to market, faster bringing up of compute clusters. And so as you bring new solutions to the market like Scorpio P and Scorpio X Series products and continue to expand your product line, is the team continuing to leverage its COSMOS software? And is this still a major competitive advantage for the team.

Absolutely, 100%. In fact, I'll make an anecdote, I talk to somebody one of the hyperscaler customers and say, what's easier, setting up a 100,000 cluster or continuing to run 100,000 cluster and they didn't even blink that running a 100,000 cluster is far more difficult than setting 1 up. So while everything you said this is very important, Harlan, in how you bring this up and how you debug this, the ability to provide just really intense detailed diagnostics on what's going on, on this cluster is super critical. And so with our Aries products, we are able to provide 1 level of diagnostics and link health monitoring, environmental sensing and whatnot. But now with Scorpio in the mix, these 2 products are better together in that not only can you get linked level diagnostics, you can also figure out where is congestion in the system, how full the buffers are inside of the switch. And so together, we provide even more diagnostics all under the same COSMOS umbrella. So it just makes the solution more sticky. And we are investing a lot in our software. Our software team is already bigger than our design team.

Let's talk about your Taurus product family of AEC networking connectivity, 800 gig Ethernet connectivity, obviously, strong adoption with current AI scale-out networking connectivity, right? But we're not actually starting to see the adoption curve of 800 gig within general-purpose cloud and data center footprint starting to pick up, right? Very early days, but starting to see some momentum. 200 gig, 400 gig within the data center starting to maybe lose a little bit of steam, right? And so I guess the question for the team is, are you still on track to shipping your 800-gig AEC solution in volume production this year? And will it be with multiple customers?

Yes. So just for some background, our current solution is based on 50 gigabit per second per lane technology that gets deployed as 200-gig and 400-gig modules. And that is largely with 1 hyperscaler who has chosen to deploy this active electrical cable technology. However, within that 1 hyperscaler, it is actually quite diversified, both on AI platform as well as our robust compute platform, different form factors, straight cable, X cable, cable, et cetera. Now as we go from 50 gig per lane to 100 gigabit per second per lane, you get to 800 gig modules, and we are ready to get those deployed towards the end of the year. And now what will happen with 800 gig is there will be some fragmentation still in the market. Some customers will choose to deploy it with passive copper cable, thicker cables maybe they are shorter or sometimes you need to go to the end of the row so you need optical solutions because copper just doesn't reach it. But there will be that sweet spot in the middle where active electrical cable solutions will be applicable, and we are 100% ready to address those sockets.

And to your point, there's an ongoing sort of market debate regarding the choice between AEC versus ACC solutions. But as you engage with your customers, how are they evaluating these 2 options and how do you envision these technologies sort of evolving over time?

So the difference between AEC and ACC is the following. AEC uses Retimer, which is a component that completely cleans up the signal. It receives signals, it understands it and retransmit completely a copy of that. ACC uses something a lot simpler called a redriver, which is effectively an amplifier. So it takes in the signal that's coming in, amplifies it and amplifies all the noise and other imperfections that are coming with the signal and sends it out on the other side. So it is a kind of a not as powerful as signal conditioning technique. However, the advantage of redrivers is that they are low power. So if you can share the simplest things in my mind at least, if you can afford the power, you will always go with a retimer solution because it provides you with all of the diagnostics and the performance that you need. But if power is a constraint, then redriver solutions are more applicable and typically, they work when you control both sides of the equation. So if you control both the source and the destination with difficulty, you can make a redriver work. And to the extent that our customers want it, we will build 1 ourselves as well. they are much easier to build than a retimer class solution.

On the financial strong gross margins, the team drove strong gross margins. 74% is the guidance for the June quarter, that's tracking 400 basis points above your long-term target model of 70%. While you plan to increase the mix of hardware products over the next several years, which could exert some pressure on gross margins. You also actually have a whole bunch of new products that we just talked about, right, like the Aries 6 solution, the Scorpio products that are beginning to ramp. And I believe Scorpio does carry very high gross margin profile, given the performance, complexity of the chipsets and the software attached. And so Scorpio obviously is going to continue to scale higher. So does the team have the potential to sustain margins above the long-term target of 70%.

We still like people to expect the long-term gross margin target of 70%. As we diversify into multiple product lines, we'll have a wider range margins per product line as well. And just given all the opportunities we have, we think, over time, we'll trend towards 70% still.

On operating expenses, I mean, they've been biasing higher as you continue to maintain a very, very strong product cadence and expand your networking and connectivity portfolio. How should we think about the trajectory of your operating expenses as we think about it over a multiyear period, as the team continues to build scale, how do we think about your OpEx growth relative to your revenue growth?

We're still aggressively investing in R&D and expect OpEx to grow. But with the revenue growth ahead of us, we do expect to see leverage over the long term and deliver 40% operating margins in our target model.

Jitendra, Mike, thank you very much for your participation and look forward to monitoring the progress of the team this year.

Thanks, Harlan.

Thank you.