AMD brings rack-scale agentic AI to OCP: Helios ramps production as Helion targets November 2026

AMD moved rack-scale infrastructure for agentic AI from design to actual delivery, with the Helios rack ramping production while customer deliveries, deployments and workload validations proceed. Helios first appeared at the 2025 Open Compute Project summit as a design debut, and now enters production and customer delivery. Initial systems are installed and running at OpenAI, Meta, Microsoft and Oracle, and Helios was co-engineered with Meta and the broader OCP community using shared designs for the rack, power systems and hardware management.

The Helion rack lets customers tailor CPU core density, performance and power to agentic workloads, separating CPU-only configurations for hyperscale deployments from integrated CPU/GPU configurations for enterprises. Helios targets large-scale model training and inference, while Helion extends the same open approach to agentic AI and adds rack configurations customized to the workload. Helion is based on an Open Compute Project spec-design and powered by AMD compute, and its integrated model router directs agents' requests across frontier and open-source models, with configurations under development with ecosystem partners and planned for November 2026.

AMD integrates computing, memory and networking as one system in the rack, with each Helios rack combining Instinct GPUs, EPYC processors, Pensando networking and the ROCm software ecosystem. Large AI models require substantial computing power, enough memory to hold models and their working data, and fast connections within and between racks, needs previously met by separate components. AMD plans to contribute the Helion architecture to OCP, and constantly updated ROCm software can increase performance on the same hardware.

The report covers the open full-stack infrastructure AMD presented in the 2026 OCP Global Summit keynote, for researchers and engineers following AI infrastructure. Helios is ramping production with customer deliveries, deployments and workload validations underway, and initial systems are installed and running at OpenAI, Meta, Microsoft and Oracle; Helion configurations remain under development with ecosystem partners and are planned for November 2026.

Helion's agentic rack configurations remain under development with ecosystem partners and are planned for November 2026, with no deployed customer list or performance data given; Helios production ramp, customer deliveries and workload validations are underway, and the material gives no specific scale or performance comparison.