Astera Labs announced Orion, a compute-agnostic rack-scale reference design that lets neoclouds and enterprises pool up to 160 accelerators of their choosing on a PCIe Gen 6 fabric. Futurum examines whether an open, bring-your-own-GPU rack can win inference deployments against fully integrated single-vendor systems.
What Is Covered in This Article:
- Orion launch as a production-ready, compute-agnostic rack-scale reference design for AI inference
- Six validated reference designs spanning rack, compute tray, fabric tray, memory tray, link board, and fabric board
- A 320-lane Scorpio X-Series PCIe Gen 6 fabric scaling from 8 to 160 accelerators per pod
- Leo CXL fabric-attached memory targeting the KV cache bottleneck in long-context inference
- ODM and OEM delivery through Accton Technology and Hyve Solutions with availability starting Q1 2027
- Competitive positioning against NVIDIA MGX and NVL racks, AMD Helios, and merchant PCIe fabric rivals
The News: Astera Labs announced Orion, a modular rack-scale reference design with fully validated connectivity that allows customers to choose their own GPU and CPU. Orion comprises six production-ready reference designs, the Orion-CT compute tray, Orion-FT fabric tray, Orion-MT memory tray, Orion-FB fabric board, Orion-LB link board, and the rack itself, all managed through COSMOS Rack Manager and COSMOS Fabric Manager software. At the core is a high-radix, 320-lane PCIe Gen 6 switch fabric built on Astera Labs’ Scorpio X-Series that scales pools from 8 to 160 accelerators per pod, which the company says helps the same GPU extract more tokens per second across AI inference workloads. “We’re entering a new era of AI inference, where operators are judged on the tokens they deliver per dollar, not the hardware inside their racks,” said Pravin Patel, Vice President and General Manager of the Platform Solutions Group at Astera Labs. Astera Labs designs and validates the architecture but does not sell racks: complete systems are built, licensed, and sold through ODM and OEM partners, with Accton Technology and Hyve Solutions validated today. Orion designs will be available through partners starting Q1 2027, and Astera Labs will showcase a powered Orion rack with interchangeable, pre-validated compute trays, in collaboration with Accton, at OCP Global Summit 2026, October 12-15 in San Jose.
Astera Labs Orion Reimagines the XPU Rack as a Compute-Agnostic Platform
Analyst Take: The Astera Labs Orion launch completes the steepest portfolio climb in AI infrastructure silicon. A company that shipped its first retimer in 2019 now hands ODMs a validated design for the entire rack. Orion arrives at OCP Global Summit 2026 as a counter-proposal to the integrated single-vendor rack in which the GPU vendor defines the fabric, the memory hierarchy, and increasingly the business terms. Astera Labs inverts that structure. The GPU becomes a plug-in card, the fabric becomes the product, and the rack becomes a reference design any qualified ODM can build. A deployed Orion rack ships Scorpio X fabric switches, Aries retimers on every link board, Leo CXL controllers in the memory tray, and COSMOS software across all of it, so Astera Labs monetizes nearly every socket except the compute it deliberately declines to choose. The implication is that Orion is less a new product than a new sales motion with reference designs that pull the full connectivity portfolio into the rack at once.
Compute-Agnostic Racks Target the Inference Buyer Underserved by GPU Scarcity
Orion’s thesis is that inference economics reward flexibility over peak performance. Operators must commit to accelerator capacity long before they know which GPU will be available or affordable, and Orion lets them deploy rack infrastructure once and populate it with whichever compute is cheapest via PCIe plug-in cards, moving procurement from a platform bet to a spot decision. The pitch also reaches accelerator startups that design competitive silicon but lack a rack-scale platform to ship it in.
The incumbents are closing the openness gap from the other direction. NVIDIA contributed the Vera Rubin NVL144 MGX rack specification to OCP while keeping NVLink as the scale-up fabric, and AMD built Helios on Meta’s OCP Open Rack Wide specification around 72 Instinct MI450 GPUs. Neither will populate a competitor’s accelerator in its rack, and that distinction is Orion’s clearest differentiation. Whether openness converts into orders depends on inference-focused neoclouds growing fast enough to disrupt the committed single-vendor installed base.
A 160 XPU PCIe Domain Is a Contrarian Fabric Bet Awaiting Token Benchmarks
Orion scales up over PCIe Gen 6 rather than a proprietary or emerging fabric. The Scorpio X-Series’ 320-lane radix supports a two-tier switching topology reaching 80 GPUs in a single rack and 160 across a pod in one PCIe domain, removing the protocol translation to Ethernet or InfiniBand that costs latency and power in conventional designs. Leo CXL controllers in fabric-attached mode give GPUs shared memory for KV cache reuse, a lower-cost alternative to buying HBM capacity through more GPUs, making Orion the first rack-level reference design to treat fabric-attached CXL memory as a standard tray. The caveats are material. Rack-level tokens-per-watt results are still being computed. Orion 1.0 is PCIe Gen 6 and Open Rack only, with UALink deferred to a future generation, and NVLink delivers far higher per-GPU scale-up bandwidth today. Orion’s case rests on sufficient bandwidth per dollar for inference, an argument the unpublished benchmarks must win.
The Reference Design Model Trades Rack Revenue for Silicon Pull-Through
Orion changes what Astera Labs sells without changing what it bills for. ODMs manufacture and OEMs sell, so Orion revenue arrives as Scorpio, Aries, Leo, and COSMOS attach, keeping gross margins at silicon levels and avoiding channel conflict with the hyperscalers behind Q2 FY 2026 revenue of $392.4 million, up 104% year over year, with Q3 guidance of $540 million to $560 million, as Futurum covered in its analysis of Astera Labs’ Q2 FY 2026 results. Orion gives Scorpio X a captive rack-scale socket that no longer depends on winning each fabric design individually.
The risk sits in the channel. Accton and Hyve own the customer relationship, and reference designs are copyable, since Broadcom’s PCIe switches and Marvell’s Structera CXL line supply the same ingredients to any ODM that prefers multi-sourcing. Astera Labs’ defense is validation scope and software, because six pre-validated systems with COSMOS fleet telemetry compress an ODM’s integration cycle, and that time-to-market compression is the product. Q1 2027 availability pits Orion against the Vera Rubin ramp, which is also the moment accelerator scarcity makes the bring-your-own-GPU argument most persuasive.
What to Watch:
- Whether rack-level tokens-per-watt and tokens-per-dollar benchmarks publish before Q1 2027 availability
- Whether named neocloud deployments commit to Orion racks in the first half of 2027
- Whether additional ODM and OEM partners beyond Accton and Hyve complete qualification
- Whether UALink support lands in Orion 2.0 before UALink-native accelerators reach volume
- Whether fabric-attached Leo CXL memory wins production KV cache deployments against GPU-local HBM scaling
Read the complete press release on the Orion launch on the Astera Labs website.
Sources
- Astera Labs Unveils Orion: A Compute-Agnostic Rack-Scale Architecture for Token-Efficient AI Factories, AsteraLabs, October 2026
Declaration of generative AI and AI-assisted technologies in the writing process: This content has been generated with the support of artificial intelligence technologies. Due to the fast pace of content creation and the continuous evolution of data and information, The Futurum Group and its analysts strive to ensure the accuracy and factual integrity of the information presented. However, the opinions and interpretations expressed in this content reflect those of the individual author/analyst. The Futurum Group makes no guarantees regarding the completeness, accuracy, or reliability of any information contained herein. Readers are encouraged to verify facts independently and consult relevant sources for further clarification.
Disclosure: Futurum is a research and advisory firm that engages or has engaged in research, analysis, and advisory services with many technology companies, including those mentioned in this article. The author does not hold any equity positions with any company mentioned in this article.
Analysis and opinions expressed herein are specific to the analyst individually and data and other information that might have been provided for validation, not those of Futurum as a whole.
Read the full Futurum Group Disclosure.
Other Insights From Futurum:
Astera Labs Bets on PCIe 7 as AI’s Next Signal Bottleneck
Astera Labs Aries 7 Goes Optical. Can PCIe 7 Outgrow the Retimer Plateau?
Microsoft Turns Dynamics 365 Into an Ambient CRM Layer
Author Information
Brendan is Research Director, Semiconductors, Supply Chain, and Emerging Tech. He advises clients on strategic initiatives and leads the Futurum Semiconductors Practice. He is an experienced tech industry analyst who has guided tech leaders in identifying market opportunities spanning edge processors, generative AI applications, and hyperscale data centers.
Before joining Futurum, Brendan consulted with global AI leaders and served as a Senior Analyst in Emerging Technology Research at PitchBook. At PitchBook, he developed market intelligence tools for AI, highlighted by one of the industry’s most comprehensive AI semiconductor market landscapes encompassing both public and private companies. He has advised Fortune 100 tech giants, growth-stage innovators, global investors, and leading market research firms. Before PitchBook, he led research teams in tech investment banking and market research.
Brendan is based in Seattle, Washington. He has a Bachelor of Arts Degree from Amherst College.
