Summary
AMD has introduced Helion, an open rack-scale blueprint for agentic AI infrastructure, with CPU-led and combined CPU/GPU configurations. Rack systems are under development with partners and are planned to be available in November 2026.
AMD introduced Helion on October 12, 2026, describing it as an open rack-scale blueprint for agentic AI infrastructure. The design is based on an Open Compute Project (OCP) specification design and offers two broad approaches: CPU-led racks for agent execution, or systems that combine CPUs and GPUs for both execution and model inference.
The distinction reflects the different computing tasks involved. AI agents can use models to plan or respond, then execute code, call tools or coordinate multi-step workflows. Model inference—the processing that produces a model’s output—can be assigned to GPUs, while CPUs handle general-purpose execution and supporting tasks. AMD says Helion is intended to let customers match these resources to their workloads and power requirements.
Two rack configurations
For hyperscale deployments, AMD describes a CPU-only rack powered by 6th Gen AMD EPYC processors, supporting up to 32,768 CPU cores. The design is intended to provide dedicated capacity for agent execution, connecting to separate infrastructure for model inference.
The enterprise-oriented option combines AMD EPYC CPU nodes, AMD Instinct MI350P GPU nodes and general-purpose compute in one rack. AMD presents this as an all-in-one approach to agentic AI, bringing agent execution and GPU-based inference onto the same platform.
Helion’s configurations can use air or liquid cooling, according to AMD. The company also says customers and partners can tailor CPU and GPU configurations to the performance, power and cost requirements of particular deployments. AMD describes Helion as compatible with the idea of building on existing IT infrastructure, though the specific rack designs will depend on deployment needs.
The announcement also describes an integrated token router that directs agents’ model calls across frontier and open-source models. In an agent system, routing can determine which model handles a given request; AMD’s announcement does not detail how the router makes those decisions.
An open blueprint, with systems still in development
AMD says it plans to contribute the Helion specification to OCP so the community can develop and adopt it further. The announcement presents Helion as a blueprint rather than a single fixed rack configuration: partners can build systems around the design and adapt them for different workloads.
Rack configurations are currently under development with ecosystem partners and are planned to be available in November 2026. That makes the near-term development the partner-built systems based on the blueprint; the announced configurations are not described as shipping products.