32,768 CPU Cores in a Single Rack: AMD Pitches Helion, an Open OCP Reference Design, as the Missing Half of Agentic AI Infrastructure

AMD on Monday announced Helion, a rack-scale reference design built specifically for agentic AI workloads, and said it will contribute the specification to the Open Compute Project. The pitch is a deliberate counterpoint to the GPU-centric data center: as AI shifts from chatbots to agents that plan, call tools and execute multi-step workflows, AMD argues a growing share of the work is ordinary CPU computation that never needed a GPU in the first place.

The headline configuration is a CPU-only rack holding up to 32,768 cores, built from 128 sixth-generation EPYC processors with 256 Zen 6 cores each, aimed at hyperscalers that need to run enormous numbers of agent sessions. An integrated configuration combines EPYC CPU nodes with Instinct MI350P GPU nodes and general-purpose compute in the same rack, for enterprises that want agent execution, model inference and conventional workloads side by side. Both versions support air and liquid cooling.

Dan McNamara, AMD senior vice president and general manager of the Cloud and Enterprise AI business group, said agentic AI is not one workload but a spectrum spanning personal agents, enterprise applications, code execution and complex workflows, each with different requirements. His argument tracks what operators have been reporting anecdotally since coding agents went mainstream: the bottleneck is often not the model but the scaffolding around it, the code execution sandboxes, database calls, tool APIs and state management that sit between model calls.

Helion also includes an integrated token router, software that directs each agent model request to the appropriate backend, whether a frontier model or an open-source model, across different inference providers. That makes the rack a scheduling layer as much as a hardware blueprint, and it reflects how enterprises actually operate agents today: a mixture of APIs, self-hosted weights and cost tiers that changes month to month.

The partner list covers most of the server market: Cisco, Dell Technologies, HPE, Lenovo, Supermicro, QCT, Wiwynn and Hon Hai (Foxconn) are all developing Helion configurations, with availability planned for November 2026. Because the design aligns with OCP specifications, AMD says it can deploy on existing data center infrastructure, lowering the migration cost for customers who have already spent heavily on general-purpose compute.

The announcement is also a competitive statement. NVIDIA has pushed the opposite direction with full-stack GPU racks, and Microsoft this week said its own Maia 300 silicon will carry most of its AI inference. AMD, which supplies MI-series accelerators to OpenAI and others, is betting that the agent era revalues the CPU, and it wants the industry standard for that bet written in open specifications rather than any single vendors product line. Whether hyperscalers dedicate floor space to CPU-dense agent racks at scale will be the test.