Astera Labs (ALAB.US) Q3 Conference Call: CFO responds to whether Scorpio X ramp-up can be sustained; CEO declares "Each generation of XPU doubles the content"

According to Zhitong Finance APP, semiconductor company Astera Labs (ALAB.US) released its FY2026 Q3 results after-hours on Tuesday, November 3, and held a conference call. Revenue for the quarter reached a record $584.9 million, up 154% year-on-year and 49% quarter-on-quarter, surpassing the consensus estimate of $550 million and exceeding the company's prior guidance range of $540-560 million. Non-GAAP EPS was $1.32, beating the expected $1.19; GAAP EPS was $0.99. Non-GAAP gross margin was 72.4%, exceeding guidance of 72%; non-GAAP operating margin was 44.7%, up 560 basis points QoQ.

Strong Q4 guidance: The company expects revenue of $730-750 million, with the midpoint representing about 26% QoQ growth, above the consensus estimate of $655.7 million. Non-GAAP EPS is expected to be $1.62-1.66, well above analyst forecasts of $1.43. GAAP EPS is guided at $1.13-1.17, with gross margin about 72%.

Significant changes in product mix: Scorpio smart switches surged to become the company’s largest revenue product line this quarter, primarily driven by the ramp of the Scorpio X-Series 320-lane switch into mass production. During the same period, the company also launched the Aries 7 PCIe 7 signal conditioning portfolio (sampling to selected customers expected in Q4), the Leo 2 and Leo X-Series memory expansion products, and the Orion reference architecture for rack-level AI inference. CEO Jitendra Mohan said the acceleration in design activities and new design wins position the company to increase the chip value per accelerator.

Key Market Focus:

After Scorpio X's Steep Ramp, What Will Sustain Growth in 2027?

The main topic during the Q&A was the sustainability of growth. Joe Moore of Morgan Stanley Research asked directly: Scorpio X’s revenue run rate has exceeded $1 billion annualized, with a very steep ramp; will there be a plateau before switching to UALink? Trip Smith of Barclays referenced market expectations, probing: Main customer deployments could peak in Q1 2027, while the next-gen platform may not ramp until close to 2028—how will 2027’s revenue cadence look?

Management did not comment on end-customer ramp schedules, but provided three responses: (1) The initial Scorpio X platform has largely completed its ramp and will continue contributing at scale; (2) In Q3, shipments of X-Series to a second hyperscale cloud customer began (in pre-production), and shipments for backlog orders from multiple new customers are scheduled for Q4, with onboarding and certification expected to close in H1 2027 and revenue starts contributing in H2 2027; (3) Existing customers are deploying Scorpio X on more new platforms. CEO Jitendra Mohan also provided a long-term narrative: content value per XPU is already over $1,000, and the company aims to double chip content per XPU with every AI platform generation.

From an external perspective, the market disagreement on Astera is not about current growth, but valuation and its sustainability. 24/7 Wall St. data shows that 69% of the 26 covering analysts rate the stock a ‘buy’, with a consensus price target of $387.57, well above the current price. However, AllMind’s September analysis points to core risks: customer concentration (top five customers to account for 84% of revenue before 2025); any change in main customer deployment may cause a revenue “gap,” and gross margin slippage towards 70% will impact valuation multiples more than incremental revenue. The weak after-hours stock movement partly reflects market caution over visibility into 2027 growth off a high base.

Product Mix: Scorpio Surges to No.1 Line, Annualized Revenue Tops $1 Billion

In Q3, the Scorpio smart fabric switch family became the company’s largest revenue product line, with X-Series revenue surpassing P-Series, and the entire Scorpio family reaching over $1 billion annualized revenue run-rate. The X-Series shipped to two customers across multiple platforms in Q3; P-Series saw double-digit QoQ growth, with client base expanding to two U.S. hyperscale cloud providers. For signal conditioning, Aries saw record growth driven by PCIe Gen 6 adoption in both scale-up and scale-out scenarios; Taurus 800G solutions began volume shipments to key customers in Q3. CFO Desmond Lynch expects P-Series to maintain similar growth in Q4, but X-Series, with higher ASPs, will be next year’s primary growth engine.

The ‘Content Doubling’ Chip Narrative

Mohan clarified the long-term growth logic: In high-config systems, the content value per XPU combining fabric switches and copper signal conditioning already exceeds $1,000; the next-gen AI fabric alone can add over another $1,000 per XPU, achieving the first doubling. Further upside will come from integrating optical components, Leo memory controllers, increased attach rates from larger cluster sizes, and ASP increases with each generational speed upgrade—the company aims to double chip content per XPU each generation. Sanjay Gajendra added that every time data rates double, ASP typically rises 20-30% per generation.

Timelines for the Four New Growth Drivers

Management described four incremental growth paths beyond core: (1) Optics: Sampling of smart TIAs and laser drivers for Near-Packaged Optics (NPO) to select customers begins this quarter, with ~six-month qualification, expecting revenue contribution H2 2027; single accelerator content worth several hundred dollars; Co-Packaged Optics (CPO) ramps expected 2028-2029; scale-up optical interconnect market estimated at $40 billion by 2030. (2) Custom chips: Over 10 active custom projects (custom fabric, memory controllers, IO chiplets), with shipments by end-2027 and ramping in 2028. (3) Memory solutions: Leo 2 in pre-production shipment, ramping to high volume by mid-2027. (4) Orion rack-level reference architecture: One 80-GPU rack represents over $100,000 Astera content (includes hardware, software, and license fees); market entry through ODM/OEM partners in 2027.

Q4 Guidance and Operating Leverage

Q4 non-GAAP guidance: Revenue $730-750 million; gross margin ~72%; OpEx $194-200 million; operating margin ~45%; diluted shares about 187 million; EPS of $1.62-1.66. GAAP EPS guidance $1.13-1.17. Q3 non-GAAP OpEx was $162 million, up 19% QoQ, but Lynch stressed OpEx growth was slower than revenue growth, with long-term OpEx target about 30% of revenue, currently well below that; headcount doubled YoY. Q3 operating cash flow was a record $179.2 million; cash and marketable securities at period-end were $1.38 billion, up $125.3 million QoQ.

Full Meeting Transcript

Chief Executive Officer and Co-Founder Jitendra Mohan: Thank you, Leslie. Good afternoon, everyone, and thank you for joining our FY2026 Q3 call. Today, I will discuss Q3 results, market trends, and progress on our AI fabric and signal conditioning portfolio, followed by Sanjay discussing long-term growth, and Des reviewing Q3 financial results and Q4 guidance.

Astera Labs delivered strong results in Q3: revenue at a record $584.9 million, up 49% QoQ and 154% YoY. Operational execution was robust. We broadened our customer base with new design wins and expanded the intelligent connectivity platform with several new product launches. Growth was broad-based, led by the mass deployment ramp of Scorpio X-Series for scale-up networks. Scorpio became our largest product family in Q3, with X-Series revenue surpassing P-Series. The Scorpio line now represents over $1 billion in annualized revenue—a significant milestone on our journey to becoming a critical solution provider for rack-scale AI infrastructure.

Since our last earnings call, we have meaningfully expanded our product portfolio with a record number of launches: 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 cases; and Orion, a new compute-agnostic rack-level architecture targeting enterprise AI and inference applications. These innovations expand our addressable market and use cases, creating more value for hyperscale and enterprise customers while underpinning a diversified long-term growth profile. Q3 saw significant increases in design-in activity both for new and existing customers.

Scorpio continues to establish itself as a diversified AI fabric platform, with momentum across both X-Series and P-Series. In Q3, X-Series shipped to two customers across multiple platforms; we expect pre-production volume for more new customers this quarter. P-Series similarly expanded across multiple platforms, covering two U.S. hyperscale cloud providers. Beyond AI fabric, diversification momentum continues: Taurus 800G completed qualification and began volume shipments to key customers, with sampling to others underway; Leo’s momentum continued, with certifications completed at multiple customers including two hyperscale cloud providers, ramping to high-volume shipment by mid-2027. Custom solutions design activity grew significantly, with over 10 active projects. Finally, our Orion reference architecture-based platform business is seeing robust early customer interest, with first shipments expected to begin in 2027.

AI infrastructure demand remains strong, with hyperscale cloud customers reporting robust revenue growth from AI services. However, physical input factors such as wafer supply, electrical power, and memory remain tight. This makes utilization especially critical. Workload patterns are shifting toward inference, with sustained inference cycles driving fabrics to carry much higher traffic. Operators are now optimizing for two things: latency (shortening time-to-first-token) and accelerator utilization (boosting revenue per token). Our Scorpio AI fabric is purpose-built for these metrics.

As token economics become more important, let me highlight Astera’s differentiators in AI fabric. First, radix: our Scorpio X devices, including the industry’s largest 320-lane configuration, enable PCIe-like scale-up AI fabric supporting massive clusters. Scorpio X delivers high lane count to maximize domain size and support all-to-all flat topologies for full-bandwidth, low-latency accelerator communication. This is especially important for AI inference, as fewer hops reduce collective communication latency, keeping resources allocated to compute vs. shuffling data between switches. Improved utilization translates to higher tokens-per-second, maximizing ROI on accelerator investment.

Second, intelligence: unlike traditional switches, Scorpio X features AI workload awareness and participates in collective communication and computation. Hardware-accelerated HyperCast and in-network compute offload operations like all-reduce and broadcast from accelerators into the fabric itself, delivering up to 2x collective performance improvement. Similarly, fabric-mounted Leo memory controllers offload KV cache, improving time-to-first-token and tokens-per-second. With Scorpio and Leo’s advanced AI-awareness, customers achieve better token economics and higher token throughput with the same number of accelerators.

Third, software: the COSMOS layer transforms our AI fabric and intelligent connectivity solutions into infrastructure that customers can efficiently deploy, manage, and operate at scale. COSMOS spans our entire portfolio, widely adopted in the AI ecosystem; it covers device-level diagnostics as well as fabric and rack management. COSMOS accelerates rack deployment by resolving connectivity issues in software, preventing costly hardware re-spins, and maintains rack uptime with full-fabric telemetry and predictive failure detection. COSMOS also manages HyperCast and in-network compute—directly interacting with AI workloads running on XPUs. Faster revenue conversion and improved AI infrastructure utilization are why COSMOS drives new design wins across our portfolio.

These differentiators are optimizing performance, increasing XPU utilization, rapidly expanding our opportunity funnel, and accelerating design-win conversion. We thus expect robust, diversified growth in the AI fabric portfolio through increased customer value and chip content per accelerator. Today, per-XPU content in fabric switches and copper signal conditioning exceeds $1,000; next-gen AI fabric content alone will contribute well above $1,000 per XPU. Over time, our goal is to cover the full scale-up switch TAM, expanding from PCI Express and NVLink Fusion platforms today to UALink and other protocols-based systems. With memory controller and optical products, rising AI fabric attach rates spurred by protocols like UALink, larger clusters, and rising ASP from every speed migration, we see a clear path to doubling XPU chip content with every new AI platform generation.

In summary, Q3 was another pivotal point in Astera Labs’ evolution. Record results were led by the Scorpio X-Series mass ramp, with continued growth in the signal conditioning portfolio. We saw strong design-win momentum at both new and existing customers and launched a record number of new products. Looking ahead, we see major opportunities in optics, custom chips, memory solutions, and rack-scale infrastructure—built on our core businesses, expanding addressable markets, and underpinning long-term growth. I’ll now turn it over to our President and COO, Sanjay Gajendra.

President, COO, and Co-Founder Sanjay Gajendra: Thank you Jitendra, and good afternoon everyone. Q3 was an inflection point for Astera Labs: we’ve evolved from a device-level vendor to a rack-level infrastructure platform supplier. We’re now integrating chips, hardware, and software, solving cluster-wide connectivity for AI racks and pods with both copper and optical interconnects, improving XPU utilization and ROI. As the smart connectivity portfolio broadens, we’re lifting per-accelerator chip content to thousands of dollars while expanding addressable market opportunities.

Our foundation today comprises two core revenue engines—AI fabric and signal conditioning—while rapidly expanding along four incremental paths: optics, custom chips, memory solutions, and rack-level platforms. With AI systems growing in scale, the challenge is no longer just connecting more accelerators inside a server. As Jitendra noted, customers are building larger scale-up domains and need more bandwidth, lower latency, and smarter data movement to optimize token economics. Our real-world scale-up deployment experience helps us define target features for specific workloads, and thanks to these factors and rising attach rates, we believe we can more than double per-XPU content with each generation.

As data rates double and AI clusters grow, signal conditioning remains a critical challenge to solve. We therefore expect the new PCIe Gen 7 family (covering copper and optics) to surpass the prior generation in both ASP and attach rate. In addition, our Aries gearboxes and redrivers will open up new socket opportunities in AI platforms.

Taurus for 800G started mass production shipments in Q3; as ASP rises and new projects ramp, we expect robust growth in scale-out Ethernet AEC in 2027. We’re also sampling 3.2T retimers to several customers.

Looking forward, we’re developing four new engines to significantly expand TAM. Optics first: our optical interconnect business is progressing well. At the recent OCP Summit, we launched our first optical products, including smart optical drivers and TIAs optimized for scale-up PCIe Gen 7 topologies, targeting NPO and linear pluggable optics (LPO) use cases. We also demoed integration of photonics with the Scorpio X-Series fabric switch. Sampling of smart TIA and laser driver to select hyperscale and OEM partners for NPO begins this quarter; we’re working closely with customers to extend PCIe fabric beyond the rack into pod-level interconnects. For NPO in scale-up, TIA and drivers unlock several hundred dollars in additional content per XPU, with first revenue contribution expected by end-2027. Longer-term, we’re developing fully integrated optical engines to enable optical Scorpio X-series fabric switches. By 2030, the scale-up optical interconnect market is a $40 billion TAM, and we’ll serve it with high-density optical attachment and optical engines for NPO and CPO use cases.

Second, custom: The landscape of AI accelerators keeps evolving, driving demand for tailored smart connectivity. Custom business interest and activity are accelerating for 2026, with over 10 projects in flight. We’re developing custom fabric solutions, memory controllers for dedicated inference, and IO chiplets for next-gen platforms. We expect first shipments by end-2027, ramping in 2028, with a focus on high-value projects leveraging our connectivity IP, robust supply chain, and system-level expertise.

Third, memory. We’ve expanded the Leo intelligent controller family for new applications and use cases. Leo X-Series is a new fabric-attached controller that, together with Scorpio X fabric, creates a dedicated memory tier for KV cache and agent context. We also launched Leo 2 E-Series and P-Series, doubling bandwidth and capacity versus prior generations, supporting both DDR4 and DDR5. As memory bottlenecks increase, Leo enables customers to expand memory, pool commodity memory across racks, and reliably mix DDR4 and new DDR5. Engagement remains strong with hyperscale and inference cloud customers. Leo 2 is already in pre-production shipment, with ramp starting mid-2027.

Fourth, platforms. We recently launched Orion, a compute-agnostic reference architecture for enterprise AI inference clouds seeking turnkey infrastructure. Orion includes six production reference designs, scaling from 8 to 160 accelerators per cluster, integrating COSMOS with enhanced rack and fabric management. These designs utilize our smart connectivity (fabric, signal conditioning, advanced memory, COSMOS) for tightly-coupled, high-performance platforms. We’ve seen strong interest from GPU/XPU vendors; Orion will enter the market in 2027 via ODM and OEM partners. A full Orion rack carries over $100,000 in Astera content, creating significant incremental TAM.

In summary, we remain execution-focused, delivering substantial value in chip, hardware, and software for AI and cloud infrastructure customers. Design activity is accelerating with broad project wins across customers. A record number of new products is broadening the portfolio, deepening solution synergy, and lifting market opportunity and per-platform chip content. Over to Desmond.

CFO Desmond Lynch: Thank you, Sanjay, and good afternoon. I will review Q3 results and discuss Q4 guidance, both on a non-GAAP basis.

Q3 was a strong quarter: record revenue of $584.9 million, up 49% QoQ and 154% YoY, with broad-based growth across AI fabric and signal conditioning portfolios led by Scorpio X-series mass production ramp. Revenue further diversified across a wider base of customers and platforms. Scorpio was the largest product family by revenue in Q3, with X-Series revenue surpassing P-Series, expanding across more platforms and radix configurations; P-Series grew sequentially with both expanded deployments and new projects moving to volume.

On signal conditioning, Aries grew strongly driven by PCIe Gen 6 in both scale-up and scale-out; Taurus momentum improved as well, ramping 400G and 800G solutions in scale-out applications. Q3 non-GAAP gross margin was 72.4%, ahead of 72% guidance. Non-GAAP OpEx was $162 million, up 19% QoQ, reflecting continued R&D investment to expand core and new growth markets. Non-GAAP operating margin was 44.7%, up 560bps from 39.1% in Q2, reflecting significant leverage from ramped revenue. Interest and other income was $13 million; non-GAAP tax rate was 11.4%.

Non-GAAP net income was $243.4 million; diluted non-GAAP EPS was $1.32, up $0.52 QoQ, a 65% increase. Cash, equivalents, and marketable securities at quarter-end were $1.38 billion, up $125.3 million, driven by record $179.2 million in operating cash flow.

Turning to Q4 outlook. We expect revenue of $730-750 million, midpoint representing 27% QoQ growth. Growth will be driven by Scorpio X and P-Series ramping across customers and platforms. Signal conditioning growth will come from Aries PCIe Gen 6 retimers and Taurus 800G ramp. Non-GAAP Q4 gross margin is expected around 72%; OpEx $194-200 million, reflecting continued investment in core AI fabric, signal conditioning, and incremental opportunities described by Sanjay. Non-GAAP operating margin should be about 45%; interest and other income about $13 million; tax rate about 12%. Diluted shares expected at 187 million. Overall, non-GAAP fully diluted EPS is expected to be $1.62-1.66.

Last, Q3 was a truly exceptional period: record revenue, significant operating leverage, and Scorpio becoming the largest revenue family marks a strategic milestone. Looking forward, our growth drivers keep widening: Scorpio X and P ramp at both new and existing customers and platforms; Aries and Taurus benefit from next-generation connectivity adoption; beyond the core, optics, custom, memory, and platform opportunities are emerging—reflecting strategic R&D to broaden the portfolio and full-market coverage. We’ll continue investing across the portfolio, executing with discipline, remaining strongly profitable, and creating long-term value for shareholders. That concludes the formal remarks; let’s move to Q&A.

Q&A Session

Harlan Sur (J.P. Morgan Research): Good afternoon, thanks for taking my question—team execution continues to impress. Regarding the new Orion rack-level reference architecture: institutions like ours are evaluating on-prem deployment of open, open-weight models for lower token inference cost, control over fine-tuning, and, importantly, data security. Such enterprises seek computing platforms that integrate with private DC architectures: air-cooled, compact clusters, flexible support for XPUs and CPUs. Orion appears well-matched. What is current market interest? Have you engaged large enterprises? Across configurations, how much Astera content can you capture?

Sanjay Gajendra: Excellent question, Harlan, thanks. We are very excited about Orion. As you noted, enterprise AI is a significant, fragmented market—diverse workloads, all need out-of-box compute, and critically, tailored to their own workloads. We built Orion as a GPU-agnostic rack: bring your preferred hardware or whatever is available, with all other infrastructure, power, cooling, connectivity, and design pre-integrated as part of the reference.

Astera provides three things: (1) complete reference architectures, whether at node/box or rack level; (2) baseboards populated with our chips, integrating fabric, memory, signal conditioning, and components in a fully optimized and validated manner; (3) software, very exciting, because COSMOS has matured from device-level to rack/fabric management, handling diagnostics and GPU performance orchestration. In summary, Orion directly addresses your use case. There will be multiple generations, content will rise over time. The baseline Orion 1.0 on PCIe scale-up is estimated at about $100,000 in Astera content, monetized via components, boards, and software.

Joe Moore (Morgan Stanley Research): Thanks, and congrats. On Scorpio X: you’re over $1 billion run-rate, very steep ramp. Do you expect a platform plateau after Scorpio X matures and before UALink? Or how should we think about this transition, given PCI Express’s long tail and the migration to other protocols?

Desmond Lynch: Hi Joe, this is Des. We’re very pleased with the execution in 2026—the portfolio is broad-based. Second-half growth is led by scale-up deployments and Scorpio X production ramp, which began in Q3 and continues into Q4. This drove nearly 50% sequential revenue growth in Q3; with Q4 guidance midpoint, we expect another 27% sequential increase.

Longer-term, the initial Scorpio X platform will continue to contribute at scale after the initial ramp. We expect Scorpio X to keep growing, driven by new platforms and continued customer adoption throughout 2027. Even further ahead, thanks to our experience with large deployments, advanced feature sets, and strong customer relationships, we’re well-prepared for next-gen platform growth. Beyond Scorpio X, we’re excited about growth in new areas discussed earlier: Aries 7, Leo, optics—all will layer in during 2027, providing future growth opportunities. Overall, we’re very happy with the portfolio’s strength and momentum.

Brian Curtis (Jefferies Research): Thanks for taking my question—great results. On NPO (Near-Packaged Optics): you referenced design progress and new analog devices at OCP. Can you walk us through how these components will layer on? Are you designing full NPO solutions now, or is that a bit further out?

Sanjay Gajendra: It’s underway now, Brian. As referenced in our remarks, sampling of NPO products with lead customers starts this quarter. Qualification and system integration takes about six months, so we expect NPO deployments to start contributing revenue in H2 2027. These are NPO module components—TIAs, laser drivers, etc. We expect NPO class products to add several hundred dollars of content per accelerator. Further out, NPO will continue to drive growth post-2027, with CPO ramping mainly in 2028-2029 timeframe.

Tore Svanberg (Stifel Research): Thanks, congrats on another record quarter. You now have four new growth engines, all potentially huge TAMs. You said you’ll materially raise OpEx this December quarter. I’m curious: for these four opportunities, does most OpEx investment still lie ahead, or have you already invested heavily?

Desmond Lynch: Hi Tore, this is Des. We’ll continue to invest at appropriate levels for the large growth opportunities ahead. Yes, Q3 OpEx rose to about $162 million; we expect Q4 to increase further. But note, OpEx growth is slower than revenue growth, so there’s clear model leverage. Second half OpEx growth is driven by three factors: ongoing headcount expansion to support customer projects (headcount doubled YoY); variable R&D (wafer, mask, IP, etc.); and increased variable compensation as results beat plan. Again, leverage is clear in our model. Long-term, 30% of revenue is the right reference point for OpEx; however, with today’s high revenue growth, we’re well below that.

Ananda Baruah (Loop Capital Research): Hi, congrats on the progress—thanks for taking my question. Just a clarification: you reference “content per XPU doubling each generation”—does “each generation” mean each new XPU generation, or each new Astera product (like Gen 7, Gen 8), or something else? How should we think about that?

Jitendra Mohan: Good question, this is Jitendra. Today, in some high-config systems, combined fabric plus signal conditioning gives over $1,000 per XPU of content. In the next gen, switching alone aims to add another $1,000 per XPU—this is the first doubling. After that, with the next AI platform or XPU generation, we add more: optical networking, Leo memory controllers for KV cache, larger clusters with ASP uplift, higher speed bringing ASP uplift. Aggregate modeling of these factors gets us doubling per-XPU content each successive generation.

Ananda Baruah: That’s a big content uptick. Thanks.

Jitendra Mohan: Yes, our customers keep building larger clusters; the bigger the cluster, the higher our attach rate. Thanks.

Papa Sila (Citi Research): Thanks for taking my question—congrats on a very strong Q3. For Jitendra or Sanjay, can you elaborate on the 10+ Scorpio X projects you mentioned previously? How many could meaningfully ramp in H1 2027? Also, any update on your UALink timeline? Still view that as more second-half 2027? If so, is it reasonable to expect some Scorpio X projects to convert to UALink?

Sanjay Gajendra: Thanks for the question. As disclosed earlier, 10+ customers are collaborating with us on Scorpio X, all progressing actively. The lead hyperscale platform is now in production, as reflected in Q3 numbers and Q4 guidance. In Q3, we started shipping X-Series to another hyperscale (pre-production), expecting revenue in 2027. In Q4, we have Scorpio X backlog from several new customers, to be shipped this quarter—these will complete development and certification in H1 2027, contributing revenue in H2.

Moreover, customers brought into Scorpio X are expanding to more platforms—scaling up usage. So, not only is the number of Scorpio X customers growing, but deployments are expanding across multiple platforms per customer. Overall, Scorpio X is scaling robustly—visible in both pipeline and revenues: the Scorpio family reached a $1+ billion annualized run rate in Q3.

Jitendra Mohan: I’ll take UALink. We continue to see the UALink ecosystem strengthen: more IP available, more vendors launching products. From a customer perspective, little change: those focused on memory-semantic protocols (like PCI Express) stay on course and will, as you note, transition to UALink over time. Our UALink product work continues—Taurus family launched earlier this year, with ongoing work on switch anchor sockets to align with customer XPU launches. Little change overall: those following UALink keep on UALink, those on Ethernet remain on Ethernet. Our strategy is cover the full scale-up switch TAM: starting with PCI Express, expanding to NVLink Fusion, and actively pursuing UALink. As interest grows, we’ll also cover further scale-up protocols.

Sebastian Nagy (William Blair Research): Thanks, good afternoon. Sticking with Scorpio, but about P-Series. In Q3, X-Series revenue surpassed P-Series. How should we think about P-Series for Q4 and 2027? X-Series carries higher content but is more concentrated; P-Series has lower content but broader customer coverage. How should we view the structure of these two lines going forward?

Desmond Lynch: Hi Sebastian, this is Des. We’re very pleased with the P-Series. In Q3, it achieved double-digit growth, and I expect that rate to continue in Q4 and the future. Looking at next year, hyperscale cloud customers remain a growth engine—two hyperscales have contributed meaningfully in H2 this year, and this continues into next year, with multiple customer projects ramping. So, P-Series continues as a meaningful part of the Scorpio ramp. Of course, with higher ASPs, X-Series will be the primary growth driver next year, but P-Series remains strong.

Quinn Bolton (Needham Research): Congrats on the guidance. Long-term question: you presented four new areas—optics, custom, memory controller expansion, and Orion rack-level platform. Not asking for firm commitments, but do any of these four look likely to reach 10% of revenue in the next few years? How should we view their relative growth rates?

Sanjay Gajendra: First, all four are large TAMs: optics and memory/inference use cases are obvious; custom represents big opportunity in XPU attach and similar; rack is a unique value prop and format for enterprise and inference. All could be big for us, depending on product development, execution, and delivery—a matter of timing.

Near-term, optical ramp, especially NPO, is the most imminent revenue contributor. Custom chips are an area where we’re building capability—over 10 active projects, including custom fabric and memory controllers—a major opportunity in 2028. Orion is seeing fantastic feedback since the OCP launch, especially from enterprise customers and XPU/accelerator vendors; will launch via ODM/OEM in 2027. Exact Orion numbers are TBD, with more details likely next quarter.

Key takeaway: all are material TAMs, requiring innovation and execution, both proven capabilities for us. We expect each to drive growth engines beyond our current core. If all six engines grow strongly, we’ll have achieved our goal of being a great, iconic company.

Sam Feldman (BNP Paribas Research, for Karl Ackerman): Hi, Sam Feldman here for Karl Ackerman. Aries growth is strong—can you break out how much comes from increased content in the PCIe Gen 6 transition vs. unit volume? As optics move closer to compute, how do you view retimer content per XPU?

Sanjay Gajendra: Great question. Aries set records this quarter, and the family continues to grow. We started with Aries PCIe Gen 4, now moving to Gen 5, 6, and Gen 7 announced at OCP. As you noted, ASP expansion comes from two sources: faster data rates drive 20/25/30% ASP uplift per generation; and, we ship in multiple forms: chip-down, smart cable modules (with the chip on a small PCB inside an AEC), the latter carrying higher ASP. Importantly, these AECs plug into scale-up, driving higher attach rates and overall revenue.

Overall, we’re very satisfied with Aries growth, which is expected to continue through 2024, 2025, 2026, and beyond into 2027.

Your last point: in future generations, links previously copper will be optical, allowing higher per-link dollar content since optics is valued higher than copper. Furthermore, some links that previously required no retimer, will need both copper and optical components—expanding the Aries addressable market.

Jeff Koch (Raymond James Research, for Simon Leopold): Thanks, I’m Jeff Koch for Simon Leopold. Short question on Marvell’s merger with Lumentum: clearly aimed at tighter chip-aligned photonics integration and potential for better co-design and power consumption. Do you see this as a competitive advantage? Could this affect your view of vertical integration?

Jitendra Mohan: Marvell and Lumentum are excellent companies, and we respect both—no doubt they have given this strategic move plenty of thought; I'll let them comment on their own strategy. But let me share our optical strategy. Our focus is scale-up, not scale-out/across like others. This TAM is not yet large but, as discussed, expected to grow into a $40B market. We’re building all required components in-house: starting with electronic ICs, laser drivers, and TIAs for the NPO era, all the way to more complex EICs for CPO. At OCP, we displayed silicon photonics tech for some NPO modules and the connectors required for full NPO/CPO optical engines. These engines will deploy in hyperscale clouds, which will continue sourcing laser modules from their usual suppliers.

Trip Smith (Barclays Research, for Tom O’Malley): Hi, Trip Smith for Tom O’Malley. Congrats on the results. Question on 2027 revenue cadence: some market expectations are for main customer deployments peaking around Q1 2027; but the next-gen platform won’t really ramp until late 2028. Any insight on this would be appreciated, thanks.

Desmond Lynch: Hi, this is Des. Thanks for the question. Clearly, we don’t comment on end-customer ramp schedules. Previously I mentioned: Scorpio X is now at scale, initial deployment complete, and will continue contributing at scale; new platforms and customers will continue layering in. Important to note is the size of Scorpio X: Jitendra mentioned a $1B run-rate in Q3, and as we move to the next-gen platform, we remain optimistic on future growth.

Bastien Faucoun (Susquehanna Research, for Mehdi Hosseini): Hi, thanks for taking my question—for Mehdi Hosseini. Going back to your point on Orion rack content: you mentioned $100,000 per rack, and PR says clusters (pods) range from 8 to 160 accelerators. What does the $100,000 represent—how many GPUs/accelerators does that estimate cover? And what level of gross margin should we expect from rack-level offerings vs. your 70% long-term target?

Sanjay Gajendra: Let me break it down. Orion rack designs support up to 10 compute trays, each holding up to 8 GPUs/accelerators. Our content—fabric, switches, etc.—is on the compute tray level for interconnect, representing our content. Some customers join 10 trays as a 80-GPU cluster (as one vs. independent trays), which requires scale-up switching and extra trays, needing more hardware. At 160 GPUs, two racks connect in east-west topology.

In such systems, our monetized content includes all compute/rack-level fabric and switching, all required signal conditioning such as retimers and AEC, and, for the first time, monetized COSMOS software—upgraded from device to rack/fabric-level, providing an additional layer of monetization on top of the $100,000. All components are delivered as production-ready reference designs—with royalties from OEMs/ODMs delivering to end enterprise/inference cloud buyers.

In summary, there are three monetization points: baseboards, software, and reference design royalties; that’s the business picture. The $100,000 refers to one rack (80 GPUs), though the reference design supports up to 160. Key: $100,000 per rack. This is just the beginning—future versions may feature UALink, liquid cooling, etc., generating more value by integrating software and components.

As for gross margin, we anticipate this business to be accretive—based on the model I’ve described: providing chips, plus software, plus additional licensing revenue. Overall, it drives incremental revenue and lifts gross margin. I’ll note the platform is in early days with strong market interest; we’ll keep you updated, expecting to provide more details on its P&L impact in Q1.