AMD introduces Helion, an open rack blueprint aimed at running “agentic” AI workloads rather than training models. The proposed CPU-only configuration can scale to 32,768 CPU cores in a single rack, with the underlying approach focused on maximizing how many agents can run per unit of power.
Multiple outlets report Helion is designed around an Open Compute Project (OCP) specification and pairs AMD EPYC processors with optional Instinct GPUs for configurations that can be tailored to different deployment needs. Quartz and The Next Web both point to a November timeline for partner systems. The Financial Post adds further details on software integration, including an integrated token router intended to direct agent model calls across different model sources, such as frontier and open-source options. Across coverage, the common emphasis is on rack-scale, configurable hardware density and power efficiency for workloads that orchestrate calls to AI models rather than concentrating on model training.
