AMD Unveils Helion Open Rack With 32,768 EPYC Cores For Agents

AMD said it will ship a rack-scale reference design next month that packs up to 32,768 EPYC CPU cores into a single rack. The design, called Helion, is pitched as the unit of agentic-AI buying rather than another chip SKU, and in its announcement AMD frames it as an open blueprint for agent fleets.

"Agentic AI is not a single workload. It spans personal agents, enterprise applications, code execution and complex workflows, each with different compute requirements," said Dan McNamara, SVP and GM of Compute and Enterprise AI.

Helion ships in two shapes. A CPU-only configuration uses 6th Gen AMD EPYC processors and scales to the 32,768-core number for hyperscale customers running agent orchestration. A combined all-in-one version pairs EPYC CPU nodes with AMD Instinct MI350P GPU nodes for enterprises that want one rack to carry both the agent control plane and the model inference. Both variants come air- or liquid-cooled, and AMD has bolted an integrated token router into the stack to route agents' calls across frontier and open-source models.

The partner list is wide. Cisco, Dell, HPE, Lenovo, Supermicro, Hon Hai, QCT and Wiwynn are developing configurations. Supermicro's statement specifies ORv3 compliance; the AMD release does not pin the overall rack to a specific OCP revision.

The configurations are planned for November 2026, and AMD said it will contribute the Helion specification to the Open Compute Project. A reference design donated to OCP is an invitation to the ODMs to build without paying for a proprietary rack-level stack, which is a posture Nvidia's current rack-scale products do not take. Against the 457 AI-infrastructure stories our tracker has logged in the last 90 days, Helion stands out as an open-rack blueprint written for agent orchestration rather than for training throughput.