AMD Unveils Helion: A 32,768-Core CPU Rack for AI Agents

AMD has published Helion, an open rack design for running AI agents rather than training models. The accelerator-free, CPU-only version fits up to 32,768 cores in a single rack. The first partner systems are due in November. AMD is selling Helion on efficiency: how many agents fit into each watt.

A rack for a different workload

For the past three years, the data-center industry has been built around training: dense GPU racks, liquid cooling, fast networks between accelerators. Agentic systems work differently. An agent rarely occupies a GPU continuously. It calls tools, hits APIs, runs code in a sandbox, holds context, waits for responses and coordinates other agents. That adds up to many parallel, relatively light processes, and a large share of them are bound by the CPU, not the accelerator.

According to the announcement, Helion is built on that observation. AMD is offering a standard rack design that partners can reproduce. The format is open, meaning the specifications are available to anyone who builds servers.

What we know so far

• Name and format: Helion, an open rack blueprint.
• Purpose: running and scaling agents, not training models.
• CPU-only configuration: up to 32,768 cores in a single rack.
• Timing: partner systems are expected in November 2026.
• Key metric: number of agents per watt.
• Not disclosed in the announcement: the exact CPU model, rack power consumption, networking, cooling and price.

That last item matters. For now we have a claim and one headline number, but no product spec sheet.

The arithmetic and the metric

The number 32,768 is 2 to the 15th power, and it looks like a convenient round configuration. To gauge the scale: at 256 cores per socket, that works out to 128 processors per rack. This is our own calculation for illustration, not AMD data. The actual layout may be different.

The "agent per watt" metric raises more questions. Training has established units: FLOPS, tokens per second, time to convergence. Agents have no standard. What counts as one agent: a simple loop with a couple of API calls, or a complex process with a browser and code execution? What workload, what latencies, what share of idle time? Without answers, AMD's figure is comparable only to itself. Independent measurements will not be possible until the hardware reaches customers.

Where this leaves the market

Helion shows that chipmakers are looking for a second source of demand beyond expensive accelerators. If agentic workloads really do shift to the CPU, the winners will be those with strong server processors and an ecosystem of assembly partners. An open blueprint lowers the barrier to entry for integrators and raises fewer vendor lock-in worries than a closed platform. But competitors also offer server CPUs, including Arm-based ones, and each will make its own efficiency case on its own benchmarks.

What teams building agents should do

Before waiting for November, measure your own workload. How much time does an agent actually spend on the GPU, and how much waiting for tools and running on the CPU? If the second share is large, paying for accelerators to do orchestration is a poor deal, and a CPU rack may turn out cheaper per completed task. Count the cost of a task, not the cost of a server.

When partners announce configurations, ask them for power consumption under your workload, not the spec-sheet figure. Separately, check software licenses and per-core pricing: with tens of thousands of cores per rack, they can eat up the energy savings. For now, Helion remains a claim, but the direction is clear: infrastructure for agents is starting to separate from infrastructure for training.