AMD introduced Helion on October 12 as an Open Compute Project spec-based rack blueprint for agentic AI. Its two designs let hyperscalers run agent execution on CPU-only racks connected to separate model-inference infrastructure, while an enterprise configuration combines CPU and GPU capacity. AMD’s announcement describes both as configurations still being developed with ecosystem partners.
On the hyperscale path, AMD says a CPU-only rack based on its sixth-generation EPYC processors can support up to 32,768 CPU cores and connects agent execution to separate model-inference infrastructure. The enterprise concept combines EPYC CPU nodes, Instinct MI350P GPU nodes and general-purpose compute in one rack. AMD says partners can tune the mix for each deployment’s performance, power and cost needs, but the announcement does not report a measured result.
For infrastructure buyers, the distinction is where to place the work. A CPU-only rack can add capacity for agent execution beside an existing inference system; the enterprise design puts agent compute, GPU acceleration, and general-purpose workloads in one platform. AMD also describes a token router that directs agent calls across frontier and open-source models. NeonControl’s CPU-vs-GPU explainer covers the chip-level roles; Helion packages that choice at rack scale.
AMD says partner configurations are under development and planned for November 2026, and that it plans to contribute the specification to OCP for further development. The company provides no system pricing or benchmark data in the announcement, so Helion is a design direction to track rather than a system customers can yet compare on cost or measured throughput.
