Astera Labs (Nasdaq: ALAB) announced the Orion rack-scale reference architecture, a modular system designed to allow customers to select their own GPUs and CPUs for AI inference workloads, according to a company press release.

Orion is built on open standards and uses a 320-lane PCIe switch fabric capable of connecting up to 160 accelerators per pod. The architecture includes six production-grade reference designs covering the rack, compute tray, fabric tray, memory tray, link board, and compute board. All components are managed through Astera Labs' COSMOS Rack Manager and COSMOS Fabric Manager software.

The system is intended for neocloud and enterprise customers seeking to deploy AI inference infrastructure without committing to a single accelerator vendor. Customers can swap compute resources via PCIe plug-in cards as availability, cost, or workload requirements change.

Orion will not be manufactured by Astera Labs directly. Instead, complete racks and individual systems are expected to be manufactured, licensed, and sold through OEM and ODM partners, with availability targeted for Q1 2027.

Astera Labs plans to extend Orion's roadmap to additional open interconnects, including UALink.

"Orion gives AI inference cloud providers a production-grade, compute-agnostic architecture so they can quickly bring inference capacity online, provide outcome-based pricing for their customers, and adapt as the industry and compute landscape evolves," said Pravin Patel, vice president and general manager of the AI Platform Solutions Group at Astera Labs.

The company plans to display a powered Orion rack at the 2026 Open Compute Project Global Summit, taking place October 12–15 in San Jose, California.