AMD introduces Helion agentic AI rack blueprint: OCP-based design, up to 32,768 CPU cores for hyperscale, MI350P all-in-one for enterprise, partner configs due November 2026

Context

Agentic workloads are not one monolithic GPU inference job: personal agents, enterprise apps, code execution and long tool loops stress CPU density, orchestration and model routing differently. On 2026-10-12 AMD published Helion as an open rack blueprint for agentic AI, alongside the Helios training/inference rackscale line, so hyperscalers and enterprises can size CPU vs GPU to the work agents actually do.

Architecture

OCP-based blueprint with air or liquid cooling; AMD plans to contribute the Helion spec to OCP. Example CPU-only racks use 6th Gen EPYC up to about 32,768 cores for agent execution tied to separate inference. Example all-in-one racks combine EPYC, Instinct MI350P and general-purpose nodes. Open software plus an integrated token router to steer agent calls across frontier and open-source models. Ecosystem quotes include Cisco, Dell, Hon Hai, HPE, Lenovo, QCT, Supermicro and Wiwynn; partner configs are in development for November 2026.

Benchmarks

The official release does not publish tokens/s, TCO or third-party rack bake-offs. Hard numbers to cite are vendor claims: ~32,768-core CPU-only ceiling, MI350P in the all-in-one concept, and the November 2026 partner window. Treat unverified bandwidth figures from secondary summaries as non-authoritative.

Developer takeaway

Separate Helion (agent blueprint) from Helios (training/large-scale inference). Hyperscalers should evaluate CPU agent pools plus external inference; enterprises should check whether EPYC+MI350P all-in-one covers local agents and inference. Track OCP contribution and OEM/ODM SKUs—Helion is a reference design, not AMD selling finished racks in the press copy. Ask partners what the token router actually routes and how it fails closed.