The Zhitong Finance App learned that semiconductor company Astera Labs (ALAB.US) announced the results for the third quarter of the 2026 fiscal year and held a conference call after the US stock market on November 3 (Tuesday). The quarter's revenue of US$584.9 million was a record, up 154% year on year and 49% month on month, higher than the market's agreed expectations of about US$550 million, and also exceeded the company's previous guidance limit of US$540 million to US$560 million; non-GAAP earnings per share were US$1.32, higher than expected; GAAP earnings per share were US$0.99. The non-GAAP gross profit margin was 72.4%, higher than the 72% guideline; the non-GAAP operating margin was 44.7%, up 560 basis points from month to month.
Strong guidance for the fourth quarter: the company expects revenue of US$730 million to US$750 million, with a median increase of about 26% month-on-month, higher than the market's general expectation of US$655.7 million; the company expects non-GAAP earnings of US$1.62 to 1.66 per share, far higher than analysts' forecast of US$1.43; the company also expects GAAP earnings of US$1.13 to 1.17 per share, with a gross margin of about 72%.
There has been a landmark change in the product structure: The Scorpio smart switch has jumped to become the company's largest product line in terms of revenue this season, mainly driven by the Scorpio X-Series 320-lane switch entering mass production. At the same time, the company also released the Aries 7 PCIe 7 signal conditioning product portfolio (samples are expected to be delivered to designated customers in the fourth quarter), Leo 2 and Leo X-Series memory expansion products, and Orion reference architectures for rack-level AI inference. CEO Jitendra Mohan said that the acceleration of design activities and new design orders have made the company expect to increase the value of chips corresponding to individual accelerators.
The core focus of market attention:
After Scorpio X climbs steeply, how will it continue to grow in 2027
The most focused topic of the question and answer session at this conference was the sustainability of growth. Joe Moore of the Morgan Stanley Research Department directly asked: Scorpio X's annualized operating rate of revenue has exceeded 1 billion US dollars, and the climb is so steep, whether there will be a platform period before switching to ULink. Tripp Smith of the Barclays Research Department quoted market expectations to further ask: The deployment of this generation of major customers may peak in the first quarter of 2027, and the next generation platform will not be launched until close to 2028. How do you view the revenue pace for the 2027 calendar year.
The management did not comment on the volume pace of end customers, but gave a three-tier response: first, the initial Scorpio X platform has completed most of the initial deployment and will continue to contribute revenue on a large scale; second, the X Series has been shipped to the second hyperscale cloud customer in the third quarter (in pre-production stage), and a backlog of orders from several new customers will be shipped in the fourth quarter. It is expected that the introduction and certification will be completed in the first half of 2027 and contributed revenue in the second half of the year; third, imported customers are expanding Scorpio X to more new platforms. CEO Gitendra Mohan further gave a long-term story: Currently, the content value corresponding to a single XPU has exceeded 1,000 US dollars, and the company aims to double the content per XPU chip with each generation of AI platforms.
From an external perspective, the market's disagreement over Astera did not grow during the quarter, but rather about valuation and sustainability. According to 24/7 Wall St. data, 69% of the 26 analysts covered gave a buying rating, with a consistent target price of $387.57, which is significantly higher than the current stock price; AllMind previously suggested in the September analysis that the core risk is customer concentration (the top five customers account for 84% of revenue in 2025). Once the pace of main customer deployment changes, it may cause an “empty window” in revenue, and the impact of gross margin of 70% on the valuation multiplier will outweigh the increase in revenue. The stock price did not rise but fell after the market, reflecting to a certain extent the market's cautious attitude about the visibility of 2027 growth under a high base.
Product structure: Scorpio jumped to the largest product line, with annualized revenue exceeding 1 billion US dollars
In the third quarter, the Scorpio intelligent switching matrix (fabric switch) family became the company's largest product line in terms of revenue, and the X series's revenue surpassed that of the P series, and the entire Scorpio family's annualized revenue operating rate exceeded 1 billion US dollars. The X series was shipped to two customers and across multiple platforms in the third quarter; the P series achieved double-digit month-on-month growth, and customer coverage expanded to two US hyperscale cloud vendors. In terms of signal conditioning, Aries was driven by the adoption of PCIe Gen 6 in vertical expansion (scale-up) and horizontal expansion (scale-out) scenarios; the Taurus 800G solution began to be shipped in batches to major customers in the third quarter. Chief Financial Officer Desmond Lynch expects the P series to maintain a similar growth rate in the fourth quarter, but the X series will be the dominant growth engine next year due to its higher average selling price (ASP).
The “doubling of each generation” chip content narrative
Mohan clearly quantified the long-term growth logic: Currently, in high-configuration systems, the content value of a single XPU with switching matrices plus copper connection signal conditioning has exceeded 1,000 US dollars; the next generation of AI fabric products alone is expected to contribute more than 1,000 US dollars to double the unit content for the first time; in the future, the higher mount rate (attach rate) brought about by superimposed optical components, Leo memory controllers, and the ASP increase brought about by each generation of speed upgrades. The company aims to double the unit chip content of each generation of XPU. Sanjay Gajindra added that with every doubling of the data rate, ASP usually increases by 20% to 30%.
Timeline of the four new growth engines
Management outlined four incremental curves outside of the core business: First, optical connectivity, smart transimpedance amplifiers (TIA) and laser drivers for near-package optics (NPO) began to be delivered to designated customers this quarter. It is expected to start contributing revenue in the second half of 2027 after a six-month certification period. A single accelerator corresponds to hundreds of dollars, and co-packaged optics (CPO) emissions are expected to focus on 2028 to 2029; the company estimates that the scale-up optical interconnect market will reach 40 billion US dollars in 2030. Second, custom chips. Currently, we have over 10 custom projects (custom fabrics, custom memory controllers, IO chips). Shipment is expected to begin at the end of 2027 and release in 2028. The third is the memory solution. Pre-production and shipment of Leo 2 has begun, and mass climbing will begin in mid-2027. Fourth, Orion's rack-level reference architecture. A single 80 GPU rack supports over $100,000 worth of Astera content (including three monetization points for the board, software, and reference design license fees), and will enter the market through ODM/OEM partners in 2027.
Fourth quarter guidance and operating leverage
Fourth quarter guidance (non-GAAP): revenue of $730 million to $750 million; gross margin of approximately 72%; operating expenses of $194 million to $200 million; operating margin of approximately 45%; number of diluted shares of approximately 187 million shares; and earnings per share of $1.62 to $1.66. GAAP earnings per share guidance of $1.13 to $1.17. Non-GAAP operating expenses for the third quarter were US$162 million, up 19% month-on-month, but Lynch emphasized that the cost growth rate was lower than the revenue growth rate. The long-term operating expenses rate target was about 30% of revenue, which is currently significantly lower than this level; the number of employees doubled year on year. Operating cash flow for the third quarter reached a record of US$179.2 million, with cash and marketable securities at the end of the period of US$1.38 billion, an increase of US$125.3 million over the previous quarter.
Full Meeting Minutes
CEO and Co-Founder Gitendra Mohan: Thank you Leslie, good afternoon everyone, and thank you for attending our third quarter 2026 conference call. Today, I will introduce the third quarter results, market trends, and progress of the AI fabric and signal conditioning product portfolio, then Sanjay will discuss our long-term growth profile, and Des will introduce the third quarter financial results and fourth quarter guidance.
Astera Labs delivered an outstanding report card for the third quarter: record revenue of US$584.9 million, up 49% month-on-month and 154% year-on-year. Operational execution was strong during the quarter, and we broadened our customer base through new design wins (design wins) and expanded our intelligent connectivity platform with various new product launches. Revenue growth in the third quarter was widespread, led by a massive climb in the Scorpio X series aimed at scale-up (scale-up) networks. Scorpio became our largest product family in the third quarter, and the Scorpio X series has already surpassed the P series in revenue. In total, Scorpio represents more than $1 billion in annualized revenue. This is an important milestone in our journey to become a supplier of key solutions for rack-scale AI infrastructure.
Since our last performance call, we've substantially expanded our product portfolio with a record number of new product releases, including: Aries 7, the industry's first PCIe Gen 7 copper and optical connectivity solution; Leo 2 and Leo X series intelligent memory controllers for AI and general use scenarios; and Orion, our new compute-agnostic (compute-agnostic) rack-scale architecture for enterprise AI and inference applications. Together, these innovations broaden the market and application of our services, create greater value for hyperscale cloud vendors and enterprise customers, and support a diversified long-term revenue growth profile. In the third quarter, our design introduction activities for both new and existing customers heated up significantly.
Scorpio is increasingly establishing its position as a diversified AI fabric platform, with kinetic energy spanning the X Series and P Series. In the third quarter, the X Series was shipped to two customers across multiple platforms, and we expect to deliver pre-production batches to more new customers this quarter. The P series also expanded its footprint on multiple platforms, covering two US hyperscale cloud vendors in the third quarter. In addition to AI fabrics, we are also seeing more diversified paths: the Taurus 800G solution has been certified and shipped in batches to major customers, and samples are currently being delivered to many other customers; Leo's momentum continues, and multiple customer certifications, including two hyperscale cloud vendors, have been completed, and we expect mass shipments to begin in mid-2027. Our custom solutions business design activities have also increased dramatically, and our opportunities have continued to expand throughout the year, and we are currently working on more than 10 active projects. Finally, the platform business anchored by Orion's reference architecture is receiving strong early customer interest, and the first shipments are expected to begin in 2027.
Demand for AI infrastructure continues to strengthen, and our hyperscale cloud customers have reported strong revenue growth from AI services. However, physical input factors, including wafer production capacity, grid power supply, and memory supply, remain highly tight. This combination makes utilization particularly critical. At the same time, the load structure is shifting to inference, and the continuous inference cycle is multiplying the traffic through the fabric. Operators are now optimizing two things: latency to reduce time to first token (time to first token); and accelerator utilization to increase revenue per token. Our Scorpio AI fabric is tailored for these metrics.
As token economics become increasingly important to customers, let me highlight Astera's differentiating ability to drive success in the AI fabric field. First, port radix (radix). Our Scorpio X device portfolio, including the industry's largest 320-channel (lane) configuration, enables PCIe-like scale-up AI fabrics to support large-scale clusters. Scorpio X's high channel count helps maximize the size of the scale-up domain and supports an all-to-all (all-to-all) flat topology, allowing AI accelerators to communicate with each other with full bandwidth and low latency. This is particularly important for AI inference, as fewer hops reduce collective communication delays and allow resources to be used for computation rather than moving data between switches. Increased utilization results in a higher number of tokens per second, ultimately making the same accelerator investment more profitable.
Second, intelligence. Traditional switches only carry data packets. The Scorpio X series has AI load sensing capabilities and participates in collective communication and computation. Hardware-accelerated Hypercast and in-network computing (in-network computing) offload AI operations such as all-reduce and broadcast directly from the accelerator to our fabric, improving aggregate performance by up to 2 times. Similarly, the Leo memory controller attached to Fabric is used for KV cache (KV cache) offload, which significantly improves the initial token time and number of tokens per second. With advanced AI perception capabilities brought by Scorpio and Leo products, customers can use the same number of accelerators to improve token economy and increase token output.
Third, software. The COSMOS software layer transforms our AI fabric and intelligent connectivity solutions into infrastructure that customers can efficiently deploy, manage, and operate on a large scale. COSMOS covers our entire product portfolio and is widely used in the AI ecosystem, with functions extending from device-level diagnostics to fabric and rack management. COSMOS accelerates rack deployment by resolving connectivity issues in software, avoiding costly hardware revisions, and keeps the rack online with full fabric telemetry and predictive fault detection. It also configures and manages HyperCast and in-network computing, so in addition to managing the connectivity infrastructure, our software is directly involved in the AI load running on the XPU. Faster revenue conversion times and higher AI infrastructure utilization rates are what drove COSMOS to win bids for new designs across our entire product portfolio.
These differentiating capabilities are optimizing performance, improving XPU utilization, rapidly expanding our opportunity funnel and speeding up design bid transformation. Therefore, we expect to achieve strong diversified growth in the AI fabric portfolio by continuously increasing value to customers and expanding the dollar content of individual AI accelerators. Today, our single XPU content on fabric switches and copper signal conditioning can exceed $1,000, and the next generation of AI fabric products alone is expected to contribute far more than $1,000 per unit of XPU. Over time, our goal is to cover the entire scale-up exchange TAM, from today's PCI Express and NVLink Fusion based platforms to systems based on UALink and other protocols. Coupled with memory controllers and optical connectivity products, higher AI fabric installation rates driven by next-generation scale-up protocols including UALink, larger cluster sizes, and higher ASP brought about by each data rate transition, we are seeing a path: with each new AI platform, our dollar content per XPU chip doubles.
In summary, the third quarter was another important point in the evolution of Astera Labs. The record performance was led by the Scorpio X series's high-volume climbs, and the signal conditioning portfolio continued to grow. We saw strong design winning momentum in both old and new customers, and released a record number of products during the quarter. Looking ahead, we see significant opportunities in optical connectivity, custom chips, memory solutions, and rack-level infrastructure. They build on our core business, expand our accessible markets, and lay the foundation for long-term growth. Now it's up to our President and Chief Operating Officer Sanjay Gajindra.
President, Chief Operating Officer and Co-Founder Sanjay Gajindra: Thank you Gitendra and good afternoon everyone. This quarter was a major inflection point for Astera Labs: we have evolved from a device-level vendor to a rack-level infrastructure platform vendor. We are now integrating chips, hardware, and software to solve AI rack and cluster (pod) wide connectivity challenges through copper-optical interconnections to improve XPU utilization and return on investment (ROI). As the smart connectivity portfolio expands, we are increasing the dollar content of the chip corresponding to a single accelerator to the level of several thousand dollars, while expanding accessible market opportunities.
Our roots today are comprised of two core revenue generating businesses: AI fabric and signal conditioning; and we are rapidly expanding along four incremental market paths: optical connectivity, custom chips, memory solutions, and rack-scale infrastructure platforms. As AI systems expand in scale, the challenge is no longer simply connecting more accelerators within the server. As Gitendra said, customers are building larger scale-up domains and need more bandwidth, lower latency, and smarter data handling methods to improve token economics. Our hands-on experience in deploying scale-up networks on a production-level scale is helping us define a set of high-value features for specific workloads. Based on these factors and increasing load rates, we believe we can more than double the dollar content per unit of XPU with each subsequent generation.
As data rates double and AI clusters expand, signal conditioning remains a key problem that must be solved. Therefore, we expect that the new PCIe Gen 7 combination (covering optical and copper) will surpass the ASP and payload rate of the previous generation. Furthermore, our Aries gearbox (gearbox) and heavy drive (redriver) solutions are expected to open up new socket (socket) opportunities in AI platforms.
Taurus for 800G applications began mass shipments in the third quarter, and we expect steady growth in the scale-out Ethernet active cable (AEC) sector in 2027 as higher ASP and new projects climb. We are also sending samples of 3.2T retimers (retimers) to a number of customers.
Looking ahead, we are adding four growth drivers to substantially expand accessible markets. First is optics: we continue to make good progress in the fiber connectivity business. At the OCP (Open Computing Project Summit) last month, we unveiled our first optical products, including intelligent optical drivers and TIA optimized for scale-up PCIe Gen 7 topologies for near-package optics (NPO) and linear pluggable optics (LPO) applications; we also demonstrated the integration of optical technology with Scorpio X series fabric switches. Our smart TIA and laser drivers will begin sending samples to designated hyperscale cloud vendors and OEM partners this quarter for NPO applications; we are actively working with a number of customers to expand the PCIe fabric from within the rack to the pod level. In terms of near-package optics for scale-up applications, we expect TIA and drivers to unlock additional content opportunities for individual XPUs, and expect these optical products to start contributing initial revenue by the end of 2027. In the longer term, we are developing a fully integrated optical engine (optical engine), with the goal of creating optically-enabled Scorpio X series fabric switches. We estimate that the 2030 scale-up optical interconnect opportunity is a $40 billion market, and we will serve this market through high-density fiber mount solutions and optical engines for NPO and CPO applications.
The second is customized solutions. The AI accelerator landscape continues to expand and diversify, driving the demand for dedicated intelligent connections. Interest and activity in our bespoke business accelerated markedly in 2026, and we currently have more than 10 opportunities at our disposal. We are developing custom fabric solutions, custom memory controllers for dedicated inference applications, and IO chips (chiplets) for next-generation accelerator platforms. We expect the first batch of these plans to be shipped at the end of 2027, and the volume will rise in 2028. We made trade-offs and focused on high-value custom projects to leverage our deep connectivity IP portfolio, robust supply chain partnerships, and system-level expertise.
The third is the memory solution business. We recently expanded our Leo smart memory controller family to cover new applications and use cases. The Leo X series is our new fabric-mounted memory controller, which works with the Scorpio X series AI fabric to create a dedicated memory layer for the KV cache and agent context (agent context). We also released the Leo 2 E series and P series, which double the memory bandwidth and capacity compared to the previous generation, while supporting DDR4 and DDR5 memory sticks. As memory constraints increase, Leo allows customers to expand memory capacity, pool standard memory within a rack, and reliably mix and reuse deployed DDR4 with the new DDR5. Customer interaction and design activity remains strong among the growing number of hyperscale cloud vendors and inference cloud customers. I'm happy to report that pre-production shipments of Leo 2 products have now begun, and mass climbing is expected in mid-2027.
The fourth is the platform plan. We recently released Orion, a computationally neutral, rack-level reference architecture designed to provide a turnkey AI infrastructure solution for inference cloud and enterprise customers. Orion covers six production-level reference designs, can be expanded from 8 to 160 accelerators in a single cluster, and integrates the COSMOS software management platform, which has now been enhanced to include rack and fabric management capabilities. These reference designs use our portfolio of intelligent connectivity solutions, including AI fabrics, intelligent signal conditioning, advanced memory solutions, and COSMOS to provide a tightly coupled, high-performance platform. We're already seeing strong interest from GPU and XPU vendors, and Orion is expected to enter the market through our ODM and OEM partners in 2027. The fully equipped Orion rack-class design includes over $100,000 worth of Astera content, creating another major incremental market opportunity.
In summary, we remain focused on execution to deliver substantial value to AI and cloud infrastructure customers at the chip, hardware, and software levels. Design activities are being accelerated, winning bids on a wide range of projects and applications from multiple customers. Record new product launches are broadening portfolios, deepening collaboration between solutions, and expanding market opportunities and the chip dollar content of individual AI platforms. Let's leave it to Des.
CFO Desmond Lynch: Thanks Sanjay and good afternoon everyone. Today I'm going to review financial results for the third quarter and then discuss the fourth quarter guidance, both presented in a non-GAAP manner.
We had excellent results in the third quarter, with record revenue of US$584.9 million, up 49% month-on-month and 154% year-on-year. The growth is widespread in AI fabric and signal conditioning combinations, and is being led by a steep rise in mass production of the Scorpio X series. Revenue for the quarter was further diversified based on a wider range of customers and platforms. Scorpio became the largest product family by revenue in the third quarter, with X series revenue surpassing P series at the same time. Revenue of the X series increased significantly, and shipments expanded to multiple platforms and port base configurations; the P series also achieved month-on-month growth, expanding existing deployments and entering mass production of new projects.
In terms of signal conditioning, ARIES' growth was led by strong adoption of PCIe Gen 6 in scale-up and scale-out applications; Taurus also increased kinetic energy, and 400g and 800g solutions climbed a slope in scale-out applications. Non-GAAP gross margin for the third quarter was 72.4%, higher than the 72% guideline. Non-GAAP operating expenses for the third quarter were US$162 million, up 19% month-on-month, reflecting continued investment in R&D to expand accessible markets in core businesses and emerging growth areas. The non-GAAP operating margin for the third quarter was 44.7%, up from 39.1% in the previous quarter, and increased by 560 basis points, reflecting significant operating leverage brought about by the volume of revenue. Interest and other income of $13 million, non-GAAP tax rate of 11.4%.
Non-GAAP net profit of US$243.4 million corresponds to non-GAAP diluted earnings of US$1.32 per share, an increase of US$0.52 over the second quarter and an increase of 65% over the previous quarter. Cash, cash equivalents and marketable securities at the end of the period were US$1.38 billion, an increase of US$125.3 million over the second quarter, mainly driven by a record operating cash flow of US$179.2 million.
Now let's look at the fourth quarter outlook. We expect revenue to be between $730 million and $750 million, with a median increase of 27% month-on-month. Growth is expected to be driven by the continued expansion of the Scorpio X Series and P Series in both customer and platform dimensions. The growth in signal conditioning is expected to be led by the adoption of PCIe Gen 6 retimers in Aries and the climbing of the Taurus 800G solution. We expect non-GAAP gross margin of about 72% for the fourth quarter, and non-GAAP operating expenses between $194 million and $200 million, reflecting continued investment in core AI fabrics and signal conditioning portfolios, as well as additional investment in the growth opportunities described by Sanjie. We expect a non-GAAP operating margin of approximately 45%, interest and other income of approximately $13 million, and a non-GAAP tax rate of approximately 12%. The number of diluted shares is expected to be 187 million in the fourth quarter. Overall, we expect fully diluted non-GAAP earnings per share between $1.62 and $1.66.
Finally, the third quarter was an extraordinary quarter: record revenue, significant operating leverage, and a strategic milestone for Scorpio to become the largest product family by revenue. Looking ahead, our growth drivers continue to expand: the Scorpio X series and P series are being deployed to new and existing customers and platforms; in terms of signal conditioning, Aries and Taurus benefit from the adoption of next-generation connectivity standards; opportunities for optical, customization, memory, and platform solutions are emerging outside of our core business. These are all reflected in our strategic R&D investment to expand our portfolio and cover the complete market. We will continue to invest in our entire portfolio, execute with discipline, and maintain strong profits to create long-term value for our shareholders. This concludes the formal statement. Let's move on to the question and answer session.
Q & A session
Harlan Suhl (J.P. Morgan Chase Research): Good afternoon, thanks for taking the questions. The team's execution is excellent as always. Regarding the new Orion rack-level reference architecture: Large organizations like ours are beginning to evaluate locally deploying open source, open weighted models to obtain lower token inference costs, control model fine-tuning, and more importantly ensure internal data security. These customers are looking for computing platforms that can be integrated into their private data center architectures: air cooling, compact cluster size, and flexible support for XPUs and CPUs. Orion looks like that's the plan. What is the current market interest? Have you worked with large companies? How much dollar content can Astera capture in the different configurations you offer?
Sanjay Gajindra: Great question, Haran, thank you. We're very excited about Orion. As you correctly pointed out, enterprise AI is a major market. It is highly fragmented and has diverse workloads. They all need to obtain computing power in a ready-to-use form, and more importantly, target specific workloads. Our idea behind developing Orion is a GPU-neutral (GPU-agnostic) rack: you can bring any GPU that fits the load or is selected based on availability, the rest of the infrastructure is all defined, the connections are all defined as part of the reference design, and the same goes for power supply, cooling, and other necessary components.
Astera provides three main things: first, a complete reference architecture, whether at the node (box) level or rack level; second, using the board of our chip to combine fabric solutions, memory solutions, signal conditioning, and other necessary components in a fully optimized, fully tuned, and fully verified manner; and third, software, which makes us very excited because CosMOS has upgraded from device-level software to the rack level and fabric level to manage all fabric, rack-level diagnostics, and orchestrate other aspects of GPU performance. Overall, Orion responds positively to the kind of use cases you've described. The Orion family will have multiple generations, and the content will obviously get higher and higher. But the original platform we call Orion 1.0 was based on PCIe scaling, and we estimate that the Astera content of around $100,000 will be monetized through components, boards, and software.
Joe Moore (Morgan Stanley Research): Thank you and congratulations. Can you talk about Scorpio X? You've crossed the $1 billion threshold and climbed such a steep slope. Do you expect that after Scorpio X's product matures and before entering UALink, there will be some kind of platform period? Or help us understand what this process of evolution probably looks like? It sounds like you're expecting PCI Express to have a long end while also transitioning to other protocols. Can you talk about this dynamic?
Desmond Lynch: Hi Joe, I'm Des. We were very satisfied with the execution throughout 2026, and the product portfolio showed a broad range of strengths. Looking at the growth in the second half of the year, it was mainly led by scale-up deployment and mass production climbing of the Scorpio X series. The climbing began in the third quarter and will continue into the fourth quarter. This enabled us to achieve close to 50% month-on-month revenue growth in the third quarter. Based on the median guidance for the fourth quarter, we expect a further 27% increase over the previous quarter.
Looking further into the future, the initial Scorpio X platform has completed most of its initial deployment and climbing, and will continue to contribute on a large scale in the future. We expect Scorpio X to continue to grow, driven by new platforms and more customers that will continue to add to the model throughout 2027. Looking further, with our experience in large-scale deployments, advanced feature sets, and strong relationships with our customers, we are well prepared for the next generation of platforms to grow. Of course, in addition to Scorpio, we are also excited about the growth opportunities described above: new product fields such as Aries 7, Leo, and optical products will continue to grow, all of which will be superimposed into the model starting in 2027, providing future growth opportunities. Overall, we are very satisfied with the performance and momentum of the overall combination.
Brian Curtis (Jefferies Research): Thanks for taking the questions, great results. I'd like to ask about NPO (Near Package Optics). You mentioned that there are some design developments over there, and new devices on the analog side have also been released on OCP. Can you take us to sort out how these different components will be gradually superimposed? Are you currently designing a complete NPO plan, or is it going to be a little later?
Sanjay Gajindra: It's happening right now, Brian. As we mentioned in our statement, we plan to begin sending samples of NPO products to leading customers this quarter. We expect around six months of certification and system-level integration. Therefore, we expect the NPO deployment to start contributing revenue in the second half of 2027. These are still the components we supply internally for the NPO module, i.e. TIA, laser driver, etc. We expect that a single accelerator will bring in hundreds of dollars through NPO products and solutions under construction. Looking back, we expect NPOs to continue to drive growth after 2027, and the volume of CPO volumes will probably be mainly focused on the 2028 and 2029 time frames.
Torre Swanberg (Stifel Research): Thank you and congratulations on another record quarter. You now have four new growth engines, and I agree these are all huge TAMs. You will also raise operating expenses (OpEx) significantly during the December quarter. What I'm curious about is, are most of the operating expenses corresponding to these four new business opportunities still behind, or have you already invested heavily in these four major opportunities?
Desmond Lynch: Hello Torre, I'm Des. We will continue to invest in the business at the right level to support the high growth opportunities that lie ahead. You're right, operating expenses did rise to about US$162 million in the third quarter, and we expect continued growth in the fourth quarter. However, I would like to point out that our operating expenses are growing at a slower rate than revenue growth, so the model shows significant leverage. The increase in operating expenses in the second half of the year was mainly driven by three factors: first, we continued to expand our personnel to support customer projects, and the number of employees doubled year on year; second, variable R&D expenses in the form of wafers, masks, and IP increased; third, as performance continued to exceed the profit and loss model, variable compensation increased accordingly. Once again, the model shows significant leverage. I think in the long run, a reasonable reference level for operating expenses is around 30% of revenue, and given the current high revenue growth, we are clearly below this level.
Ananda Barua (Loop Capital Research Department): Good afternoon, congratulations on the progress made, and thanks for taking the questions. One clarifying question: You've mentioned many times that each new platform generation doubles the amount of XPU per unit. Does “each generation” here refer to each new generation of XPU, or is each generation of Astera's own products (such as Gen 7 and Gen 8)? or something else? Would love to know how to understand it.
Gitendra Mohan: Good question, I'm Gitendra. Looking at our content in some high-configuration systems today, fabric switches plus signal conditioning products, the unit XPU content has exceeded 1,000 US dollars. As we enter the next generation of products, we aim to increase the XPU content per unit by more than $1,000 just by exchanging (switching) opportunities, which constitutes the first time that is currently being said to double. Further, as we enter next-generation AI platforms or next-generation XPUs, we will bring more content to these platforms, such as optical network components that Sanjie just mentioned, Leo memory controllers for KV cache offload, etc. In addition, there are single-product ASP improvements brought about by larger clusters, and single-product ASP improvements brought about by higher speed. After modelling all of these factors, we expect to double the amount of XPU per unit with each successive generation of XPU.
Ananda Barua: That's quite a bit of content. Thank you.
Gitendra Mohan: Yes, our customers are constantly building bigger and bigger clusters. The bigger the cluster, the higher our load rate. Thank you.
Papa Sila (Citi Research): Thanks for taking the questions and congratulations on the strong third quarter results. I'd like to ask Gitendra or Sanjay to talk more about the more than 10 Scorpio X collaboration projects you mentioned earlier. How many of these are expected to generate significant volumes in the first half of 2027? Related to this, have your views on ULink timelines changed? Still think it's more of a story from the second half of 2027? If so, is it reasonable to assume that some of your current Scorpio X collaboration projects will turn into UALink opportunities in the future?
Sanjay Gajindra: Thank you for asking. We previously disclosed that more than 10 customers are working with us on Scorpio X, and these partnerships are currently progressing in a very positive way. We mentioned that the main hyperscale cloud platform is now in mass production, which is reflected in the third quarter figures and our fourth quarter guidance. We also mentioned that in the third quarter, we began shipping X-series Scorpios to another hyperscale cloud vendor, which is currently in pre-production and should help us generate revenue in 2027. We also had a backlog of Scorpio X orders from many new customers in the fourth quarter. These products will be shipped this quarter, then complete the development and certification phase in the first half of 2027, and contribute revenue in the second half of the year.
What's also happening is that with the results and performance Scorpio X delivered to customers, as well as the many unique features we emphasized in our statement, customers who have already introduced it are adding new platforms, that is, using Scorpio X on the new platform. So we're not only seeing an increase in the number of customers using Scorpio X, but we're also seeing customers rapidly expanding their use of Scorpio X on the multiple platforms they've developed. Overall, Scorpio X took off very smoothly, which is reflected both in the project pipeline and in revenue: the Scorpio family's annual operating rate in the third quarter has exceeded 1 billion US dollars.
Gitendra Mohan: UALink's questions are answered by me. We continue to see the UALink ecosystem grow: more IPs are available, and more vendors are launching UALink products. Overall, UALink hasn't changed much from a customer perspective. Customers who focus on memory semantic (memory semantic) protocols such as PCI Express will stay on their tracks and will, as you say, switch to UALink over time. Our UALink product development also continues: we released the Taurus family earlier this year, and continue to work on key anchor sockets (anchor sockets) such as swaps to enable docking when the customer's UALink XPU is released. Not much has changed: the one on the UALink circuit stays on the UALink track, and the ethernet track stays on the ethernet track. Overall, our strategy is to cover the entire scale-up exchange TAM: starting with PCI Express, expanding to NVLink Fusion, and actively advancing UALink. As customer interest increases, we will also cover other scale-up agreements.
Sebastian Nagy (William Blair Research): Thank you and good afternoon. I'd like to continue the topic of Scorpio, but ask about the P series. The Scorpio X series surpassed the P series in revenue this quarter. Can you talk about how we view the growth of the P Series in the fourth quarter and 2027? It seems that the X series has a higher content but is more concentrated in customers, while the P series has a lower content but a wider range of customers. How should the structure between these two products be viewed in the future?
Desmond Lynch: Hi Sebastian, I'm Des. We are very pleased with the performance of the Scorpio P series. It achieved double-digit growth in the third quarter, and I expect it to continue at a similar rate in the fourth quarter and beyond. Looking ahead to next year, we continue to see hyperscale cloud vendors continue to grow: we've talked about two hyperscale cloud vendors contributing revenue growth in the second half of this year, which will continue until next year; at the same time, multiple projects for different customers continue to climb. So the P series will continue to be a meaningful part of Scorpio's overall climb. Obviously, given the higher ASP, the Scorpio X series will be the dominant growth driver next year, and the P series is still performing very well.
QUINN BOLTON (Needham Research): Congratulations on this guide. I'd like to ask a longer-term question. You presented four new categories at the conference: optical, custom, memory controller extensions, and Orion rack-scale platforms. I'm not asking for promises from you, but are any of these four more likely than the others to reach 10% of revenue within the next few years? Or how should we view the relative growth rate of these four new opportunities?
Sanjay Gajindra: First of all, I want to say that the four are huge TAMs: not to mention optical and memory inference use cases; customization is a huge opportunity in terms of XPU mounting and other XPU opportunities; the rack is aimed at multiple enterprises and inference applications, and the value proposition and form are completely different. These four are big markets for us, and the next depends on whether we can develop the right product and ensure execution and delivery. I think it's a matter of time.
In other words, the most recent opportunity in time is the optical climb we expect in the second half of the year, particularly on the NPO side, which will be the most immediate contributor to revenue. Custom chips are what we have been refining, and great progress has been made. Currently, we are working on more than 10 projects, covering custom fabrics, custom memory controllers, etc. I think this is a huge opportunity for us, but it will probably be in 2028. Finally, Orion. Frankly speaking, it has had an excellent response since it was released in OCP. Many people have come to the door, especially corporate customers and XPU and other accelerator manufacturers. In this area, we will start contributing revenue in 2027 by launching products for our ODM and OEM partners who build systems. We haven't fully figured out what the Orion platform numbers will look like, and we expect to provide more specific data during the next quarterly conference call.
But overall, the key message is: these are substantial big market opportunities that require innovation and execution, and we as a company have proven our ability to deliver both. We look forward to participating in all of these markets and contributing more growth engines beyond the existing main business of signal conditioning and fabric. If all six engines can output strong growth, I think we have achieved our goal of becoming a great iconic company.
Sam Feldman (BNP Paribas Research Division, Daikal Ackerman): Hello everyone, I'm Sam Feldman. Karl Ackerman (Karl Ackerman) asked. You Aries are growing strongly. Can you split this growth into two parts: the increase in content brought about by the switch to PCIe Gen 6 and the increase in shipment volume? More generally, as optics move closer to the computational side, how do you view the retimer content per unit XPU?
Sanjay Gajindra: Very good question. Aries set a record season and the family continues to grow. Our company actually started with Aries PCI Express Gen 4, and is now moving towards Gen 5 and Gen 6, and released Gen 7 packages at OCP. As you correctly pointed out, ASP has two factors: one is the increase in the data rate, and ASP can usually be increased by 20%, 25%, and 30% between generations; the second is that we ship these devices in various forms: both in the chip-down form where the chip is directly welded to the board and in the form of a smart cable module (Smart Cable Module). The chip is sold in the form of a small PCB board and then installed in an active cable (AEC), which has a higher ASP. Equally important, these active cables are attached to the scale-up portion of the network, thus leading to higher uptake rates and overall revenue growth.
Overall, we are very happy with the growth of the Aries family, which continues to grow from 2024 to 2025 to 2026, and we expect 2027 to grow as well.
The last point you mentioned is that after entering the next generation, some links previously connected by copper will be carried by optical connections, which gives us the opportunity to obtain higher content, because the dollar value of a single link on an optical link is much higher than that of copper links. Additionally, some links that previously didn't require a retimer at all will now require both copper and light content, which is a great opportunity for us to continue growing the Aries portfolio.
Jeff Kocher (Regent Financial Research Department, Desimon Leopold): Thank you, I'm Jeff Koche, Simon Leopold (Simon Leopold) asked. Just a quick question about Maywell's merger with Lumentum. Obviously, the deal is intended to make chips more closely integrated with photonic technology, possibly enabling closer collaborative design and better power consumption. I'd like to hear your thoughts on this strategy: Do you think it could constitute a competitive advantage? Will this change the way you think about vertical integration?
Gitendra Mohan: Maywell and Lumentum are excellent companies. We have great respect for both. I believe they have carefully thought out their strategies before they came together; leave this question to them. But I'd love to talk about our own optical strategy. Our optical strategy is more rooted in scale-up: unlike other companies, we don't focus on scale-out or scale-across; our focus is on scaling up. This TAM isn't big today, but as we mentioned in our statement, we expect it to grow into a $40 billion market. We are developing all the components needed to cover this TAM: starting with electronic chips (EICs), laser drivers, and TIA for the NPO generation, to more complex EICs for the CPO generation. At OCP, we also demonstrated silicon photonics (silicon photonics) technology for some NPO modules, as well as connectors that allow us to build complete light engines for NPO and CPO. These light engines will then be deployed to our hyperscale cloud customers, and hyperscale cloud vendors will continue to procure laser components from existing vendors as they do today.
Tripp Smith (Barclays Research, Tom O'Malley): Hello, I'm Tripp Smith. I'm Tom O'Malley (Tom O'Malley) asking a question. Very good results, thank you for taking the questions. What I want to ask is the pace of revenue growth entering the 2027 calendar year. There are some predictions in the market that the current generation of your main customers will peak around the first quarter of 2027, while the next generation platform is not expected to actually scale up until close to the 2028 calendar year. Any info on this would be very helpful, thanks.
Desmond Lynch: I'm Des. Thanks for asking. Obviously, we're not going to comment on the volume pace of end customers. I have already addressed this in part in the previous Q&A: we have reached a large scale on the Scorpio X platform, the initial deployment has been completed, and we will continue to contribute on a larger scale in the future; we will also see new platforms and more customers continue to be added to the model in the future. It's also important to remember the size the Scorpio X family has reached: Gitendra talks about achieving a $1 billion operating rate in the third quarter, and we're optimistic about future growth prospects as we move into the next generation platform.
Bastian Faucon (Susquehanna Research Division, DeMehdi Hosseini): Hello, thank you for accepting the question. I am asking for Mehdi Hosseini (Mehdi Hosseini). I want to go back to your statement about Orion's rack-level design content: you mentioned that each rack captures about $100,000 of content, and the press release says that these clusters (pods) range from 8 to 160 accelerators. I'd like to know more about how to understand this $100,000: How many GPUs or accelerators does this estimate correspond to? My additional question is: How should I understand the gross margin level of the rack-level plan compared to your long-term gross margin target of 70%?
Sanjay Gajindra: Let me disassemble it. Orion's rack design supports up to 10 compute trays (compute trays), and each compute tray can insert up to 8 GPUs or accelerators. Our content, including fabrics, switches, etc., is located at the compute tray level and is used for GPU interconnection. This part includes our content. Instead of running the 10 compute trays independently, some customers want to build a larger cluster, that is, run 80 GPUs as a whole. This requires a scale-up switch and corresponding tray, which requires additional hardware in these cases. And when it comes to 160 GPUs, the two racks are interconnected via an east-west (east-west) topology.
Overall, in such a system, the components we can integrate include: calculating all tray-level and rack-level fabric and switching products; all required signal conditioning products, such as retimers, including AEC; then we monetized COSMOS software for the first time: as I mentioned, COSMOS has upgraded from a component-level solution to a rack-level solution, fabric manager, etc., which provides additional monetization points, superimposed on what I said about $100,000. Finally, we are providing all of this in the form of production-ready reference designs. Obviously, we will receive a certain license fee (royalty) from OEM and ODM partners who adopt these designs and then deliver them to terminal enterprise AI and inference cloud customers.
All in all, there are three monetization points: the board we sell, the software we sell, and the license fee for the reference design. This is the complete picture of how to monetize. We're talking about $100,000 for a single rack, or 80 GPUs, although the reference design itself can support 160. $100,000 refers to a single rack, which is something I want to emphasize. Overall, this is just the beginning of the Orion story, and more versions such as UALink, liquid cooling, etc. will be introduced in the future. Overall, we expect more value to be added through vertical integration of components and software.
As for gross profit margin, we expect this business to increase gross margin. This is based on the business model I just described: we not only provide chips, but also software, and earn additional revenue through the licensing model. Overall, it is intended to increase revenue and enhance existing gross margins. I would also like to point out that this platform is still in its early days, and the market interest is very good. We will of course keep you updated on business progress, and we expect to provide more details on its impact on overall profit and loss in the first quarter.