AMD introduces Helion rack blueprint for CPU-led agentic AI

Two approaches to agent infrastructure

AMD introduced Helion on October 12, 2026, describing it as an open rack blueprint for agentic AI. Configurations are being developed with ecosystem partners, with availability planned for November 2026. AMD also intends to contribute the specification to the Open Compute Project; the announcement describes that contribution as a plan.

One configuration uses sixth-generation EPYC processors in a CPU-only rack with up to 32,768 cores. AMD positions it for hyperscale agent execution linked to separate inference infrastructure. An enterprise concept combines EPYC CPU nodes, Instinct MI350P GPU nodes and general-purpose compute within one rack. The blueprint supports air-cooled and liquid-cooled configurations.

Source: Introducing AMD Helion, Agentic AI Compute at Rack Scale

Local and external inference

AMD’s product documentation describes routing agent inference requests to frontier models in the cloud or to open-source models within the enterprise rack. It presents Helion as a configurable reference design, with CPU and GPU blade choices matched to deployment requirements.

Source: AMD Helion Agentic AI Rack: product and configuration documentation

LANGOM perspective

For a purchasing team, the distinction between running agents and serving their models should shape the first capacity study. Measure the time spent executing code, retrieving data and waiting for model responses. A core-count target alone will not reveal whether the bottleneck sits in execution, memory, storage or inference.

Evaluate the two configurations against the same representative workflow. Include concurrency, response-time targets and the behavior of external model connections during an outage. Where model calls leave the premises, document data-handling rules and assign responsibility for the connection and credentials.

Before selecting a rack, obtain a partner-specific bill of materials, cooling requirements and delivery commitment. Treat November as AMD’s announced planning window until the chosen supplier confirms the exact configuration. Define a workload-based acceptance test and support boundary across compute, networking and software. These are LANGOM’s procurement considerations, rather than measured performance results for Helion.

← All industry news