AMD's Helion rack design comes in CPU-only and CPU-GPU versions for agentic AI

A team that buys a CPU-only Helion rack for its agents still has to find somewhere for the models to run. TechRepublic wrote that such a rack does not remove the need to plan for model inference elsewhere ^[6].

AMD describes the CPU-only version for hyperscalers that already operate separate accelerated infrastructure for model inference. It describes the integrated CPU-GPU version for enterprises that want agent execution and inference in one rack ^[5]. On that description, an enterprise with no inference capacity of its own is the intended buyer of the integrated version, and the CPU-only rack is aimed at an operator with accelerated infrastructure already running.

AMD's Dan McNamara, senior vice president and general manager of Compute and Enterprise AI, said: "With AMD Helion, we are bringing AMD EPYC performance, open software and flexible rack designs together so our partners can build systems around customer needs and make efficient use of available power." ^[11] Lenovo's Scott Patti, a vice president in its Infrastructure Solutions Group, said Helion "will give added versatility to meet customers where they are" ^[12].

Both statements describe intent. Helion itself is a reference design, and it is partners who tailor it to workloads ^[2]. TechRepublic noted that nothing in the announcement shows Helion will cut costs, raise utilization or beat competing systems ^[13].

We'd cross two facts a team can check today. The first is whether it already has model capacity it can reach, owned or rented. The second is whether its agent steps are mostly orchestration, tool calling and code execution, or mostly inference ^[4].

- Has models, CPU-heavy agents: the CPU-only case. Measure the network connections, latency and operating cost of reaching those models before anything else ^[6]. - Has models, inference-heavy agents: the models already have a home, so ask what a second set of GPUs would add. - No models, CPU-heavy agents: the gap is inference. Test representative workflows with the cost of any external inference service included ^[15]. - No models, inference-heavy agents: this is the buyer AMD describes for the integrated CPU-GPU rack ^[5].

We recommend placing the models first and sizing the rack second. The tradeoff is patience: TechRepublic advises asking suppliers for validated hardware configurations, supported software and clear responsibility for system support before committing ^[15], and those answers come from partners whose systems AMD says are planned for November 2026 ^[10].