AMD introduces Helion rack design for agentic AI
At the Open Compute Project Summit 2026, AMD announced Helion, a new rack-scale reference design for AI agents. Helion can be configured with only CPUs or a mix of CPU and GPU blades, and supports liquid or air cooling. The design targets enterprise deployments that combine agent execution and inference, with options for up to 32,768 Zen 6 cores and 16 Instinct MI350P GPUs.
AMD's Helion blueprint, shown at OCP's 2026 San Jose gathering, differs from its earlier Helios line, which was aimed at frontier training and very large inference. Helion instead targets enterprise racks that handle both agent processing and model serving, with either CPU-only hyperscale setups or combined CPU/GPU enterprise builds.
Configurations can reach 32,768 Zen 6 cores through 256-core Venice Epyc nodes, 16 Instinct MI350P accelerators, and 102 Tbps of network bandwidth. Blade choices include 400 W dense, 600 W balanced, 5 GHz performance, 5U eight-GPU inference, and general-purpose database/storage/technical compute options. Partners include Cisco, Dell, HPE, Foxconn, Lenovo, and Supermicro; partner-built versions are slated for November 2026.
Helion may make agentic AI deployments more practical for enterprises by letting them mix CPU and GPU capacity in one rack, potentially lowering integration complexity and energy waste. This could affect IT teams, cloud providers, and businesses adopting automation, while also increasing demand for power, cooling, and skilled operators. If agent systems spread, customers and workers may encounter more automated services, though benefits and disruptions will depend on how organizations deploy them.