Summary
At the 2026 OCP Global Summit, AMD said its Helios AI racks are ramping production, with systems installed at four named customers. The company also said Helion rack configurations for agentic AI are under development and planned for November 2026.
AMD used the opening keynote at the 2026 OCP Global Summit on October 12 to describe two rack-scale AI systems: Helios, which the company says is ramping production, and Helion, whose agentic-AI configurations are planned for November 2026. AMD CEO and chair Lisa Su framed both around open hardware designs, shared standards and a mix of processors, networking and software.
Helios moves into customer deployments
AMD says Helios debuted at the Open Compute Project (OCP) in 2025 and is now ramping production. Customer deliveries, deployments and workload validations are underway; the company says initial systems are installed and running at OpenAI, Meta, Microsoft and Oracle.
Each Helios rack combines AMD Instinct GPUs, EPYC processors, Pensando networking and the ROCm software ecosystem. AMD says it co-engineered the system with Meta and the wider OCP community, drawing on shared designs for the rack, power systems and hardware management.
The rack approach treats computing, memory and networking as parts of one system. GPUs perform much of the parallel computation used in AI, while CPUs can support other computing tasks; memory holds models and working data, and networking moves information among components and racks. Integrating these elements is important for large AI workloads, which need them to work together rather than relying on processors alone.
AMD also argued that data centres will use a mix of CPUs, GPUs and purpose-built chips rather than a single processor type. An open design can give system builders and customers a common basis for adapting infrastructure, while leaving room to combine components and software from different parts of the ecosystem.
Helion is aimed at agentic AI workloads
AMD described Helion as a rack framework for agentic AI, in which software agents can retain context, use tools and make repeated model requests while carrying out tasks. A workflow might involve calling a business application, querying a database or running code, so the computing demand can extend beyond the model’s inference—the process of running a trained model to produce an output.
The company says Helion configurations will let customers tailor CPU core density, performance and power to different workloads. Options under development with ecosystem partners include CPU-only systems for hyperscale deployments and combined CPU-and-GPU systems for enterprises. AMD also describes an integrated model router that directs agent requests among frontier and open-source models.
AMD plans to make Helion rack configurations available in November 2026 and says it intends to contribute the architecture to OCP. The proposal reflects a broader infrastructure challenge: supporting AI agents requires coordinating model execution with the processors, software and systems that handle the surrounding tasks. The planned configurations and Helios deployments show how AMD is trying to address those needs with rack-scale systems built through partner collaboration.