TD Cowen: Astera Labs Eyes $1,000 Per GPU Content and a Full Optical Engine Built In-House
TD Cowen 54th Annual TMT Conference, May 27, 2026 — VP of Finance and IR Nicholas Aberle outlines the company's expanding connectivity portfolio and optical roadmap
From $50 to $1,000 Per Accelerator — And Still Climbing
The most important number Astera Labs wants investors to hold in their heads is $1,000. That is the company's current blended content opportunity per GPU or XPU shipped, up from a sub-$100 figure around the time of its IPO. Nicholas Aberle, VP of Finance and IR, walked TD Cowen through the arithmetic in granular terms: Aries PCIe retimers initially contributed $50 to $100 per accelerator. Adding scale-up Aries doubled that figure. Taurus AEC cables layered in additional hundreds of dollars. The Scorpio P-Series fabric switch added more again, and Scorpio X pushed the total to roughly $1,000. "As we think about growing the business and expanding the business going forward, it's going to be layering on additional parts of the portfolio from a connectivity standpoint to continue to grow that $1,000 up much higher than that," Aberle said. The statement is not idle ambition — the optical engine program described later in the session is precisely the vehicle intended to drive that number materially higher.
Scorpio X: Early Volume Production Now, Back-Half Inflection Ahead
Scorpio X, the scale-up fabric switch, is the single most consequential near-term revenue driver. Aberle confirmed that preproduction volumes were shipping in the second half of 2025, that early volume production commenced in the first quarter of this year, and that both smaller and larger Radix configurations are now shipping in Q2. The addressable market for Scorpio X alone is characterized as $10 billion or more, which Aberle noted is nearly half of Astera's total stated addressable market. Given the richness of content and ASP in that product, even incremental share gains move consolidated revenue meaningfully. The company is targeting ten or more customers for scale-up deployments of Scorpio X. Only the lead customer has been publicly identified, but Aberle was explicit that incremental sockets are in development and will pull through PCIe retimer revenue alongside the switch itself. The remaining execution risk is system-level integration — these are platforms with thousands of components that must operate in unison — but Aberle expressed confidence in the current standing, and TD Cowen's framing of the back half as "hitting the knee in the curve" appeared to go unchallenged.
NVLink Fusion: A Project-Based Win With NVIDIA and AWS, and More in the Pipeline
Perhaps the most strategically novel disclosure was the NVLink Fusion collaboration involving NVIDIA, AWS, and Astera Labs. The product addresses a specific architectural gap: certain hyperscaler-designed accelerators are not natively capable of communicating over NVIDIA's proprietary NVLink scale-up protocol. Astera's solution provides a translation layer — "almost fabric level," in Aberle's words — sitting between the customer's accelerator and an NV switch on a one-to-one basis. "Any time you get to work with guys like NVIDIA and AWS on a project like this is something pretty meaningful and kind of shows you there has been a track record of execution and trust built up between our companies," Aberle said. Critically, this is not a multi-sourced commodity socket. Wins are project-specific and go entirely to the winner. Aberle confirmed that additional engagements beyond the publicly announced AWS collaboration are "percolating," noting that "NVIDIA can be a good matchmaker." Investors should treat this as an emerging revenue stream whose competitive structure is structurally favorable for incumbents with demonstrated execution.
Building the Full Optical Engine In-House — CPO Ambitions Extend to Both Sides of the Link
The optical program is more advanced internally than the public narrative has conveyed. Aberle confirmed that Astera has been investing in optical capabilities for at least two years, assembling teams across electrical IC, photonic IC, fiber coupler, and packaging disciplines. The aiXscale acquisition brought coupler technology that Aberle described as "a little bit more of an art than a science" — a component that is underappreciated relative to the EIC and PIC but that is "very meaningful and influential to the overall scalability of the overall optical engine." A leading AI infrastructure provider has already designed the fiber coupler into a co-packaged optics solution set to ramp in 2027, meaning near-term coupler revenue is visible. The larger ambition, however, is an end-to-end optical engine. Astera is developing its own EIC, driving its own organic PIC, and building the associated packaging, test, and manufacturing flow entirely in-house. The intended output is a CPO solution wrapped around Scorpio X — and the same optical engine deployed on the accelerator side of the link as well. "We want to be in a position where we can provide full link accountability across the entire domain," Aberle said. The company has also adopted a photonics-agnostic architecture, meaning hyperscalers that wish to bring their own PIC can integrate it with Astera's EIC and coupler, while customers who prefer a turnkey solution can use Astera's organic PIC. Near-term product announcements are expected over the coming quarters, with NPO applications and initial optical-based revenue targeted for 2027.
Aries PCIe: Still Two-Thirds of Revenue, and Gen 6 Is Just Beginning
Despite the narrative gravitating toward newer products, Aries remains the financial backbone of the company. Aberle confirmed the TD Cowen estimate of roughly two-thirds of current revenue is directionally correct, down from over 90% at IPO, and noted that the portfolio grew approximately 70% year-over-year in 2025. Gen 6 PCIe is already shipping — Aberle stated that over one-third of Aries revenue in Q1 was Gen 6, and is not aware of any competitor shipping Gen 6 product in volume today. The generational upgrade carries a roughly 20% ASP uplift on a like-for-like basis, and attach rates are expected to rise because higher signal speeds require retimers in locations that previously managed without them. As the scale-up domain migrates toward optics for multi-rack distances beyond seven meters, the scale-out or head-node domain — connecting CPUs, GPUs, storage, memory, and networking — remains firmly wired in PCIe and will continue to deepen Aries penetration.
Taurus AEC: 800G Ramp Now, Long Copper Runway Ahead
Taurus, Astera's active electrical cable product for Ethernet-based scale-out connectivity, is expected to contribute more materially to growth this year as the company's primary customer ramps 800 gigabit infrastructure. On the recurring investor debate about whether AECs have a future beyond current line rates, Aberle's view is that the optical opportunity is additive rather than cannibalistic. Copper's advantages in power, reliability, and cost mean it remains the preferred medium wherever reach requirements can be satisfied. As line rates double from generation to generation, the physics argument for active over passive cable strengthens, expanding the addressable AEC market even before accounting for new customer additions. Aberle positioned Astera as still in the early innings of both 800G and the subsequent 1.6T transition. Customer diversification in Taurus is also an active commercial initiative.
CXL Finally Getting Traction, Driven by AI Inference Economics
Leo, the CXL memory controller, has been a long-promised contributor that repeatedly deferred to AI infrastructure spending priorities. That dynamic may now be shifting. Aberle cited a new design win with a hyperscaler customer for a customized Leo variant targeted at CXL connectivity within an AI inferencing appliance — the first concrete revenue signal from this product in a meaningful AI context. The catalyst is the cost pressure created by rising DRAM prices and memory supply constraints, which are pushing hyperscalers to optimize memory utilization per dollar spent on inference workloads. Broad engagement activity has expanded significantly over the past six to nine months, moving beyond general-purpose compute into AI-specific use cases. Whether this becomes a meaningful revenue line in the near term remains to be seen, but the design win validates the product's relevance and ends what Aberle acknowledged has been a prolonged wait.
Supply Chain: Comfortable Through 2027, Optical Infrastructure Being Built Now
Astera characterized its supply position as solid through 2027, with allocation visibility on both the front-end silicon and back-end assembly, and buffer built in for demand upside. Aberle was candid that the company cannot protect against bottlenecks in other components within the same rack, and that program timing risk from third-party constraints is real, even if not specifically identified today. On the optical supply chain — a domain where Astera is effectively a new entrant — manufacturing capabilities tied to the aiXscale coupler are actively being scaled to support volume production in 2027 and beyond. Building customer confidence in optical supply is described as a parallel workstream to the product development itself.
The Conversations Investors Are Not Hearing
Aberle's closing observation is worth taking seriously. The conversations that matter most to Astera's long-term positioning are not about 2026 or 2027 revenue. They are about the architectural and connectivity challenges customers are trying to solve for 2028, 2029, and 2030 — how clusters scale, how racks interconnect, and how the copper-to-optical transition gets executed at the systems level across increasingly complex accelerator platforms. "The conversations that we're having aren't about what's going to drive revenue in '26 or '27, even — it's how we're going to solve these next-generation connectivity challenges, '28, '29, '30," Aberle said. For a company that has grown content per GPU from $50 to $1,000 in roughly three years, the implication is that the roadmap visibility customers and Astera share internally is considerably more expansive than what the public financial model currently prices in.