AMD Unveils Helion Rack Platform for Agentic AI, With Up to 32,768 EPYC Cores per Rack

On October 12, 2026, AMD announced Helion, a rack-scale computing platform designed for agentic AI, at an event in Santa Clara, California. Built around Open Compute Project (OCP) specifications and powered by AMD-designed processors, the platform includes a token router that can route agent model requests to frontier or open models and adapt to different inference backends.

Dan McNamara, senior vice president at AMD, said agentic AI spans scenarios including personal assistants, enterprise applications, code execution and complex workflows, and that Helion is designed to address varying compute requirements through a flexible architecture.

Helion is offered in two configurations. The CPU-only version uses sixth-generation AMD EPYC processors and supports up to 32,768 cores per rack for hyperscale data centers. The integrated version combines EPYC CPU nodes, AMD Instinct MI350P GPU nodes and general-purpose compute units within a single rack, targeting enterprise customers that want an end-to-end agentic AI platform. Both versions support liquid and air cooling, and partners can customize CPU and GPU configurations based on workload needs.

AMD described Helion as an open technology blueprint rather than a proprietary platform, and plans to contribute the full specification to the OCP community, continuing the open-platform approach of its earlier Helios rack solution. The company said Helion can deliver agentic AI capabilities on customers' existing IT infrastructure, helping lower migration and procurement barriers. The product is being developed with ecosystem partners and is expected to be available in November 2026.