AMD shows Helion, an open rack blueprint for AI agents

AMD introduced Helion, an open blueprint for racks of servers that operate AI agents. Agents are software that does many steps without help from a person. AMD states that it will give the Helion specifications to the Open Compute Project, a group that shares hardware blueprints. Partners include Dell, HPE and Lenovo. AMD gave no test results.

How it works

Helion has 2 types of rack. The first is for hyperscalers, which are very large cloud operators. It uses only 6th Gen AMD EPYC CPUs, with a maximum of 32,768 cores in one rack. The AI models operate on other hardware.

The other type is for companies. It puts CPU nodes, AMD Instinct MI350P GPU nodes and general computer parts in one rack. A token router sends agent calls to frontier models or to open source models.

The numbers

The one number is 32,768 cores for each CPU rack. AMD gave no data on speed or electrical supply. A rack can use liquid or air to cool the parts.

What is not known

The blueprints are in development with partners. AMD states that the last racks start in November 2026. AMD gave no price and no test results, and it gave no customer.

Insight: AMD opens the rack to win the standard

AMD made Helion open after Helios, its earlier open rack. On 10 October, AMD and TCS wanted to make a blueprint for a data center in India. On 13 October, NVIDIA opened its software that gives GPU resources to AI work. This reads as a race for the open standard for agent racks, and the blueprint that partners use first can set it. Leaders must not lock in one maker at this time.