For three years the AI hardware story has been GPUs, GPUs, and a side of GPUs. The humble CPU sat in the corner like a drummer in a band where everyone only watches the lead singer. AMD just handed the drummer a solo, and the venue was a rack with up to 32,768 cores.

One Rack, Two Personalities

At the OCP Global Summit in San Jose on October 12, AMD introduced Helion, an open rack-scale reference design built for agentic AI. It comes in two flavors: a CPU-only rack using 6th Gen EPYC processors that scales up to 32,768 cores for hyperscalers that need dedicated agent execution capacity, and an "all-in-one" rack that mixes EPYC CPU nodes with Instinct MI350P GPU nodes for enterprises.

AMD isn't selling the racks itself. Helion is a design built on Open Compute Project specifications, with Cisco, Dell Technologies, Hon Hai, HPE, Lenovo, QCT, Supermicro, and Wiwynn lined up as partners. Both liquid- and air-cooled versions are supported, and AMD says the configurations are planned to be available in November 2026.

Agents Do a Lot More Than Think

Here's the insight hiding inside the spec sheet: an AI agent isn't just a model generating tokens. It's also running code, calling APIs, parsing files, juggling tool calls, and orchestrating multi-step workflows, and much of that is classic CPU work. AMD's framing is that "agentic AI is not a single workload," spanning personal agents, enterprise apps, code execution, and complex workflows, each with different compute needs. Splitting agent execution from model inference is a bet that the GPU shouldn't be stuck babysitting a Python sandbox.

The angle most people miss is the word "open." By contributing Helion to OCP rather than shipping a proprietary box, AMD is courting hyperscalers and server makers who want options beyond a single vendor's stack. It's the same playbook that made open source win the web: make it easy to adopt, let partners build on top, and let the ecosystem do the selling.

The takeaway: if the infrastructure giants are re-architecting data centers around agents, the agent era isn't hype on a slide anymore. It's purchase orders.

Curious what agentic AI could actually automate in your own content workflows or customer experience? Let's talk through the practical version, no 32,768-core rack required.

Source: AMD Investor Relations