AMD Helion: An Open OCP Rack Blueprint for Agentic AI

AMD wants racks built for agents, not just giant training jobs. On October 12, 2026, the company introduced AMD Helion — an open blueprint for agentic AI infrastructure based on an Open Compute Project (OCP) spec-design and powered by AMD compute.

Two example configurations

• CPU-only hyperscale rack on 6th Gen AMD EPYC processors — up to 32,768 CPU cores for agent execution, with links out to separate model-inference infrastructure
• All-in-one enterprise rack combining EPYC CPU nodes, AMD Instinct MI350P GPU nodes and general-purpose compute on one platform

Helion builds on the open-platform approach of AMD’s Helios rackscale solution. AMD says customers can tailor CPU core density, performance and power — including air- and liquid-cooled versions — and highlights an integrated token router that directs agents’ model calls across frontier and open-source models.

Availability and partners

AMD plans to contribute the Helion specification to OCP. Configurations are under development with ecosystem partners and are planned to be available in November 2026. The release includes support quotes from Cisco, Dell, Hon Hai, HPE, Lenovo, QCT, Supermicro and Wiwynn.

Dan McNamara, AMD’s SVP of Compute and Enterprise AI, said agentic AI is not a single workload — it spans personal agents, enterprise apps, code execution and complex workflows, each with different compute needs. Helion’s pitch is matching CPU density, performance and power (plus the cited token router) to those jobs instead of forcing one rack design.

Rocks take

Agent stacks spend a lot of cycles on planning, tools and orchestration — not only GPU tokens. An open rack recipe that separates agent execution from inference is the kind of plumbing Home Lab and cloud builders will watch as local MCP agents grow up.