Interconnect supplier Astera Labs unveiled Orion, designed for 8–160 accelerators per pod, with partner availability targeted for Q1 2027.

Astera Labs’ SEC-filed earnings release describes Orion as a compute-agnostic, rack-scale reference architecture that lets customers choose their GPUs and CPUs. Its six production-grade reference designs cover the rack, compute tray, fabric tray, memory tray, link board, and compute board.

Orion uses a 320-lane Peripheral Component Interconnect Express (PCIe) switch fabric designed to support 8–160 accelerators per pod. This is an architectural scaling range, not a disclosed shipment count, a rack count, or evidence of an installed customer cluster.

All designs integrate with COSMOS Rack Manager and COSMOS Fabric Manager for orchestration, management, fleet-wide telemetry, and diagnostics. Astera Labs identifies neocloud and enterprise customers building AI inference capacity as the primary audience and says the architecture is intended to increase tokens per second from the same accelerator, without supplying a numerical Orion benchmark in the excerpt.

At the 2026 Open Compute Project (OCP) Global Summit in San Jose, the company demonstrated a powered Orion rack with interchangeable, pre-validated compute trays containing different CPU and accelerator configurations. Availability through OEM and original design manufacturer (ODM) partners was targeted to begin in Q1 2027; that remains the company’s stated schedule rather than an independently verified delivery outcome.