AMD says Helios production is ramping with initial systems running

AMD said it is ramping production of its Helios rack-scale AI platform, with customer deliveries, deployments and workload validation underway. The company said initial systems are installed and running at OpenAI, Meta, Microsoft and Oracle.

The statement adds deployment status to AMD’s July launch announcement. AMD did not say how many systems have been installed, where they are located or whether they are serving production workloads. Earlier statements from the four companies outline plans to use Helios or AMD’s MI450-series hardware, but do not independently confirm AMD’s latest operating-status claim.

Helios combines AMD Instinct GPUs, EPYC processors, Pensando networking and the ROCm software stack in a rack-scale system for large AI workloads. At launch, AMD specified 18 four-GPU compute trays per rack, totaling 72 Instinct MI455X GPUs.

AMD also said a Helios rack can provide up to 2.9 exaflops of peak FP4 performance, 1.4 exaflops of peak FP8 performance, 31 terabytes of HBM4 memory and 1.7 petabytes per second of memory bandwidth. Those figures are AMD specifications, not independently observed workload results. The company separately estimated that Helios could deliver up to 30% more tokens per dollar than an Nvidia Vera Rubin NVL72 rack in an AMD-modeled test using projected pricing. AMD said configurations and results may vary.

OpenAI and AMD announced a multi-generation agreement for 6 gigawatts of AMD GPUs in October 2025, with the first 1-gigawatt MI450-series deployment scheduled to begin in the second half of 2026. Meta said in February that shipments for its first AMD GPU deployments, using the Helios architecture, would begin in the second half of 2026.

Microsoft said in July that Helios would power upcoming ND MI455X v7 virtual machines for inference workloads on Azure. Oracle said an initial 50,000-GPU Oracle Cloud Infrastructure deployment based on the Helios rack design was planned to start in the third quarter of 2026 and expand in 2027 and beyond.

AMD’s update distinguishes Helios from Helion, its separate agentic AI rack architecture. AMD said Helion is planned for availability in November 2026.