AMD puts 32,768 CPU cores in a rack built for AI agents

AMD has published Helion, an open rack blueprint for running AI agents rather than training models, with a CPU-only version holding up to 32,768 cores. Partner systems are due in November, and the pitch is how many agents fit in each watt.

AMD has published a rack blueprint for running AI agents rather than training models. It is called Helion, and partner systems are due in November.

The design appeared at the Open Compute Project summit in San Jose, The Register reported. It comes in a CPU-only version for hyperscale operators and a mixed CPU and GPU version for enterprises.

The CPU-only rack tops out at 32,768 cores. That is 128 sockets of AMD’s 256-core Venice chip on Zen 6, in 400W dense or 600W balanced blades.

The mixed rack is thinner on graphics than it sounds. It holds at most 16 Instinct MI350P accelerators, in two eight-GPU servers that AMD says run independently, with no scale-up fabric.

So this is not a frontier machine. AMD’s Helios rack carries 72 of the newer MI455X parts and 31 terabytes of HBM4, and remains the product for training.

Helion’s most interesting part is barely described. AMD says the enterprise rack carries a token router that sends agents’ model calls either to frontier models in the cloud or to open-source models running in the rack.

That detail appears only in AMD’s own release. Dan McNamara, who runs compute and enterprise AI at AMD, said the point was to let partners make efficient use of available power.

Power is the whole pitch. Matt Kimball at Moor Insights and Strategy said in a field note that the job is packing as many agents as possible into every watt.

He named sovereign AI operators among them, which puts Europe in the frame. Denmark’s grid operator paused new connections in March, with about 60GW of projects queued against peak demand near 7GW.

The constraint is not only Danish. Brussels adopted a Cloud and AI Development Act in June, aiming to at least triple the bloc’s data centre capacity within seven years.

Operators say the same thing. In a survey of 300 data centre decision-makers in the UK, Ireland and the Nordics, power availability came out as the leading obstacle to delivering sovereign capacity.

Helion is less finished than it reads. AMD has published no pricing, plans to contribute the specification to OCP rather than having done so, and named no European system builder among its partners.

Europe is still buying. The bloc has opened bidding for up to seven gigafactories at about €30B, each meant to hold 100,000 chips from suppliers including AMD.