AMD introduces Helion, an open rack blueprint for running AI agents at data-centre scale

AMD introduced Helion on 12 October 2026 as an open blueprint for rack-scale infrastructure designed for AI agents. The company says customers can tune CPU performance, core density and power to the work agents perform, with optional GPU acceleration.[1]

The announcement describes two configurations: one CPU-only rack for agent execution that connects to separate model-inference infrastructure, and an all-in-one rack with CPUs, GPUs and general-purpose compute. AMD says partner configurations remain under development and are planned to become available in November.[1]

Helion is an open rack blueprint based on an Open Compute Project specification design. AMD says partners can adjust CPU core density, performance and power to match agent workloads. It is a system-level design, not a new standalone chip or a finished rack that a customer can order today.[1]

The first configuration pairs 6th Gen AMD EPYC processors with up to 32,768 CPU cores. AMD says it is intended for hyperscale customers that want dedicated capacity to execute agents while connecting those agents to separate infrastructure for model inference.[1]

The second option combines EPYC CPU nodes, AMD Instinct MI350P GPU nodes and general-purpose compute in one rack. AMD says designs may use air or liquid cooling. The announcement describes examples rather than a fixed product specification, so buyers should not assume every partner will offer the same components.[1]

AMD describes agentic AI as several kinds of work, from personal and enterprise agents to code execution and complex workflows. Its design separates agent execution from the model that reasons over a request. In a CPU-only rack, the agent process can run on Helion while calls reach a separate model-inference system.[1]

The combined design brings CPU and GPU nodes into one rack. AMD also describes an integrated token router intended to direct agents’ model calls across frontier and open-source models. The company has not published independent tests, workload limits or comparative operating costs for these configurations.[1]

This flexibility may help infrastructure teams match different workloads to available hardware, but that is a design goal rather than a measured customer result. AMD says partners can tailor configurations to customer needs, power requirements and existing systems. The final performance will depend on the partner’s hardware, software, cooling and workload choices.[1]

AMD says Helion rack configurations are being developed with ecosystem partners and are planned for November 2026. It plans to contribute the specification to OCP for further development and adoption. The announcement does not name a Malaysian deployment, reseller, price, delivery schedule or service agreement.[1]

For an SME that uses a cloud or data-centre service, the practical question is whether a provider can supply the agent execution and model inference capacity the business needs. Buyers can ask where each part runs, how usage is priced and what limits apply. AMD’s release does not answer those local service questions.[1]

Large infrastructure buyers should compare a partner’s final specification with their own workload and facilities. Check memory, networking, power draw, cooling, software support and maintenance terms before comparing a rack with cloud capacity. AMD has announced a design and partner development plan, not a verified local offer.[1]