AMD Helion packs up to 32,768 CPU cores into one open rack for AI agents

AMD has introduced Helion, an open rack blueprint for agentic AI, built on an Open Compute Project design. Its CPU-only version scales to 32,768 cores of 6th Gen EPYC processors, with partner systems planned for November 2026, according to AMD.

What AMD announced

The company announced Helion on October 12, 2026. It is a reference design rather than a single product: partners can adjust CPU core density, performance and power to the work AI agents do, and choose liquid or air cooling. AMD also says Helion includes an integrated token router that directs an agent’s model calls across frontier and open-source models.

AMD described two example configurations:

• A CPU-only rack with 6th Gen EPYC processors and up to 32,768 cores. It targets hyperscale customers and is meant as dedicated capacity for running agents, linked to separate model inference infrastructure.
• An all-in-one rack that combines EPYC CPU nodes, Instinct MI350P GPU nodes and general-purpose compute on one platform, aimed at enterprise agentic AI.

AMD says the design lets customers deploy agentic AI in IT infrastructure they already have. The company plans to contribute the Helion specification to the Open Compute Project. Configurations are under development with ecosystem partners and are planned for November 2026.

Dan McNamara, AMD’s senior vice president and general manager of Compute and Enterprise AI, framed the idea this way: “Agentic AI is not a single workload.” He listed personal agents, enterprise applications, code execution and complex workflows as cases with different compute needs.

Who is backing it

AMD’s announcement includes supportive statements from Cisco, Dell Technologies, Hon Hai Technology Group, HPE, Lenovo, QCT, Supermicro and Wiwynn. Supermicro said it will bring Helion to market through its Data Center Building Block Solutions and described the architecture as ORv3 compliant. Wiwynn said it will contribute rack-level integration and cooling expertise for hyperscaler customers. None of the statements gives pricing, order dates or customer commitments for Helion itself.

Why it matters

The technical shift is where AMD puts the processors. Its existing Helios rack is a GPU system: SDxCentral reported in October 2025 that it holds 72 MI450 GPUs and 31 TB of HBM4 memory, and is designed on Meta’s Open Rack Wide specification. That is a machine for training and large-scale inference. Helion starts from the opposite end, with CPUs doing the work and GPUs optional.

The bet rests on a claim about where agents spend their time. TrendForce says today’s AI data centers run at CPU-to-GPU ratios of roughly 1:4 to 1:8, and expects that to shift to between 1:1 and 1:2 as agents spread. The same piece cites a November 2025 paper, A CPU-Centric Perspective on Agentic AI, which found that tool processing on CPUs can account for up to 90.6% of total latency. TrendForce also relays an Arm estimate that data centers need about 30 million CPU cores per gigawatt today and could need 120 million in the agent era. Arm sells CPU designs, so that figure is a vendor projection, not a measurement.

That is the problem Helion targets: thousands of small, mostly idle-or-waiting tasks that must run in parallel, not one large calculation. The Register’s reported 256-core Venice parts make the density plausible, but they do not settle which workloads benefit. A rack tuned for thread count suits agents that fan out into many tool calls less well if the bottleneck is single-thread speed, and no independent benchmarks have been published.

The open approach matters as much as the silicon. Because the blueprint is meant to go to the Open Compute Project, vendors such as Dell, HPE and Supermicro can build around the same reference rather than each designing their own. Whether that yields interchangeable systems or lightly differentiated variants will depend on how much of the specification is published and when.

What’s next

AMD says partner configurations are planned for November 2026, and the specification is to be contributed to the Open Compute Project. The release is explicitly forward-looking, so timing and features could change. Independent figures on performance, power draw and pricing have not been published.