AMD designs its Helion rack for running AI agents on up to 32,768 CPU cores

Whoever has to find room for agents on Monday will look first for the number the pitch promises. Matt Kimball of Moor Insights and Strategy said in a field note that the job is packing as many agents as possible into every watt ^[12]. The only wattage in the coverage is the blade rating, 400W dense or 600W balanced ^[2]. AMD did not publish an agents-per-watt figure.

The mixed rack is the harder one to size. Helios, which AMD keeps for training, has 72 accelerators ^[10], so the mixed Helion rack has about 22% as many ^[15], though the parts differ. Because AMD says the two eight-GPU servers run independently ^[4], we'd expect eight GPUs to be the practical ceiling for any one model instance.

AMD's own description of the work is wider than model serving. "Agentic AI is not a single workload. It spans personal agents, enterprise applications, code execution and complex workflows, each with different compute requirements," said Dan McNamara, who runs compute and enterprise AI at AMD ^[11]. DatacenterDynamics reports that users can combine CPU and Instinct GPU nodes to match agent execution and model inference ^[7]. That puts the weight on the token router, which decides whether a model call lands on the rack's GPUs or in a cloud. The Next Web notes the router appears only in AMD's own release ^[5].

The Next Web calls Helion an open rack blueprint, but AMD only plans to contribute the specification to OCP ^[8]. The configurations are still under development with partners ^[16]. The Next Web also notes that no European system builder is on AMD's partner list ^[13]. Power is the constraint European operators report: in a survey of 300 data centre decision-makers in the UK, Ireland and the Nordics, cited by The Next Web, power availability was the leading obstacle to delivering sovereign capacity ^[14].

We'd put two questions to each agent workload before it goes on a rack plan. The first is whether its model calls stay in the rack or go to a hosted model. The second is whether the site is limited by power or by budget.

- Hosted models, tight power: the CPU-only rack is the candidate. Ask the partners for agents per kilowatt on your own task mix. - Hosted models, ample power: on the reporting so far, Helion is one more CPU server. Price decides. - In-rack models, tight power: the mixed rack fits, if the model fits in eight GPUs. - In-rack models, ample power: the 16-GPU cap and the missing scale-up fabric are what to test against a GPU rack.