Astera Labs Unveils Orion: A Compute-Agnostic Rack-Scale Architecture for Token-Efficient AI Factories
Astera Labs unveiled Orion, a compute-agnostic rack-scale reference architecture that lets AI inference operators mix and match GPUs and CPUs via PCIe plug-in cards without redesigning racks. The architecture centers on a 320-lane PCIe switch fabric scaling from 8 to 160 accelerators per pod, delivered as six production-grade reference designs (compute, fabric, and memory trays plus link and compute boards) managed by COSMOS software. Orion-based racks will be manufactured and sold through OEM/ODM partners with availability on track for Q1 2027; the company is showcasing a powered Orion rack at the OCP Global Summit this week.
Key figures
- Availability
- Q1 2027 through OEM/ODM partners
- Fabric lanes
- 320
- Reference designs count
- 6
- Max accelerators per pod
- 160
AI analysis
The rest of the AI analysis, red flags and sentiment are part of Signal8 Pro.
AI-generated analysis of a public disclosure. Not investment advice; verify against the original document.