AMD says Helios AI racks are running at OpenAI, Meta, Microsoft and Oracle

AMD’s rack-scale AI push now includes working systems at four major customers, not just deployment commitments. In its October 12, 2026 account of Lisa Su’s OCP Global Summit keynote, AMD said initial Helios systems are installed and running at OpenAI, Meta, Microsoft and Oracle. Customer deliveries and workload validation remain underway.

Helios moves from rack design to customer workloads

The deployment disclosure came one year after Helios made its 2025 Open Compute Project debut. AMD says production is ramping. The milestone concerns initial systems, with the rollout still progressing through delivery, deployment and testing against customer workloads—not a completed deployment program.

Helios brings several parts of the computing system into one rack. It combines AMD Instinct graphics processors, EPYC central processors, Pensando networking and ROCm software. AMD positions it for large-scale model training and inference—the work of running trained models to produce responses. Computing, memory and networking are developed together to support those workloads.

AMD’s system-level argument goes beyond processor speed. Large models need enough memory to hold their data and fast connections within and between racks. Customers also need equipment they can deploy, maintain and expand reliably. Helios is designed to address those requirements through an integrated rack built on open standards.

There are going to be a lot of accelerators out there. It's not one-size-fits-all. It's not one company has all the best parts. What we want to do is make it really easy for a heterogeneous data center environment.

Lisa Su, AMD chair and CEO, at the 2026 OCP Global Summit

An integrated rack built on shared designs

AMD says it co-engineered Helios with Meta and the broader Open Compute Project community, using shared designs for the rack, power systems and hardware management. That gives its integrated system a community-designed foundation, rather than requiring AMD to develop every surrounding component on its own.

Microsoft’s July 20 Azure announcement supplies earlier context. It said it was bringing Helios to Azure to power an upcoming offering for AI inference. That statement established Microsoft’s adoption plans; AMD’s October disclosure goes further by saying initial systems are now installed and running there.

AMD also points to continued ROCm software updates as a way to improve performance on existing hardware. The company’s pitch is that customers can get more work from the same investment, rather than replacing hardware for every improvement. That is a stated potential benefit, not a quantified performance result.