AMD Publishes Helion Rack Blueprint Aimed at Running AI Agents Under Tight Power Budgets

AMD has published Helion, an open rack design tailored for running AI agents rather than training foundation models, with partner systems scheduled to arrive in November. Presented at the Open Compute Project summit in San Jose and reported by The Next Web (https://thenextweb.com/news/amd-helion-agentic-rack-power), the specification targets data centers facing severe power bottlenecks by maximizing the volume of active agents supported per watt.

The blueprint arrives in two distinct configurations: a CPU-only build for hyperscale operators and a hybrid CPU-GPU setup for enterprise deployments. In its densest CPU-only layout, the Helion rack accommodates up to 32,768 cores across 128 sockets of AMD's 256-core Venice processors based on the Zen 6 microarchitecture, housed in either 400W dense or 600W balanced blades.

By contrast, the mixed enterprise rack limits graphics acceleration to a maximum of 16 Instinct MI350P chips split across two independent eight-GPU servers operating without a scale-up fabric. Training workloads remain relegated to AMD's Helios design, which incorporates 72 higher-tier MI455X accelerators alongside 31 terabytes of HBM4 memory.

A central component of the enterprise Helion system is an integrated token router. According to AMD's product release, the router evaluates outbound requests from software agents and directs calls either to frontier foundation models hosted in the cloud or to local open-source models executing directly on the rack. Dan McNamara, head of compute and enterprise AI at AMD, framed the design as a way for infrastructure operators to make efficient use of constrained power allocations.

Power availability has become a primary hurdle for European data center expansion. In Denmark, the national grid operator halted new connections in March amid roughly 60 gigawatts of queued project applications against a peak national demand of approximately 7 gigawatts. A separate survey of 300 data center decision-makers across the United Kingdom, Ireland, and the Nordics ranked grid constraints as the primary barrier to delivering sovereign compute capacity. Those limits run parallel to policy ambitions under the European Union's Cloud and AI Development Act adopted in June, which seeks to triple regional data center capacity within seven years and open bidding for up to seven AI gigafactories valued at approximately €30 billion.

While AMD has outlined the architecture, the rollout remains early: the company has not published pricing, has not yet formally contributed the finished specification to the Open Compute Project, and has named no European system builders among its initial manufacturing partners.