AMD announced Helion on October 12, 2026, an open rack design for running AI agents. It can pack up to 32,768 CPU cores in one rack. The design separates the agent work—database queries, code execution, waiting—from the expensive GPU inference that runs the AI model itself.
AI agents spend most of their time not talking to the model. They check your calendar, search your files, run calculations, and wait for APIs to respond. All of that runs on CPUs. Helion lets you buy exactly the compute you need for those tasks without paying for GPUs that sit idle. Partners including Dell, Cisco, and Supermicro plan to ship configurations in November 2026. AMD has not announced pricing.
What you can do with this: if you run AI tools that work for hours, not seconds, ask your hosting provider if they're adding Helion configurations. The cost difference shows up when an agent runs 200 steps to book a supplier meeting versus answering one question. The server that runs those 200 steps used to need GPU headroom it rarely touched.
What to keep in your hands: check what November pricing actually lands at before you commit to a new host. The rack supports up to 32,768 cores, but you're buying what the vendor configures, not building it yourself. Ask which AI models the vendor routes to, whether you can bring your own, and whether the contract locks you to their model list or lets you switch when a cheaper one works.