AMD is introducing Helion, an open rack blueprint designed specifically for agentic AI. The design follows an Open Compute Project specification and combines EPYC CPUs with Instinct GPUs as needed. The first configurations are expected in November.
Helion builds on Helios, the rack-scale platform that AMD previously introduced for large-scale AI training and inference. That platform contains 72 Instinct GPUs per rack. While Helios focuses on GPU computing power, Helion emphasizes the balance between CPU and GPU. Companies tailor CPU performance, core density, and power consumption to the GPU acceleration required by their AI agents.
Agents constantly invoke tools, execute code, and coordinate workflows. Much of that work falls to the CPU. GPUs primarily handle model inference.
“Agentic AI is not a single workload. It spans personal agents, enterprise applications, code execution and complex workflows, each with different compute requirements,” says Dan McNamara, senior vice president and general manager of Compute and Enterprise AI at AMD.
From CPU-only to all-in-one
AMD says the blueprint supports virtually any configuration, whether air-cooled or liquid-cooled. The company provides two examples. The first is a rack consisting exclusively of sixth-generation EPYC processors, capable of supporting up to 32,768 CPU cores. This design is intended for hyperscalers who want to run agents on dedicated capacity and offload model inference to separate infrastructure.
The second example is an all-in-one rack. It combines EPYC CPU nodes, Instinct MI350P GPU nodes, and general-purpose compute on a single platform for enterprise environments. AMD states that organizations can deploy Helion within their existing IT infrastructure. Partners tailor the CPU and GPU mix to each deployment’s performance, power, and cost requirements.
AMD plans to contribute the Helion specification to the OCP so that the community can further develop the design. AMD is currently developing rack configurations in collaboration with ecosystem partners, and they expect to be available in November 2026.