AMD introduces Helion rack blueprint for agentic AI

AMD introduced Helion, an open rack-scale reference design for agentic AI infrastructure. The blueprint is designed to let partners vary CPU capacity and GPU acceleration for the work AI agents perform.

AMD said partner-developed configurations are still in development, with availability planned for November 2026. The announcement did not provide pricing, an exact availability date, order status, shipment volume or a named customer deployment.

Helion has two example configurations. AMD says the CPU-only version uses sixth-generation EPYC processors and supports as many as 32,768 CPU cores. In AMD’s design, those processors provide capacity for agent execution while model-inference requests go to separate infrastructure. The core count is a vendor specification, not an independently measured performance result.

A mixed enterprise configuration combines EPYC CPU nodes, Instinct MI350P GPU nodes and general-purpose compute in one rack. AMD presents that arrangement as an all-in-one design for agent execution, model inference and supporting compute. The announcement does not specify the final configuration that each partner will offer.

AMD also describes a token router that can direct agents’ model calls across frontier and open-source models. The released materials do not identify supported model services or explain the routing logic, licensing or security controls, and they provide no test results showing an advantage for agent workloads.

That makes Helion different from AMD’s Helios rack-scale system. AMD launched Helios in July for frontier-model training, fine-tuning and large-scale inference. Helios centers on 72 Instinct MI455X GPUs connected through a scale-up network, alongside sixth-generation EPYC processors, Pensando networking and AMD’s ROCm software. Helion instead offers a CPU-only configuration or a mixed CPU/GPU design for agent execution and model inference. This comparison describes AMD’s stated architectures; it is not evidence that either system performs better for a particular workload.

AMD says Helion’s OCP specification-based blueprint can use air or liquid cooling. The Open Compute Project’s rack and power materials describe the project’s work on 21-inch Open Rack architectures but do not themselves list Helion. AMD said it plans to contribute the Helion specification to the project, so the announcement does not establish that OCP has accepted the design.