Photo: Steve A Johnson / Pexels
AMD launches Helion open rack blueprint for agentic AI workloads
The chipmaker's new rack-scale design leans on CPUs, open standards and a built-in token router to take on Nvidia's GPU-heavy systems
AMD has a new rack, and it is not trying to out-GPU Nvidia. The company unveiled Helion, an open rack-scale blueprint for agentic AI infrastructure, on October 12, 2026, at the Open Compute Project (OCP) Summit in San Jose, California.
What AMD actually built
Helion is a reference design for an entire server rack, tuned for agentic AI. That is the category of AI systems that do more than answer a prompt. They plan, call tools and chain tasks together.
The rack comes in two main flavors. One is a CPU-only setup aimed at hyperscale operators. The other mixes CPUs and GPUs for enterprise buyers who want a bit of both.
On the CPU side, the numbers are large. Helion supports up to 128 EPYC 9006 Series processors, adding up to as many as 32,768 Zen 6 cores in a single rack.
The enterprise version pairs EPYC CPUs with up to 16 Instinct MI350P GPUs. Those GPUs are spread across independent servers rather than fused into one giant accelerator block.
Helion is compliant with OpenRack v3 standards. It supports up to 102 Tbps of networking bandwidth, which is the plumbing that keeps all those chips talking to each other.
AMD says the configurations are designed to optimize core density, performance and cooling. Each knob can be turned depending on which agentic workloads a customer plans to run.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
The token router is the sleeper feature
The most interesting piece may be the least flashy. Helion includes an integrated token router that directs calls between different types of AI models.
The goal is better coordination between those model types. Rather than forcing every request through one expensive model, the router can match the job to the right tool, routing between frontier models, cloud-hosted models, and open-source models running locally inside the rack.
Building on Helios
Helion is not AMD’s first rack-scale swing this year. It builds on Helios, the rack solution the company launched earlier in 2026.
Helios was aimed at frontier AI training and large-scale inference. Helion shifts the focus toward agentic deployments.
AMD also plans to share Helion’s specifications with the OCP community. The company has lined up a roster of hardware partners to build specific Helion configurations. Cisco, Dell Technologies, HPE, Lenovo and Supermicro are all on board.
Availability is targeted for November 2026.
What this means for the AI infrastructure race
Nvidia currently sets the terms in AI hardware, with GPU-focused offerings like its NVL72 systems. AMD is not trying to beat that on Nvidia’s home turf.
By publishing specs through OCP and aligning with OpenRack v3, AMD lowers the switching cost for customers wary of locking themselves into one vendor.
AMD is also pitching Helion as a way to reuse existing IT infrastructure. For enterprises that already own racks, power and cooling built around standard servers, that could be more appealing than ripping everything out for a bespoke AI system.