Astera Labs, a US company that makes chips for connecting parts inside AI data centres, announced Orion on 12 October 2026 [1]. Orion is a blueprint for a full rack of AI inference hardware that lets buyers choose their own GPUs and CPUs, whatever the brand. Astera says it can scale to up to 160 accelerators in a single pod, and that racks based on it should be available through manufacturing partners from the first quarter of 2027 [1].
A blueprint, not a product on sale
Orion is a reference design. That is a detailed, tested blueprint that other companies can use to build real hardware, so they do not have to design and test every part from scratch. Astera is clear that it will not build the racks itself. In a question-and-answer section of its announcement, it answered “No” when asked whether it would manufacture them, saying it designs and validates the architecture while complete racks and systems are expected to be made, licensed and sold by partners [1]. These partners are OEMs (companies that sell finished servers under their own brand) and ODMs (contract makers that build hardware for others).
Astera named no partners, customers, prices or performance benchmarks.
How it works
Most top-end AI racks come from a single vendor and are built around that vendor’s chips. Orion takes a different route. Accelerators and CPUs plug in as PCIe cards, so an operator can swap in whichever chip is cheapest or easiest to get, without redesigning the rack [1].
PCIe is the standard connection that links processors, graphics cards, memory and storage inside a computer. A PCIe switch fabric is a network of switch chips that lets many of these devices talk to each other directly, a bit like a telephone exchange for chips. According to the announcement, Orion is built around a 320-lane PCIe switch fabric that is designed to scale from 8 to 160 GPUs per pod [1].
Astera says Orion comes as six reference designs: the rack, a compute tray, a fabric tray, a memory tray, a link board and a compute board. All are managed by its COSMOS Rack Manager and COSMOS Fabric Manager software, which handles control, monitoring and diagnostics [1]. The company also says it plans to extend Orion to other open connection standards, including UALink, an industry standard designed for linking AI accelerators [1].
Astera is showing a powered Orion rack with interchangeable compute trays at the Open Compute Project Global Summit in San Jose, which runs from 12 to 15 October [1].
Who it is for
Astera says Orion is aimed mainly at neoclouds (newer cloud companies that rent out AI computing) and enterprises building their own inference capacity [1]. Inference is the work of running a trained AI model to answer requests, as opposed to training it.
“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 AI Platform Solutions Group at Astera Labs [1]. Tokens are the small chunks of text that AI models read and write.
The claimed benefits
The benefits are Astera’s claims and have not been independently tested. The company says Orion’s connectivity is designed to help each accelerator produce more tokens per second, and that it can be deployed with “significantly lower upfront capex” than a fully loaded single-vendor rack [1]. It published no figures to back either claim.
What this means
- It is a design, not a rack. Nothing ships until partners build Orion systems, which Astera expects from Q1 2027.
- It is a bet on choice. Orion is pitched at buyers who want to mix chip brands rather than commit to one vendor’s rack.
- The numbers are still missing. With no partners, prices or benchmarks named, the cost and speed benefits are claims for now.
Sources
- Astera Labs, “Astera Labs Unveils Orion: A Compute-Agnostic Rack-Scale Architecture for Token-Efficient AI Factories”, 12 October 2026. https://www.asteralabs.com/news/astera-labs-unveils-orion-a-compute-agnostic-rack-scale-architecture-for-token-efficient-ai-factories/
Related stories
- Lumentum, Astera Labs and Corning to Show PCIe Running Over Light Inside an AI Rack
- NVIDIA Opens Vera Rubin Rack Designs to OCP, Meta Adopts Spectrum-6 and Google’s Falcon Comes to ConnectX-10
- Lightmatter Shows a 204.8 Tbps Co-Packaged Optics Switch Tray Built With Flex, and a Four-Accelerator Optical Compute Tray, at OCP
- Marvell Unveils Second-Generation Structera X CXL 3.2 Memory Expanders With Six Memory Channels
