Astera Labs has announced Orion, an open standard rack reference design built for compute agnostic AI infrastructure. The modular architecture lets data center operators customize processing hardware across different CPUs and accelerators based on supply availability. Deployments through OEM and ODM partners will begin in Q1 2027.
The platform features 6 production grade reference architectures. These include the rack, compute tray, switch tray, memory tray, connectivity board, and compute board. Data center operators can pair different CPUs with AI inference accelerators without being locked to a single vendor. This modular approach allows facilities to adjust hardware combinations based on current market inventory and specific processing demands.
The system centers on a 320 lane PCIe switch developed by Astera Labs. This switch provides predictable low latency communication and expandable memory capacity across racks and pods. Each pod supports up to 160 accelerators to maximize token throughput during AI inference tasks. System designs remain on schedule, with commercial hardware arriving through manufacturing partners in the first quarter of 2027.
