AMD introduces an AI-agent rack design with up to 32,768 CPU cores

AMD introduced Helion on October 12, 2026, giving infrastructure partners a blueprint for racks built around AI-agent work. The design offers two paths: dedicate CPUs to executing agents while models run elsewhere, or combine both jobs in one rack. AMD says partner-developed configurations are planned for November, with development still underway.

Different agent jobs, different hardware

AMD’s starting point is that agentic AI is not one uniform computing task. Dan McNamara, its senior vice president and general manager of Compute and Enterprise AI, described a spectrum spanning personal agents, enterprise applications, code execution and complex workflows. Each, he said, has different compute requirements.

Helion translates that argument into a configurable hardware design based on an Open Compute Project specification. AMD says customers can match CPU performance, core density and power with GPU acceleration according to the work their agents perform. The blueprint supports both air cooling and liquid cooling, rather than prescribing one cooling setup.

The CPU-only example uses sixth-generation AMD EPYC processors and supports up to 32,768 CPU cores. It is aimed at hyperscale customers needing dedicated capacity for agent execution. The rack connects to separate inference infrastructure—the systems that run AI models—rather than putting that model-serving capacity inside the same rack.

For enterprises, AMD describes an all-in-one configuration combining EPYC CPU nodes, Instinct MI350P GPU nodes and general-purpose compute. That puts agent execution and model inference on one platform. Partners can adjust the CPU/GPU mix around a deployment’s performance, power and cost requirements; Helion is a blueprint, not a single fixed hardware configuration.

Routing model calls, opening the specification

The announcement also includes a software layer. AMD says Helion combines open hardware and software with an integrated token router that directs agents’ model calls across frontier and open-source models. Its stated role is to route requests to models, complementing the hardware choices for executing agents and running inference.

AMD plans to contribute the Helion specification to the Open Compute Project for community development and adoption. That contribution remains a plan. Tom’s Hardware reports that AMD will not sell the racks itself: original equipment manufacturers and original design manufacturers will bring systems based on the reference design to market.

Partners take the blueprint toward deployment

AMD’s announcement includes support from Cisco, Dell, Hon Hai, HPE, Lenovo, QCT, Supermicro and Wiwynn. Their statements describe different roles in turning the design into deployable infrastructure, from networking and management to manufacturing and cooling. Three examples show how that work extends beyond supplying processors:

• Dell says its support combines Helion’s architecture with Dell engineering and options to extend deployments with PowerRack networking and storage.
• Hon Hai says it will apply rack-level integration and global manufacturing capabilities while working with ecosystem partners.
• Wiwynn points to rack integration and advanced cooling expertise to tailor systems for hyperscalers’ workloads, power requirements and data-center environments.

Availability is the immediate unresolved point. AMD’s October 12 announcement says configurations are under development and planned for November 2026. But Jake Roach’s same-day Tom’s Hardware coverage, which describes a press pre-briefing, attributes a first-half 2027 availability window to AMD. The two schedules conflict; November should be read as AMD’s published target, not confirmed delivery.