At the Open Compute Project Summit 2026 in San Jose, AMD unveiled Helion, a rack-scale reference design aimed at agentic AI workloads. Both The Register and Tom's Hardware report that Helion complements AMD's earlier Helios system: while Helios targets frontier-scale training and large-scale inference, Helion is a more flexible design focused on agent execution and inference, with CPU-only and hybrid CPU-GPU configurations. The design is OCP-compliant, based on the ORv4 specification, and supports both air-cooled and direct liquid-cooled deployments.
The CPU-only hyperscale configuration can hold up to 128 EPYC 9996 processors with 256-core Zen 6 'Venice' dies, for a total of 32,768 cores per rack. Tom's Hardware notes this setup is likely intended as an agentic AI sandbox connected to separate inference infrastructure, a model used by Meta's Muse and OpenAI Dots. The hybrid enterprise configuration instead combines agent execution, GPU inference, and supporting workloads in one rack, with up to 16 Instinct MI350P GPUs and 102 Tbps of networking, at the cost of some CPU capacity. AMD also offers dense, balanced, and high-performance CPU blade options, plus general-purpose blades for database, storage, or technical compute workloads.
Both sources agree that AMD will not sell Helion racks directly; it is contributing the specifications to OCP, and partners will bring systems to market. The Register lists Cisco, Dell, HPE, Foxconn, Lenovo, and Supermicro as on board, while Tom's Hardware says AMD did not name partners during a pre-briefing and that CEO Lisa Su would announce them during the OCP keynote. The two outlets also agree on timing: partner-built Helion configurations are scheduled for November 2026, after AMD revised an earlier first-half-2027 estimate. AMD's SVP Dan McNamara said the design aims to let partners build systems around customer needs and make efficient use of available power, reflecting that agentic AI spans workloads with different compute requirements.