Key information
Astera Labs has officially announced Orion, a modular rack-scale reference design intended to streamline the deployment of AI infrastructure. By utilizing a compute-agnostic approach, this architecture allows enterprise and neocloud customers to select their preferred CPU and GPU combinations, aiming to transform standard racks into efficient AI factories.
The design addresses significant industry bottlenecks by decoupling rack configurations from specific accelerator platforms. Orion utilizes PCIe plug-in cards to support various compute resources, enabling operators to scale inference workloads without needing to redesign their hardware as market dynamics shift. The system supports up to 160 accelerators per pod, utilizing a high-performance fabric to ensure low-latency communication.
Orion includes a comprehensive suite of components, such as compute, fabric, and memory trays, alongside the COSMOS Rack Manager and Fabric Manager software. These tools provide fleet-wide telemetry, diagnostics, and orchestration capabilities. According to Pravin Patel, the architecture focuses on token-per-dollar efficiency, allowing providers to adapt to the evolving compute landscape.
The company plans to demonstrate a powered Orion rack featuring interchangeable compute trays at the 2026 Open Compute Project Global Summit in San Jose, California. Commercial availability through Astera Labs' OEM and ODM partners is currently scheduled to begin in Q1 2027.
