AMD unveils Helion, a rack built for AI agents that can run on CPUs alone

AMD on Monday, Oct. 12, 2026, introduced Helion, an open rack-scale blueprint for agentic AI built on an Open Compute Project spec, at the OCP Global Summit in San Jose. Hyperscalers can get a CPU-only rack with up to 32,768 cores of its 6th Gen EPYC processors dedicated to running agents, while enterprises can get one rack that combines EPYC CPUs, Instinct MI350P GPUs and general-purpose compute. AMD says partner-built configurations are planned for November 2026.

For three years the AI hardware story has been one word: GPUs. Helion is AMD saying the agent era needs a lot of plain old CPUs too, because an agent spends much of its life not thinking but doing: running code, calling tools, waiting on websites, shuffling files. That's CPU work, and AMD sells a lot of CPUs. A rack with 32,768 cores and no GPU at all is a pointed bet that the bottleneck for agents is execution, not just inference. It's also a game AMD would rather play: in GPUs it is still chasing Nvidia, in server CPUs it is a heavyweight. The spec goes to OCP, so anyone can build it, but nobody can buy one until November and AMD hasn't published a price or a performance number. Still, if agents really do turn into millions of little workers, someone has to build the office they work in.

AMD introduced the AMD Helion agentic AI rack on Monday, Oct. 12, 2026: an open blueprint for agentic AI infrastructure based on an Open Compute Project spec design, unveiled at the OCP Global Summit in San Jose. The Open Compute Project, or OCP, is an industry group that publishes open designs for data-center hardware so more than one vendor can build the same kind of rack. AMD says Helion builds on its Helios rack-scale system, which it launched earlier this year for frontier-scale training and large-scale inference. Training is teaching a model. Inference is running a finished model so it can answer. Those lines are AMD’s.

Example configurations AMD published: a CPU-only rack with 6th Gen AMD EPYC processors supporting up to 32,768 CPU cores, designed for hyperscale customers as dedicated capacity for agent execution, connected to separate model-inference infrastructure; and an all-in-one rack combining EPYC CPU nodes, AMD Instinct MI350P GPU nodes and general-purpose compute for enterprise agentic AI. A core is one processing unit on a chip. A GPU is a graphics chip used to run large models. Both configurations come in liquid- and air-cooled versions. Those lines are AMD’s.

AMD says Helion combines open hardware and software with an integrated token router that directs agents’ model calls across frontier and open-source models, and lets enterprises deploy agentic AI in the IT infrastructure they already have. A token is a small chunk of text a model reads or writes. A token router, in AMD’s description, is software that picks which model an agent should call. Those lines are AMD’s.

The Register reports the core count comes from 256-core “Venice” EPYC processors, AMD’s Zen 6 generation; configurations go up to 16 Instinct MI350P GPUs and up to 102 Tbps of networking bandwidth; blades come in dense versions with 400 W EPYC 9006 chips, balanced versions with 256 cores at 600 W, and performance versions with 96 cores clocked at 5 GHz, plus GPU inference blades with eight MI350P GPUs in a 5U server and general-purpose blades for databases and storage. A blade is a thin server that slides into the rack. 5U is five standard rack units of height. Those figures are The Register’s account of the launch.

Dan McNamara, AMD senior vice president and general manager, Compute and Enterprise AI, said: “Agentic AI is not a single workload. It spans personal agents, enterprise applications, code execution and complex workflows, each with different compute requirements.” That quotation is AMD’s.

AMD plans to contribute the Helion specification to OCP. Helion configurations are being developed with ecosystem partners and are planned for November 2026. The release includes statements from Cisco, Dell Technologies, HPE, Lenovo, Supermicro and Wiwynn. Those lines are AMD’s. AMD has not published a price or a performance number against a rival rack.

At the same summit, NVIDIA announced open AI-factory standards and said Meta is adopting its Spectrum-6 Ethernet switches for Vera Rubin AI factories. Those lines are NVIDIA’s OCP blog. They are same-day context, not a second AMD announcement.

The picture is AMD’s product render of a row of Helion racks, each black door lettered AMD Helion. It is the company’s official product image. The frame does not print a calendar date.

In plain terms, AMD said on Monday that Helion is an open rack design for AI agents, shown at the Open Compute Project summit in San Jose. One version is CPU-only, with up to 32,768 EPYC cores and no GPU. Another packs CPUs and Instinct GPUs in the same rack. Partners are due to ship systems in November. AMD has not printed a price.