AMD published a rack blueprint today that treats AI agents as a CPU problem. The design, AMD Helion, holds up to 32,768 CPU cores and no GPUs at all. Inference stays on separate hardware. Partner builds start in November 2026, with wider availability in 2027.
An agent loop is mostly not the model
Inside an agent, the time goes into planning, tool calls, code execution in a sandbox, retries, output repair and stitching results together. Only the model call needs an accelerator. The rest is branchy, latency-sensitive scalar work, which is CPU territory. That side stops being a rounding error once a fleet runs thousands of concurrent agents.
AMD Helion gives the split a rack. The CPU-only version targets hyperscale operators that already run inference fleets and want dedicated capacity for agent execution, wired to separate inference hardware. A second version mixes in up to 16 AMD Instinct MI350P accelerators plus up to 102 Tbps of networking for enterprises that want both in one cabinet. Blade options run from a 256-core part at 400W to the same core count at 600W and a 96-core part at 5 GHz.
NVIDIA reached the same conclusion from the other side
NVIDIA added a Vera CPU rack design to its MGX 1.2 specification for tool calling, orchestration and sandboxed execution, beside its Rubin training racks. Two competitors, one day, the same missing tier. The GPU rack was never the whole data center, and the layer driving agents needs its own power budget.
No customer has been named yet
Every quote comes from a builder, not a customer. Cisco, Dell, HPE, Foxconn, Lenovo, Supermicro, QCT and Wiwynn. AMD named no deployment partner, saying only that Helios customers will likely deploy Helion too. Helios has orders, including a 1.2 billion dollar HPE deal with Vultr. Helion has a blueprint and a date, and its density figure rests on 256-core EPYC 9006 parts AMD has announced but is not shipping in volume.
Detail sits in the AMD announcement, The Register and the NVIDIA blog. Related reading. Money, not chips, is the AI bottleneck now and the GPU rental market cover where the spending goes.
Three questions for your own environment. How many CPU cores does your agent fleet consume outside inference, and did you provision any on purpose? Do your orchestration and inference tiers need to share a failure domain? What happens to latency if both spike at once?
Get the next one before it is old news
Independent analysis of cloud-native infrastructure, Kubernetes and data center economics. No vendor spin.