AMD unveils Helion racks for AI agents at OCP summit

AMD's Helion racks offer a unified infrastructure for AI agent execution and inference, reducing complexity for enterprise deployments.

AMD has introduced Helion, a new rack-scale AI system designed for AI agents, at the Open Compute Project (OCP) Summit 2026. This system complements their existing Helios offering, which is geared towards large-scale AI training and inference. Helion focuses on enterprise AI agents, providing integrated compute infrastructure within a single rack that can handle both agent execution and model inference. Configurations include CPU-only options for hyperscale deployments and mixed CPU/GPU configurations for enterprise use. The system can support up to 32,768 CPU cores, 16 Instinct MI350P GPU accelerators, and substantial networking bandwidth. AMD plans to contribute the Helion specifications to the OCP community. Partner-built Helion systems are expected to be available in November 2026, with companies like Cisco, Dell, HPE, and Lenovo supporting the initiative.

The introduction of AMD's Helion racks signifies a strategic push to address the distinct infrastructure needs of AI agents, moving beyond the demands of large-scale AI training and frontier inference. By offering integrated CPU/GPU configurations optimized for agent execution and inference within a single rack, AMD aims to simplify deployment and improve efficiency for businesses leveraging AI for complex workflows. This move indicates a growing market segmentation within AI infrastructure, where specialized solutions for agentic AI are becoming a key differentiator.

The decision to contribute Helion specifications to the Open Compute Project (OCP) suggests AMD's strategy to foster broader ecosystem adoption and interoperability for agentic AI hardware. This open approach could accelerate the development and standardization of infrastructure specifically designed for AI agents, benefiting both AMD and its partners by creating a more robust market. The upcoming availability of partner-built systems in November 2026 will be a critical indicator of market reception and the practical impact of this new infrastructure blueprint.