Signals›Chips & compute
Chips & computePrototype

AMD's answer to Nvidia's AI racks is an open blueprint built for agents.

Unveiled at the Open Compute Project summit, Helion is a rack design rather than a chip — and AMD plans to give the specification away.

ZN ZilverNet EditorialPublished · 4 min read · 1 source Announced Oct 12, 2026
A black server rack branded AMD Helion, door open to show rows of compute nodes
Image credit: AMD
The short versionPrototype
What it is
An open blueprint for "agentic AI infrastructure" based on an Open Compute Project design, in liquid- and air-cooled versions, with a built-in router that sends agents' model calls to different models.
How far along
ResearchPrototypePilotShippingEverywhere
Prototype. Configurations are "under development" with partners and planned for November 2026.
Why it matters
Agents mix a lot of ordinary computing with AI calls. AMD's pitch is one rack built for both — "Agentic AI is not a single workload."
Still open
AMD "plans to contribute" the specification to the Open Compute Project. No performance figures were given.

A blueprint, not a box

AMD announced Helion on October 12 as "an open blueprint for agentic AI infrastructure based on an Open Compute Project spec-design." The Open Compute Project is an industry group, started by Facebook, that publishes open hardware designs any manufacturer can build — so rather than a single product, Helion is a design that server makers will build to.

It builds on AMD's Helios rack, and includes an integrated "token router" that directs agents' model calls across both frontier and open-source models.

Two example racks

  • CPU-only, aimed at hyperscalers: 6th-generation AMD EPYC processors, up to 32,768 CPU cores per rack.
  • All-in-one, aimed at enterprises: EPYC CPU nodes, AMD Instinct MI350P GPU nodes and general-purpose compute in one rack.

The emphasis on CPUs is the point. An agent spends much of its time doing ordinary work — calling tools, reading files, waiting on networks — between model calls.

"Agentic AI is not a single workload." — Dan McNamara, SVP and GM, Compute and Enterprise AI, AMD

Who's building it

Cisco, Dell, Hon Hai, HPE, Lenovo, QCT, Supermicro and Wiwynn all endorsed the design. Supermicro called it "ORv3 compliant," meaning it fits the Open Compute Project's standard rack format. HPE referred to a recent customer agreement to deploy AMD's Helios racks.

Where Nvidia's Vera Rubin arrives as Nvidia's own integrated rack, AMD is betting on an open design that many manufacturers build. Whether that wins depends on November's configurations — and on numbers AMD hasn't published yet.

Update log
  • Published.