AMD Unveils Helion Open Rack Blueprint for Agentic AI at OCP Summit

AMD pulled back the curtain on Helion, an open rack-scale blueprint for agentic AI infrastructure, on October 12, 2026, at the Open Compute Project Summit in San Jose, California. The company says the design is a reference for an entire server rack rather than a single chip, and that it targets agentic systems — AI that plans, calls tools and chains tasks instead of only answering a prompt. AMD intends to publish the specifications through the OCP community and is targeting availability in November 2026.

Helion comes in two primary configurations. A CPU-only version is aimed at hyperscale operators and supports up to 128 EPYC 9006 Series processors, which AMD says can total as many as 32,768 Zen 6 cores in one rack. An enterprise variant pairs EPYC CPUs with up to 16 Instinct MI350P GPUs, distributed across independent servers rather than combined into a single large accelerator block.

The rack complies with OpenRack v3 standards and supports up to 102 Tbps of networking bandwidth to connect the chips inside it. According to AMD, the configurations are built to optimize core density, performance and cooling, with each parameter adjustable depending on the agentic workloads a customer expects to run. AMD has lined up Cisco, Dell Technologies, HPE, Lenovo and Supermicro to build specific Helion configurations.

One of the more notable components is an integrated token router that directs calls between different categories of AI models. AMD says the router can match a job to the appropriate tool — routing between frontier models, cloud-hosted models and open-source models running locally inside the rack — instead of sending every request through one costly model. The stated aim is tighter coordination across model types.

Helion is not AMD's first rack-scale effort this year. It builds on Helios, the rack solution AMD launched earlier in 2026 for frontier AI training and large-scale inference, with Helion shifting the emphasis toward agentic deployments. Nvidia currently sets the terms in AI hardware with GPU-centric offerings such as its NVL72 systems, and AMD is not attempting to compete on that ground.

Instead, AMD is positioning openness as the differentiator. By publishing specifications through OCP and aligning with OpenRack v3, the company argues it lowers the switching cost for customers worried about locking into a single vendor. AMD is also pitching Helion as a way to reuse existing IT infrastructure, which could appeal to enterprises that already own racks, power and cooling built around standard servers.