Astera Labs launches Orion rack design supporting 160 AI accelerators per pod

AI data centers increasingly need flexible computing setups, and hardware platforms are moving toward open standards and modular designs.

Astera Labs, a company specializing in semiconductor interconnects, has released a compute-agnostic rack reference design called Orion. Built on open standards with a modular structure, it lets users mix and match computing power based on their needs and supply situation.

Orion includes six production-grade designs: rack, compute tray, switch tray, memory tray, connector board, and compute board. It supports various CPUs and AI inference accelerators. At its core is a 320-lane PCIe Switch that provides predictable low-latency communication and scalable memory between racks and pods.

Each Orion pod supports up to 160 accelerators with strong token throughput. The design will be available through OEM and ODM partners in the first quarter of 2027.