AMD introduces Helion as a configurable rack for AI agents

AMD announced Helion on 12 October, an open rack blueprint based on an Open Compute Project specification. The company release describes CPU-only configurations for hyperscale deployments and combined CPU/GPU configurations for enterprise installations.

The design includes a token router for directing model calls across hosted frontier models and open-source alternatives. A CPU-focused rack would run agents and connect to separate inference infrastructure, while the combined concept brings EPYC processors and Instinct MI350P GPU nodes into one platform.

AMD said it planned to contribute the specification to OCP and that partner rack configurations were under development, with availability planned for November. That is after this archive’s cutoff. The news is therefore a blueprint and product plan, rather than confirmation of delivered systems. Its flexible configurations address the different computing jobs inside an agent workflow, but actual capacity, power consumption and operating cost will depend on a partner’s implementation. Partner configurations will determine those details.