AMD’s Oct. 12, 2026, announcement presents Helion as an open rack-scale blueprint for agentic AI infrastructure. Its two examples split the work differently: one uses CPUs to run AI agents linked to separate model-inference infrastructure; the other combines CPU and GPU nodes with general-purpose compute. AMD says partners are developing systems based on the design, with configurations planned for November 2026. AMD’s announcement
Two configurations for different workloads
AMD describes agent execution as the work of running AI agents; inference is the step in which a model generates a response. In Helion’s CPU-only example, those jobs sit in separate infrastructure. The all-in-one example puts CPU and GPU compute together for enterprise agentic-AI use.
| Configuration | Compute components | Intended workload |
| CPU-only | Sixth-generation AMD EPYC processors; up to 32,768 CPU cores per rack | Hyperscale agent execution connected to separate model-inference infrastructure |
| All-in-one | AMD EPYC CPU nodes, AMD Instinct MI350P GPU nodes, and general-purpose compute | Enterprise agentic AI |
That distinction is the practical point: the CPU-only design is for running agents alongside separate inference capacity, while the combined example brings CPU and GPU compute into one platform. AMD has not announced a specific CPU-to-GPU mix for the combined configuration.
Routing model calls and cooling
AMD says Helion pairs open hardware and software with an integrated token router. The router is designed to direct agents’ model calls across frontier and open-source models. AMD also says the blueprint supports air-cooled and liquid-cooled versions.
Rack layout, standards, and partner plans
Tom’s Hardware reports that the rack design accommodates 44 OpenU trays. It describes each tray as 21 inches wide and 1.9 inches tall, within a 21-inch rack. The outlet also reports up to 128 EPYC 9996 CPUs in the CPU-only configuration, and up to 16 MI350P GPUs with 102 Tbps of networking in a hybrid configuration. Tom’s Hardware’s rack-layout report
The rack-standard labels differ: a Supermicro statement in AMD’s announcement calls Helion ORv3-compliant, while Tom’s Hardware describes it as ORv4-based and compliant with DC-MHS. AMD’s announcement Tom’s Hardware
AMD says it plans to contribute Helion’s specification to the Open Compute Project (OCP). It names Cisco, Dell Technologies, Hon Hai Technology Group, HPE, Lenovo, QCT, Supermicro, and Wiwynn as ecosystem supporters. AMD’s stated plan is for partner-developed configurations to be available in November 2026.
