Q3 2024 Astera Labs Inc Earnings Call
Jeremy: Hello, good afternoon. My name is Jeremy and I'll be your conference operator today. At this time I would like to welcome everyone to the Astero Labs Q3-2024 earnings conference call. I'll want to be in place to meet Revitany back around noise.
After Minitions for Marks, there will be a question in the answer session. Do you like to ask a question during this time, simply press star, followed by the number 1 on your telephone keypad.
Jeremy: Do you like to withdraw your questions simply press the pound key. We do ask if you're asking a new question to limit yourself to one question plus one follow-up. Thank you. I'll now turn the call over to Leslie Green and best of relations for Stair Labs. Leslie, you may begin.
Leslie Green: Thank you, Jeremy and good afternoon everyone and welcome to the Astera Labs 3rd Quarter 2024 earnings conference call.
Jeremy: Joining us on the call today are Jitendra Mohan, Chief Executive Officer and co-founder, Sanjay Gajendra, President and Chief Executive Officer and co-founder and Mike Tate Chief Financial Officer.
Jeremy: Before we get started I would like to remind everyone that certain comments made in this call today may include forward-looking statements regarding among other things, expected future financial results.
Jeremy: Strategies and Plans, Future Operations, and the markets in which we operate. These forward-looking statements reflect management's current beliefs, expectations, and assumptions about future events, which are inherently subject to risks and uncertainties that are discussed in today's earnings release and the periodic reports we file from time to time with the SEC, including risks set forth in the final perspective relating to our IPO.
Jeremy: It is not possible for the company's management to predict all risks and uncertainties that could have an impact on these forward-looking statements or the extent to which any factor or combination of factors may cause actual results to different materialy from those contained in any forward-looking statement.
Jeremy: In light of these risks and uncertainties and assumptions, the results, events, or circumstances related and reflected in the forward-looking statement
Jeremy: Discussed.
Jeremy: During this call may not occur and actual results could differ materially from those anticipated or implied.
Jeremy: All of our statements are made based on information available to management as of today and the company undertakes no obligation to update such statements.
Jeremy: After the date of this call to conform to these as a result of new information future events or changes in our expectations except as required by law
Jeremy: Also, during this call, we will refer to certain non-GAAP financial measures, which we consider to be important measure of the company's performance.
Jeremy: These non-GAAP financial measures are provided in addition to and not as a substitute for or superior to financial results prepared in accordance with U.S. GAAP.
Jeremy: A discussion of why we use non-GAAP financial measures and reconciliations between our GAAP and non-GAAP financial measures is available in the earnings release we issued today, which can be accessed through the investor relations portion of our website and will also be included in our filings with the SEC, which will also be accessible through the investor relations portion of our website.
Speaker Change: With that, I would like to turn the call over to Jitendra Mohan, CEO of Astera Labs.
Jitendra Mohan: Thank you, Leslie. Good afternoon, everyone. And thanks for joining our third quarter conference call for fiscal year twenty twenty four.
Jitendra Mohan: Today, I will provide a quick Q3 financial overview, discuss several of our recent company-specific product and strategic announcements, followed by a discussion around the key secular trends driving our market opportunity.
Jeremy: I will then turn the call over to Sanjay to delve deeper into our growth strategy.
Jeremy: Finally, Michael provided additional details on our Q3 results and our Q4 financial guidance.
Jeremy: Estera Labs delivered strong Q3 results, setting our 5th consecutive record for quarterly revenue at $113 million.
Jeremy: which was up 47% from the last quarter and up 206% versus the prior year. Our business has entered a new growth phase with multiple product families ramping across AI platforms featuring both third-party GPUs and internally developed AI accelerators.
Jitendra Mohan: which drove the Q3 sales upside versus our guidance.
Jeremy: We also demonstrated strong operating leverage during Q3, with non-GAAP operating margin expanding to over 32% and delivered non-GAAP EPS of $0.23.
Jeremy: Looking into Q4, we expect our revenue momentum to continue, largely driven by the Ares PCIe and Taurus Ethernet product lines, the Scorpio fabrics which is continuing to ship in pre-production volumes.
Jeremy: The criticality of connectivity in modern AI clusters continues to grow with trillion-parameter model sizes, multi-step reasoning models, and faster, more complex AI accelerators.
Jeremy: These developments present a tremendous opportunity for Estera Labs' intelligent connectivity platform to enhance AI server performance and productivity with our differentiated hardware and software solutions.
Jeremy: At the start of Q4, we announced our fourth product line, the Scorpio Smart Fabric Switch family, which expands our mission of solving the increasingly complex connectivity challenges within AI infrastructure, both for scale-out and for scale-up networks.
Jeremy: Extending beyond our current market footprint of PCI Express and Ethernet retimer-class products and controller-class devices for CXL memory, our Scorpio SmartFabric switch family delivers meaningfully higher functionality and value to our AI and cloud infrastructure customers.
Jeremy: We estimate that Scorpio will expand our total market opportunity for our four product families to more than $12 billion by 2028.
Jeremy: Scorpio family unlocks a large and growing opportunity across AI server head node scale out applications with the P-series and AI accelerator scale up clustering use cases with the X-series.
Jeremy: The P-series devices directly address the challenge of keeping modern GPUs fed with data at ever-increasing speeds, while the X-series devices improve the efficiency and size of AI clusters.
Jeremy: The Scorpio family was purpose-built from the ground up for these AI-specific workloads with close alignment with our hyperscaler and AI platform partners.
Jeremy: At the recent 2024 OCP Global Summit, we demonstrated the industry's first PCIe Gen 6 Fabric Switch, which is currently shipping in pre-production volumes for AI platforms.
Jeremy: We are happy to report that we already have design wins for both Scorpio P-Series and X-Series, and that our recent product launches further accelerated strong customer and ecosystem interest.
Jeremy: Over the coming quarters, we expect to further expand our business opportunities for the Scorpio product family across PCIe Gen 5, PCIe Gen 6, and platform-specific customized connectivity platforms.
Jeremy: Further expanding our market opportunity, Estera Labs has joined the Ultra Accelerator Link UA Link Consortium as a promoting member on the Board of Directors, along with industry-leading hyperscalers and AI platform providers.
Jeremy: This important industry initiative places us at the forefront of developing and advancing an open high-speed, low-latency interconnect for scale-up connectivity between accelerators.
Jeremy: SR Labs' deep expertise in developing advanced silicon-based connectivity solutions, along with a strong track record of technology execution, makes us uniquely suited to contribute to this compelling and necessary industry initiative.
Jeremy: In a shift towards shorter AI platform refresh cycles, hyperscalers are increasingly relying on their trusted partners as they deploy new hardware in their data center infrastructure at an unprecedented pace and scale.
Jeremy: Today, we have demonstrated strong execution with our Aries, Taurus, Leo, and now Scorpio product families.
Jeremy: Our products increase data, networking, memory bandwidth, and capacity, and our Cosmos software provides our hyperscaler customers with the tools necessary to monitor and observe the health of their expensive infrastructure to ensure maximum system utilization.
Jeremy: To conclude, Estera Labs' expanding product portfolio, including the new Scorpio smart fabric switches, is cementing our position as a critical part of AI connectivity infrastructure.
Jeremy: Delivering increased value to our hyperscaler customers and unlocking additional multi-year growth trajectories for Astera Labs.
Speaker Change: With that, let me turn the call over to our President and COO, Sanjay Gajendra, to discuss some of our recent product announcements and our long-term growth strategy.
Sanjay Gajendra: Thanks Jitendra and good afternoon everyone. We are pleased with our Q3 results and strong financial outlook for Q4. Overall, we believe we have entered a new phase at the company based on two key factors.
Jeremy: First, is the increased diversity of our business. In 3Q, our business diversified significantly with new product lines and form factors going to high-volume production.
Sanjay Gajendra: We also started ramping on multiple new AI platforms based on internally developed AI accelerators at multiple customers to go along with the continued momentum with third-party GPU-based AI platforms.
Jeremy: These together helped achieve a record sequential growth in 3Q.
Jeremy: Second, with the introduction of Scorpio smart fabric switches, our market opportunity has significantly expanded.
Jeremy: This new product line delivers increased value to our hyperscaler customers and for Estera Labs, unlocks higher dollar content in AI platforms and additional multi-year growth trajectories.
Jeremy: Let me explain our business drivers in more detail.
Jeremy: As noted, we now have multiple product lines, generations, and form factors in high-volume production.
Jeremy: This includes 80 smart DSP re-timers and smart cable modules for PCIe 5.0 and total smart cable modules for 200 and 400 gig Ethernet active electrical cables.
Jeremy: We are also shipping pre-production volumes for Leo CXL memory controllers, Ares retimers for PCIe 6.0, and Scorpio Fabric switch solutions for PCIe head node connectivity.
Jeremy: All these new products deliver increased value to our customers and therefore command higher ASPs.
Jeremy: Our first to market Scorpio PCIe Gen 6 Fabric Switch.
Jeremy: addresses a multi-billion dollar opportunity with a ground-up architecture designed for AI data flows.
Jeremy: and delivers maximum predictable performance per watt in mixed-mode PCIe headnode connectivity compared to incumbent solutions.
Jeremy: We are currently shipping Scorpio P-Series in pre-production quantities to support qualification for customized AI platforms based upon leading third-party GPUs.
Jeremy: Interest for Scorpio P-Series has accelerated since the former launch given its differentiated features and as a result we are engaging in multiple design opportunities across a diverse spectrum of AI platforms.
Jeremy: These include both PCIe Gen 6 and PCIe Gen 5 implementations on third-party GPU and internal accelerator-based platforms.
Jeremy: Overall, we are very bullish on the market for our entire product portfolio across third-party GPU-based systems with increasing dollar content per GPU on our new design wins.
Jeremy: And we believe that Estera's opportunity with internally developed AI accelerator platforms can be even larger with opportunities both in the scale-out and back-end scale-up clustering use cases.
Jeremy: This additional market opportunity has unlocked design activity for Aries and Taurus product lines.
Jeremy: for reach extension within and between racks, and for our newly introduced Scorpio X-series fabric switches for homogeneous accelerator-to-accelerator connectivity with maximum bandwidth.
Jeremy: Scorpio X-Series is built upon our software-defined architecture and leverages our Cosmos software suite to support a variety of platform-specific customization, which provides hyperscalers with valuable flexibility.
Jeremy: As we look ahead to 2025, we will begin to ramp designs across new internally developed AI accelerator-based platforms that will incorporate multiple Estera Labs product families, including Ares, Taurus, and Scorpio.
Jeremy: As a result, we will continue to benefit from increased dollar content per accelerator in these next-generation AI infrastructure deployments.
Jeremy: Though we remain laser-focused on AI platforms, we continue to see large and growing market opportunities within the general-purpose compute space for our PCIe, Ethernet, and CXL product families.
Jeremy: While the transition to faster bandwidth requirements within general purpose computing trails the leading adoption across AI systems, the market size remains substantial.
Jeremy: and the introduction of new CPU generations from Intel and AMD. We are also ramping volume production of our Taurus SEMs for front-end networking across hyperscaler general-purpose server platforms.
Jeremy: We expect to see broadening adoption of Leo CXL memory controllers across the ecosystem as CXL-capable server CPUs are deployed in new cloud infrastructure over the coming years.
Jeremy: In summary, we believe Astera Labs has entered a new growth phase, and we are well positioned to outpace industry growth rates through a combination of strong secular tailwinds and the expansion of our silicon content opportunity in AI and general purpose cloud platforms.
Jeremy: We have become a trusted and strategic partner to our customers with over 10 million smart connectivity solutions widely deployed and field tested across nearly all major AI infrastructure programs globally.
Jeremy: The introduction of the Scorpio Smart Fabric Switch family and our strategic involvement with the UALink standard for scale-up connectivity is the next critical step in our corporate journey.
Jeremy: We are hard at work collaborating with our partners.
Jeremy: to identify and develop new technologies that will expand Estera's footprint from retimer solutions for connectivity or copper within the rack to fabric and optical solutions that connect AI accelerators across the data center.
Jeremy: While we have come a long way, there's a remarkable sense of urgency and energy within the company for the opportunities that lay ahead.
Speaker Change: With that, I will turn the call over to our CFO, Mike Tate, who will discuss our Q3 financial results and our Q4 outlook.
Mike Tate: Thanks, Sanjay. And thanks to everyone for joining the call. This overview of our Q3 financial results and Q4 guidance will be on a non-GAAP basis. The primary difference in Estera Labs' non-GAAP metrics is stock-based compensation and its related income tax effects.
Mike Tate: Please refer to today's press release available on the Investor Relations section of our website for more details on both our GAAP and non-GAAP Q4 financial outlook, as well as a reconciliation of our GAAP to non-GAAP financial measures presented on this call.
Mike Tate: For Q3 of 2024, Estera Labs delivered record quarterly revenue of $113.1 million, which was up 47% versus the previous quarter, and 206% higher than the revenue in Q3 of 2023.
Mike Tate: During the quarter, we ship products to all major hyperscalers and AI accelerator manufacturers. And we recognize revenue across all four of our product families.
Mike Tate: Our Aries product family continues to be our largest sales contributor and help drive the upside in our revenues this quarter.
Mike Tate: ARIES revenues are being driven by continued momentum with third-party GPU-based AI platforms, as well as strong ramps on new platforms featuring internally developed AI accelerators from multiple hyperscaler customers.
Mike Tate: Also in Q3, Taurus Revenue started to diversify beyond 200GB applications, with an initial production ramp of our 400GB Ethernet-based systems, which are designed into both AI and general-purpose compute systems.
Mike Tate: Q3 LEO CXL revenues continue to be driven by our customers purchasing pre-production volumes for the development of their next-generation CXL-capable compute platforms.
Jeremy: Lastly, we began to ship pre-production volumes of our recently announced Scorpio Smart Fabric Switch family during Q3.
Jeremy: Q3 non-GAAP gross margins was 77.8% and was down 20 basis points compared with 78% in Q2 of 2024 and better than our guidance of 75% driven by higher sales volume and a favorable product mix.
Jeremy: Non-GAAP operating expenses for Q3 were $51.2 million, up from $41.1 million in the previous quarter, driven by greater-than-expected hiring conversion during the quarter, as we aggressively pushed to support additional commercial opportunities.
Jeremy: Within Q3 non-GAAP operating expenses, R&D expenses was $36 million, sales and marketing expenses was $7 million, and general and administrative expenses was $8.3 million.
Jeremy: Non-GAAP operating margin for Q3 was 32.4%, up from 24.4% in Q2, and demonstrated strong operating leverage as revenue growth outpaced higher operating expenses.
Jeremy: Interest income in Q3 was $10.9 million.
Jeremy: Our non-GAAP tax provision was $7.3 million for the quarter, which represented a tax rate of 15% on a non-GAAP basis.
Jeremy: Non-GAAP fully diluted share count for Q3 was 173.8 million shares, and our non-GAAP diluted earnings per share for the quarter was 23 cents.
Jeremy: Cash flow from operating activities for Q3 was $63.5 million, and we ended the quarter with cash equivalents and marketable securities of $887 million.
Jeremy: Now turning to our guidance for Q4 of fiscal 2024.
Jeremy: We expect Q4 revenue to increase to within a range of $126 million and $130 million, up roughly 11% to 15% from the prior quarter.
Jeremy: For Q4, we expect continued strong growth from our Ares product family across a diverse set of AI platforms.
Jeremy: some of which are just starting to ramp, and also from our TORIES SEM for 400 gig applications and additional pre-production shipments of our Scorpio P-Series switches.
Jeremy: We expect non-gap gross margins to be approximately 75%. The sequential decline in gross margins is driven by an expected product mix shift towards hardware solutions during the quarter.
Jeremy: We expect non-GAAP operating expenses to be in a range of approximately $54 million to $55 million as we continue to expand our R&D resource pool across headcount and intellectual property.
Jeremy: Interest income is expected to be approximately $10 million. Our non-GAAP tax rate should be approximately 10%. And our non-GAAP fully diluted share count is expected to be approximately 179 million shares.
Jeremy: This concludes our prepared remarks, and once again, we appreciate everyone joining the call, and now we will open the line up for questions. Operator?
Speaker Change: Alright, thank you. And as a reminder, if you would like to ask a question today, please simply or simply just press star followed by the number one.
Speaker Change: I'll give you one moment to compile the roster.
Speaker Change: Okay, our first question comes from Harlan Sir from JPMorgan. Please go ahead.
Harlan Sir: Good afternoon and congratulations on the.
Harlan Sir: strong results. You know, on your core retimer business looks like, you know, very strong this year, looks strong next year, majority of the XPU shipments are still going to be
Speaker Change: and your customers, you know, both Merchant and ASIC are ramping new SKUs starting second half of this year, first half of next year. You've also got the ramp of your Gen 6 products, re-timers.
Speaker Change: Scorpio Switch products with your lead GPU customer, which they're ramping now, AC firing, SCM firing. So given all of this activity, I assume your visibility and backlog is quite strong, maybe if you can just...
Speaker Change: qualitatively sort of articulate your confidence on driving sequential growth over the next several quarters or at least visibility for a strong first half of next year versus second half of this year.
Speaker Change: Yeah, thanks, Harlan. Yeah, right now our visibility is very strong on both.
Speaker Change: you know, as always with our backlogged position, but also
Speaker Change: the breadth of designs we have.
Speaker Change: Right now, we're really kind of entering a new phase of growth here where our revenue streams are clearly diversifying if you look at right now versus a year ago.
Speaker Change: We're very excited about the breadth of designs and product lines that we're ramping in.
Speaker Change: So in the back half of this year.
Speaker Change: Obviously, TORS has been very incremental, and that continues in Q4, and also...
Speaker Change: The programs that we're ramping on Taurus are just getting started, so we have good confidence that Taurus will continue to grow nicely into 2025. Aries, as he alluded to, the Gen 5...
Speaker Change: Unknown Speaker Story is still got a lot of legs. We have both with the merchant GPUs, but also on these internal ASIC designs, a lot of those programs are just starting. And with the Gen 5, you know, we have both the front end connectivity and the back end connectivity.
Speaker Change: Incrementally, Gen 6 will start to play out as well, and with Gen 6, we get an ASP boost on top of that.
Speaker Change: and then finally with LEO, you know we've been talking about LEO for quite a while but these are new types of technologies that are being adopted in the marketplace and we're happy to report that we do have line of sight to our first production shipments starting in the middle of the year next year.
Speaker Change: Thank you for that.
Speaker Change: and then on the early traction you're getting with your Scorpio switch portfolio.
Speaker Change: You know, the team has talked on some of the performance benefits both at the
Speaker Change: Silicon level and chip level, but how much of the differentiator is your Cosmos software stack that your customers have already built into their data center management systems?
Speaker Change: Adopted it with your Aries Retimer products.
Speaker Change: And now you've got the same software stack is integrated into your switch products that enables telemetry link performance
Speaker Change: link monitoring, tuning of parameters, all that kind of stuff, which is so critical for data center managers. Is the software stack familiarity, current adoption, sort of a big differentiator on the traction with Scorpio?
Speaker Change: and you are absolutely right. There are several things that differentiate Scorpio family. First and foremost, I would say that we build this ground up for AI applications.
Speaker Change: If you think about historical deployments of PCI Express switches, they were generally built for server applications, for storage, and things like that.
Speaker Change: which are quite different from AI. So we designed the chip for AI applications. We put in the performance that's required.
Speaker Change: for these AI applications.
Speaker Change: In addition, even the form factor has been designed for AI. So, rather than building a large switch, we ended up building a smaller device, such that you are not running these high-speed signals all over the board.
Speaker Change: So all of that is very purpose-built for AI. Now to your point about software, as you remember, our chips are architected with a software-first approach.
Speaker Change: So we can deliver a lot of customization based on Cosmos, which is something that our hyperscaler customers are looking for, especially for the X family.
Speaker Change: which is deployed in a more homogeneous GPU-to-GPU connectivity.
Speaker Change: Where the Scorpio family sits, we have access to a lot more diagnostics information, and we can couple that with the information that we are collecting from our other families deployed, such as Aries and even Taurus, to provide a holistic view of the AI infrastructure to the data center operators.
Speaker Change: So both from the hardware side, the kind of purpose-built nature of these devices, as well as the software stack that comes with it, is a big differentiator for us.
Speaker Change: Very insightful. Thank you.
Speaker Change: Our next question comes from the line of Joe Moore from Morgan Stanley. Please go ahead.
Joe Moore: Great, thank you and congratulations on the numbers. I wonder if you could talk about the
Joe Moore: Scorpio Business in terms of, you gave a number for 2028. Can you give us a sense for what that ramp looks like?
Joe Moore: Any kind of qualitative discussion of how big it could be in calendar 25 would be very helpful. Thank you.
Speaker Change: Yeah, thanks Joe. Yeah.
Speaker Change: What's exciting is, you know, since our public release, we're getting a lot of incremental excitement from the customer base.
Speaker Change: and what's really neat about the Scorpio product family is the diversity of the designs that we're seeing. So clearly being first to the Gen 6, there's a lot of interest in that application.
Speaker Change: But there's still a lot of Gen 5 opportunities that are developing that we're addressing as well, you know, because with a Gen 6 capable switch, it's backwards compatible. So both both design opportunities are open for us.
Speaker Change: and then incrementally, we have the X-Series that does the backend connectivity, and that's kind of a greenfield opportunity as well.
Speaker Change: So these designs are generally a little more customized system, so the bring up in the qualification cycle is a little bit more long, so we take a conservative view on how they ramp, like we always do with most of our business.
Speaker Change: But overall, you know, given that we expect Scorpio to exceed 10% of our revenues in 2025.
Speaker Change: as the deployments get into production during the course of the year and exit the year at a very good run rate and good momentum going into 2026.
Speaker Change: That's great. Thank you. And then on Leo, you talked about
Speaker Change: Some of the ramps there. I guess this application in particular of these large memory servers being able to actually reduce the CPU count and maintain this high memory bandwidth. Can you just talk about that application and are you seeing that as a material factor next year as well?
Speaker Change: Yes, so Joseph Sanjay here. Yes, I think like we've been maintaining.
Speaker Change: And with any new technology, it takes a little bit of time for things to mature. So right now, the way we look at CXL is it's a transition from the sort of the crawl stage to walk stage.
Speaker Change: There are three or four key use cases that have emerged. One of them is what you noted, which is the large in-memory database applications.
Speaker Change: and there the use case becomes one of how do you enable more memory capacity.
Speaker Change: In the past, this was done by adding additional CPUs into the server box.
Speaker Change: to provide for more memory channels.
Speaker Change: But what we have demonstrated is that by using Leo, you're not only able to get the higher performance by the added memory, but from an overall TCO standpoint, it's significantly less than adding additional CPUs.
Speaker Change: So that's one key use case that we see from a deployment standpoint.
Speaker Change: But having said that, I think at OCP this year, you might have seen some of our videos and all that, there's been tremendous amount of different platform level solutions that were being deployed for various high bandwidth applications, HPC applications.
Speaker Change: including some of the rack-level disaggregation type of use cases. So to that standpoint, there are many different ways in which the technology can develop, but just like any new technology, it will take some time before...
Speaker Change: The requisite ecosystem and software is built. So we are in that period right now, getting those pieces in place. And 2025 is when we expect the production volumes to begin.
Speaker Change: Great, thank you.
Blaine Curtis: Our next question comes from the line of Blaine Curtis from Jeffries. Please go ahead.
Blaine Curtis: Hey, good afternoon, and congrats on the results. I want to ask you, last quarter, you kind of talked about the September growth, I think it was like 20 million, and you kind of loosely said it was kind of one third retimers, one third the PCIE.
Speaker Change: between your products if you could add a little bit more color and then also just
Speaker Change: Between AI accelerators and GPUs, obviously the big GPU vendor has a ramp coming with Blackwell, but that's not exactly now. So I'm just kind of curious, you know, what's driving the strength in September and December a little bit more.
Speaker Change: Yeah, thanks, Blaine.
Speaker Change: Yeah, I you know in Q3 our business is really benefiting from the strong contribution of multiple product lines and
Speaker Change: The Ares SEM and Taurus both was a really big strong ramp quarters. Those ramps were consistent with our expectations.
Speaker Change: The upside to the guidance was driven largely from ARIES revenue, both for the third-party GPUs, but also as well with the strong ramps on new platforms.
Speaker Change: on the internally developed AI accelerators. And we're seeing that across multiple hyperscaler customers. So it's not just one. So the upside was largely driven by that Ares revenue.
Speaker Change: Thank you.
Speaker Change: Yeah, so for Scorpio, there will be a broader range of margins.
Speaker Change: There's different use cases, so it depends on the functionality and the volume. But at this point, we believe Scorpio will not impact our long-term growth margin targets at 70%, and it was kind of contemplated when we set up those targets.
Speaker Change: I'd say overall, beyond just Scorpio though, we are seeing a wider range of margin profiles across all our product portfolios. So, you know, mix will be important to contribute from a quarter to the quarter perspective, but we still feel good about the 70% target.
Speaker Change: Thanks Matt.
Speaker Change: Our next question comes from the line of Thomas O'Malley from Barclays. Please go ahead.
Thomas O'malley: Hey guys, thanks for taking my questions
Thomas O'malley: My first one is just on the X-Series for Scorpio. I think this is the first real kind of back-end switch that we've seen in the marketplace for PCIe. Can you talk about your positioning there, how far you think you're ahead, and would you expect the same kind of competitive dynamic that you're seeing in the P-Series switch?
Speaker Change: where you are competitively and just from an opportunity perspective, do you think longer term, the X series is a bigger opportunity than the P series?
Speaker Change: Unknown Speaker
Speaker Change: Tom, let me take that question. This is Jitendra. So you have three questions in there. So you don't get a follow up.
Speaker Change: So the first question, let me, let me ask with the P-series first.
Speaker Change: So P-series is actually broadly applicable to all of the accelerators. All of the accelerators require connectivity from the GPU or the accelerator to the head node or to the NICs and so on.
Speaker Change: So P-series is applicable to all of them. The P family for us supports PCI Express Gen 6.
Speaker Change: So that's where the deployment will happen. I already mentioned what are some of the advantages of the family But at the same time we should not underestimate the Gen 5 socket There are also Gen 5 designs that are taking place right now both with the third-party GPUs as well as with ASICs
Speaker Change: internally developed ASICs. So I think that's the market opportunity we see with the P-series. We estimate that the time for this P-series to be about a billion dollars plus or minus today and growing to double of that over time.
Speaker Change: But you're also correct that we do think over the longer period, X series will have a bigger TAM. The TAM today is nearly zero. It's not very commonly used outside of the NVIDIA ecosystem.
Speaker Change: We do expect many hyperscalers to start deploying this, starting with the X family and the designs for which that we have.
Speaker Change: And we are able to do that because of the architecture that we have.
Speaker Change: Because of our software-defined architecture, we can customize many parts of the X-series to cater to the specific requirements of the hyperscalers, both on the side of performance
Speaker Change: The exact configuration that they require, lane count and so on, and also the diagnostics framework that they require to monitor their infrastructure.
Speaker Change: So over time, we do expect X-Series to become larger. Now, I also mentioned during the remarks that we have joined the UA Link Consortium, and that gives us another market, another opportunity where we can play with backend interconnect.
Sanjay Gajendra: Yeah, this is Sanjay here. Let me take that. Yeah, so 2025 is the year that we think, you know, the business will get broader. As we've always said, AC or active electrical cable is a case-by-case situation, meaning it's not like every hyperscaler uses active electrical cables.
Sanjay Gajendra: So to that standpoint, you know, we do expect as data rates go higher with 800 gig and so on, that market to be much more diversified than what it is today.
Sanjay Gajendra: All right, so with that said, I mean, today, if you look at our business, we do have our AC modules or Taurus modules going into both AI and compute platforms.
Sanjay Gajendra: There are different kinds of cables in terms of various different configurations that they need to support. So overall, I want to say we are fairly diversified with our business today, but as the speeds increase,
Sanjay Gajendra: in 2025 and beyond. We do expect that the customer basis will continue to evolve.
Sanjay Gajendra: with the note that I highlighted that every infrastructure will be different.
Speaker Change: Our next question comes from the line of Torres-Vandenberg from Stifle. Please go ahead.
Torres-Vandenberg: Yes thank you and congrats on the strong results. You mentioned that PCIe Gen 6 is going to be in pre-production this quarter. When do you expect it to be an actual volume production? Would that be in the first half of next year Q1?
Speaker Change: I want to say it depends on the customer's timeline, so we don't want to speak for any of our customers on what they are communicating.
Torres-Vandenberg: But in general, what I would say, though, similar to what we've shared with you in the past, that our design ends are in the customized...
Torres-Vandenberg: RAC implementation. So to that standpoint the timing of qualification and the deployment would be based on that.
Torres-Vandenberg: But in general, between the initial design wins we had to where we are now, where we are.
Torres-Vandenberg: you know, engaging with multiple opportunities, both for Gen 5 and Gen 6, and both for, you know, third-party GPUs as well as internally developed GPUs.
Torres-Vandenberg: So to that standpoint, our opportunities on Scorpio continues to grow.
Speaker Change: and as you look at overall for 2025 like Mike suggested, we do expect our contribution from Scorpio to be in excess of 10% of our overall revenue.
Speaker Change: That's very helpful. And I had a follow-up question on Scorpio in relation to your PCIe retimer business. So,
Speaker Change: Would those typically pull each other, meaning, you know, could it be instances where there's a switch, a PCIe switch, with somebody else's read timers, or, I mean, do they pretty much go hand in hand, especially given the Cosmos software?
Speaker Change: Yes, so if you look at today's deployment with Gen 5, at least from an industry snapshot standpoint, it's mix and match, right? Meaning our re-timers get used with our switch products from other vendors.
Speaker Change: We have gone, because of our software-based architecture, it allows us to uniquely customize and optimize for different system-level configurations.
Speaker Change: So that is what it is today going forward with Cosmos. We do see an advantage
Speaker Change: because we have integrated the management framework, the customization framework, and the optimization type of feature set across all of our products, meaning if a customer is using Cosmos today for Aries,
Speaker Change: They will very easily be able to extend the infrastructure that they've already built to run on top of our Scorpio devices. That's a unique advantage we bring compared to some of the alternatives out there.
Speaker Change: Very helpful. Thank you.
Speaker Change: Our next question comes from Ross Seymour from Deutsche Bank. Please go ahead.
Ross Seymour: Hi, thanks. So yes, a couple questions and congrats on the strong results. The first question, Jitendra, you mentioned, and Sanjay, you too, about the importance of the ASIC side of the business really ramping up strongly. What was the inflection point that's really driving that?
Speaker Change: Yes, I think in terms of the ASIC designs, I think it's fairly public knowledge now that all the hyperscalers have doubled down on the amount of investment they're doing for their own internal ASIC programs.
Speaker Change: The third-party GPUs obviously have done a great job but also hyperscalers are starting to realize where the money is in terms of their AI use cases and workloads.
Speaker Change: So, to that standpoint, we have been seeing an increased investment from hyperscalers in terms of their internal programs.
Speaker Change: And we are, of course, addressing those across all of our product lines.
Speaker Change: So if you look at our business today, like we highlighted in the prepared remarks, we have truly entered a new phase in terms of our overall business.
Speaker Change: where we not only have the third-party GPU-based designs, we also have several internally developed accelerator-based AI platforms, and then we have multiple of our product lines that are ramping on these platforms.
Speaker Change: The one caveat, one additional point I would note is that for the internally developed AI platforms.
Speaker Change: We get to play not only in the front-end network, connecting the GPU to CPU and storage.
Speaker Change: We also get to play in the back-end, which generally tends to be, like I call it, a fertile land, where there are multiple connectivity...
Speaker Change: requirements that we can service with our Ares and Taurus product line and now of course with the Scorpio X-series product line.
Speaker Change: Thanks for that. And I guess as my one follow up a quarter ago, we were having significant debates about your statements about the average content per GPU. Obviously, that's not as big a topic today.
Speaker Change: Now that we know about Scorpio with more detail, but if I revisit that and you talk still on this call about the average content going up, is that just because of Scorpio? Something you had in your back pocket before that you obviously couldn't tell us about?
Speaker Change: or do you still believe the retimer content in some way, shape or form will still be bigger on most of these platforms going forward, especially for the third party merchant suppliers?
Speaker Change: Yeah, so I think when we talked about it before, there were two reasons we highlighted. One is...
Speaker Change: Generally speaking, with each new generation of a protocol like PCI going from Gen 5 to Gen 6, there is an ASP uplift. That's number one. Number two, of course, we were hinting at the Scorpio product line, which because of the value it delivers to customers, is at a higher ASP, as you can imagine.
Speaker Change: So overall, if you look at the design wins we have today, the dollar content per GPU goes up. That's.
Speaker Change: One way to look at it based on what we've shared before.
Speaker Change: The other way to also look at it is that for internally developed platform, we get to play also in terms of the
Speaker Change: The Banking Network like in O2.
Speaker Change: we get to also address some of the scale out networks that are based on Ethernet using our Taurus module. So overall, if you look at sort of the increasing speed
Speaker Change: Additional product line, as well as the fact that the internally developed platforms, AI Accel database platforms, they are starting to gain more and more traction. So when you look at all of them, on an average, our content is on the up.
Speaker Change: Thank you.
Speaker Change: Our next question comes from the line of Quinn Bolton from Needham. Please go ahead. Thanks for squeezing me in. I just wanted to follow up, clarification maybe, but for the Scorpio family being over 10% of revenues, is that largely from the P-Series or would you expect
Speaker Change: Any X-series production revenue in 2025, given the longer, I think, design and cycle for the
Speaker Change: back-end scale-up networks.
Speaker Change: We have designs for both, and both P and X will contribute to revenues. The P designs will generally be first, but we do see X starting in the back half of the year as well.
Speaker Change: And then can you give us some sense, you know, for the P-series and the X-series?
Speaker Change: You know, on the retimer, I think the market sort of, you know, generally looked at.
Speaker Change: Unknown Speaker .
Speaker Change: you know, is there any framework you can provide for us for p-series, x-series? Would you look
Speaker Change: or expect a typical, you know, attach rate per GPU accelerator, would that be one to one? Would it be less than one to one?
Speaker Change: Could it be higher than one-to-one? Any thoughts on attach rates for
Speaker Change: P series and X series. Thank you.
Speaker Change: Yes, so it's a very broad question because there's all kinds of implementation that are out there. The high level, I'll probably share three points. The first is P-series is broadly applicable.
Speaker Change: in the sense that it could work for a
Speaker Change: a third party GPU or an internally developed
Speaker Change: accelerator because every accelerator doesn't matter where it comes from needs to connect to the head node side which generally includes the networking, the storage, as well as the CPU.
Speaker Change: So, to that standpoint, that will be a very broadly used device. And when it's used, it's one-to-one, meaning every GPU would need one of our Scorpio P-series devices.
Speaker Change: That's number one. Number two is the X-series. Now, these are generally used for GPU to GPU interconnect.
Speaker Change: So, to that standpoint, depending on the configuration, the number of devices is a function of the number of ports that an X-series device exposes and really depends on how the back-end fabric is built.
Speaker Change: And to that standpoint, again, it truly depends on how the configuration is being built.
Speaker Change: And this one, like Mike noted, is a Greenfield use case, meaning if you keep NVIDIA and their NVSwitch aside, everyone else is starting to build configurations that are obviously going to need some kind of a switching functionality, which is what we're addressing with our X-Series device.
Speaker Change: So that's the second point to keep in mind. And then in general, what I would say is that, overall,
Speaker Change: you know, depending on where things are and how big of a chip that we're building, the attached state will continue to evolve.
Speaker Change: In general, the dollar content that we're talking about is expected to continue to grow, both because we are adding more functionality with devices like Scorpio. And at the same time, we are seeing additional pull in for products like Aries and Taurus.
Speaker Change: and other things that we're working on.
Speaker Change: Our next question comes from the line of Mark Lippicus from Evercore ISI. Please go ahead.
Mark Lippicus: Hi, thanks for taking my question.
Mark Lippicus: A question on the diversification. If you think just longer term, you can diversify by your customer base and then by your product line. So, pick your time in the future, three years out, five years out. What do you think your...
Mark Lippicus: Split will be between, you know, the merchant GPU player versus, you know, the custom AI accelerator.
Mark Lippicus: player, like how your products will be attached to either type of solution? And then, you know, let's just say three years out, you have, you know, five product lines, is it, would you expect to have
Mark Lippicus: I still have a skew to one or would you expect to have
Mark Lippicus: like 20% in each product line bucket or something like that. If you could help us out how you're thinking about diversification like three years out, I think that'd be helpful. And then I had a follow up. Thanks.
Speaker Change: Mark, sorry, we're not going to be too much of a help.
Jitendra Mohan: What I will say, this is Jitendra by the way, what I will say is, is we like all of the, the, you know, growth vectors that you mentioned.
Mark Lippicus: We are going to grow with the third-party GPUs. We are already seeing significant growth with the internally developed ASIC platforms.
Mark Lippicus: All of our product families are contributing to revenue, as Sanjay and Mike said, and we are hard at work building new products, both new generations of the devices we already have, as well as new product families that will come to bear fruit over the next three to five years.
Mark Lippicus: So in terms of which will become more dominant or less dominant, fortunately, we don't need to predict that. We are very happy to service both opportunities.
Mark Lippicus: What I would say, as Sanjay mentioned earlier, is the hyperscalers are putting a lot of dollars into building out their own internal solutions and deploying them and customizing them to their own
Mark Lippicus: at infrastructure requirements. And because we are able to play in the backend interconnect with those customers, we do expect that to, over time, become a larger portion of our revenue.
Speaker Change: Gotcha, very helpful. And the follow up, if I may, maybe for Mike, you guys delivered a lot of upside to the outlook you originally provided. Can you help us understand, you know,
Mark Lippicus: the mechanics of how you're able to deliver that kind of upside. Is this?
Mark Lippicus: Do you necessarily have to have the inventory on your balance sheet in order to meet that? Can you, you know, get an upside to an order and then, you know, have the chip?
Mark Lippicus: you know, fab to a dye and then package assembles and test within a quarter. If you could just help us think about how you met the upside and, you know, the potential to do that again going forward. Thanks.
Mark Lippicus: Sure.
Speaker Change: Yeah, we always want to be able to supply to upside store customers needs, given that we're such a critical component in their, their system. So as you look at our balance sheet, we typically carry a pretty healthy level of weeks of inventory.
Mark Lippicus: Now, if you see in Q3 that those weeks dropped down, you know,
Mark Lippicus: Reasonably well from the previous quarter.
Mark Lippicus: that was inventory that was there to supply into the upsides.
Mark Lippicus: We are, you know, seeing that play out in Q3, we've already started to increase our inventory buys to get that days or weeks of inventory built up for Q4. But we never want to be in a position where we can't meet our customers' demands, even if they come in in the short term. We also do that so they don't feel like they need to stock up on their end. So we try to keep a very lean inventory channel at the same time.
Speaker Change: Very helpful. Thank you.
Speaker Change: Just to add to that, one thing I would say is that the way Esterra's supply chain is designed from early days was designed to address the demand profile of what you would expect from hyperscalers.
Speaker Change: with folks that have service hyperscaler customers will understand it's a it's a very
Mark Lippicus: very volatile type of thing in terms of how the upsides come in.
Mark Lippicus: So what we have done is for each one of our product made sure that there are multiple
Mark Lippicus: sources, if you will, in terms of
Mark Lippicus: OSAT and substrate vendors and all that stuff. So it's been done very thoughtfully, very purpose built in some ways for the demand profiles that we expect in these kind of markets. So we stand to gain from some of the work that's been done as part of the initial supply chain structure and how we have laid it out.
Speaker Change: Very helpful. Thank you.
Speaker Change: Our next question comes from the line of Suji Da Silva from Roth Capital. Please go ahead.
Speaker Change: Hi Jitendra, Sanjay, Mike, congrats on the strong quarter here. On the mix of revenues, I'm not really sure where Taurus is this year, but looking ahead to next year, does that, I'm curious how much that grows and helps diversify the products? Is it going to roughly a third kind of plus or minus? Is that too high just to get a sense of?
Speaker Change: How Areas and Tours Diversify Next Year.
Speaker Change: Clearly for Q4, it's going to increase as a percentage of the Q4 revenues. Beyond that, given the broadening
Speaker Change: of Aries, Scorpio, and other contributors. I don't think we can be as granular as to predict what a representative percentage of revenues, but it definitely will be one of our nice, strong road drivers for the company.
Speaker Change: Okay, Mike, that's helpful. Thanks. And then just try to understand the share situation on Ares and Gen 5. Clearly, you're very strong here. Is it the expectation that Gen 6 you guys maintain a similar share? They'll be competitive share across that? Or just what's the thoughts on Gen 6 share for Ares versus the very strong Gen 5 share?
Speaker Change: Obviously, we have a biased look on this. We have been sampling our Gen 6 devices since
Speaker Change: since February of this year. And clearly, we have, you know, learned a lot in the last three, four years by being deployed in
Speaker Change: in several different AI platforms worldwide.
Speaker Change: So we've taken the learnings from it. We've already been engaged with the market with our Gen 6 free timers.
Speaker Change: We've already learned a lot about how some of the Gen 6 based systems are being designed and deployed, including some of the customization that are required for the various accelerators.
Speaker Change: So to that standpoint, we do believe that, you know, we have a strong start and we have backlog and everything to support that. We've been shipping that for some time now from a pre-production.
Mark Lippicus: standpoint. So overall, we do believe that we will continue to make grounds. But having said that, this is a big market, and we do expect competition to come in. There is no surprises on that front.
Speaker Change: It's very helpful, Jitendra. Thanks, guys.
Speaker Change: Our final question comes from the line of Richard Shannon from Craig Hallam. Please go ahead.
Speaker Change: Transcription by ESO. Translation –
Speaker Change: Well, thanks guys for taking my questions.
Richard Shannon: Maybe one for, I guess, for either Jitendra or Sanjay regarding Cosmos.
Richard Shannon: Maybe we can talk about...
Richard Shannon: Any competitive response you're seeing from others who are who are getting in this market either from the retimer switch side or both And it seems like their press releases have alluded to capabilities in this direction Obviously your first to market move is is it seems to have a lot of stickiness process. I'm sure you're not Standing still and moving forward here I think you can talk about what you're seeing there from a competitive response and how you're how you're trying to maintain that That barrier to switching place
Speaker Change: It's a great question. So, you know, we have not, we've been talking about our Cosmos software fairly openly.
Speaker Change: for about a year now, and so it's kind of fair to assume that other people will if they want to try to copy something similar.
Richard Shannon: But what you have to realize is Cosmos software is not just a collection of, you know, APIs and so on. It absolutely is that. There is also a lot of knowledge that has gone into the software by us being in the trenches over the last three to four years.
Speaker Change: understanding what works in the ecosystem, what issues some of the the ASICs and other products have, and how we work around them. So all of the learnings have gone into this Cosmos software to make it very rich. And as Sanjay mentioned, we've been sampling our
Richard Shannon: Gen6 product for over six months now. And all of the learnings that we've seen firsthand by being shoulder to shoulder with our customers, they have again gone in to make our Cosmos system software very rich and reliable.
Richard Shannon: So anybody who wants to come in will have to have that level of kind of experience and soak time to make their software as good.
Richard Shannon: Now, the other interesting thing is with the launch of the Scorpio family.
Richard Shannon: We get even more access to diagnostics and what's going on with the network congestion and things like that.
Richard Shannon: that we can now very uniquely enable for our customers, if somebody without having all of these components in play would find it difficult to do. So we believe Cosmos is a very differentiated software for us and we continue to make it better and better over time.
Speaker Change: And just to add to what Jitendra said, again, Cosmos, you've got to look at it in two ways. There is a part that runs in the operating stacks of the hyperscalers.
Richard Shannon: but there is also the part that runs within our check.
Richard Shannon: The point I'm trying to make is that our chips have been defined or developed in a software-centric approach.
Richard Shannon: So, to that standpoint, it's not just about the diagnostics, it's about how the fundamental architecture of a check has been done.
Richard Shannon: in order to make it.
Richard Shannon: more of the eyes and ears from a diagnostic and feed management standpoint, but also in terms of customization and various different things that hyperscalers care about in terms of making it uniquely fit into their infrastructure requirements.
Speaker Change: Okay, wonderful. Thanks for all that detail. My follow-on question, while I'd love to ask yet another question on the Scorpio line, which is obviously a very interesting product announcement, maybe I'll ask one on the Leo product line in that general area here. I guess maybe, you know, kind of looking back from, say, the beginning of the year or a year ago, when hyperscalers were looking at CXL and trying to examine the use cases that made sense,
Richard Shannon: I want to get your sense of whether you think that they're developing and getting certainty on them and the pace and the timeframe that you originally thought. You obviously talked about a design, I think, ramping by, you know, mid next year. Is that kind of the timeframes you expected back then or maybe just kind of talk about
Richard Shannon: the hyperscalers experiences and figuring out those use cases.
Speaker Change: The combination of both, right, in the sense that CXL is a new technology. There needs to be a clear use case and ROI.
Richard Shannon: And that, we believe at this point, is starting to get established based on the work that, you know, we have done and others have done from a CPU memory vendor standpoint.
Richard Shannon: So that's becoming clear. But I think probably a year, two years ago, maybe there was a little bit of overhype on things like pooling and others that that got made out. But I think those have been clarified now. So the use cases are much more real.
Richard Shannon: And to that standpoint, it's really a matter of dollars flowing back into the general compute area, the CPU launches, which, of course, both Intel and AMD did recently with support for CXL.
Richard Shannon: So I think it's all starting to add up, but I will add the caution like I noted before. I think we're still in that, you know, crawl-to-walk stage. There is some more work that needs to be done before we are up and running in full speed with CXL.
Speaker Change: Wonderful. Thanks for that detailing and congratulations on the great numbers, guys. Thank you.
Speaker Change: Thank you. Thank you.
Speaker Change: All right, there are no further questions at this time. I turn the call back over to Leslie Green for closing remarks.
Leslie Green: Thank you everyone for your participation and questions, and we look forward to updating you on our progress.
Speaker Change: This concludes today's conference call. You may now disconnect.