AMD introduced Helion on 12 October 2026 as a rack-scale design for AI agents based on an Open Compute Project spec-design. The company proposes both CPU-only and combined CPU/GPU configurations.
AMD says the CPU-only version, powered by 6th Gen AMD EPYC processors, can support up to 32,768 CPU cores for running agents while connecting to separate equipment that runs models. Its combined version puts EPYC CPU nodes and Instinct MI350P GPU nodes in one rack. An integrated token router is intended to direct agents’ model calls among frontier and open-source models, much as a switchboard sends calls to different destinations. AMD also describes air- and liquid-cooled versions.
Handling a work request could therefore involve an agent running on CPUs and making a model call either to separate equipment or within the same rack, depending on which configuration is used.
AMD plans to contribute the Helion specification to the Open Compute Project. Helion follows the company’s Helios rack system: AI Affairs reported HPE’s $1.2 billion Vultr order for Helios racks and networking gear. AMD plans to make Helion rack configurations available in November 2026.
.jpg){kind=link}