AMD unveiled Helion, a reference design for server racks tailored for agentic AI, at the Open Compute Project Summit on October 12, 2026. The architecture supports two configurations: a CPU-only setup for hyperscale operators and a hybrid version for enterprise use.
The CPU-focused configuration accommodates up to 128 EPYC 9006 Series processors, providing up to 32,768 Zen 6 cores. The enterprise model integrates EPYC CPUs with up to 16 Instinct MI350P GPUs distributed across independent servers.
Helion complies with OpenRack v3 standards and offers up to 102 Tbps of networking bandwidth. A key feature is an integrated token router designed to distribute tasks between frontier, cloud-hosted, and local open-source AI models.
AMD plans to share the technical specifications with the OCP community. Hardware partners including Cisco, Dell Technologies, HPE, Lenovo, and Supermicro are set to develop specific configurations based on the Helion blueprint.
The design aims to reduce vendor lock-in and allows enterprises to utilize existing IT infrastructure for AI workloads. This release follows the launch of the Helios rack solution earlier in 2026, which focused on large-scale training and inference.
