AMD Just Built a Rack for the Other Half of the AI Data Center: The Part That Runs the Agents
Most AI infrastructure headlines are about GPUs. On October 12, at the Open Compute Project Global Summit in San Jose, AMD announced something aimed at a different layer: Helion, an open rack design for the CPUs that actually run AI agents.
Here’s what AMD actually announced:
→ Helion is an open reference design for agentic AI, built on the OCP Open Rack V3 standard. AMD is contributing the specification to the OCP community instead of selling a locked-down system
→ The CPU-only configuration packs up to 32,768 Zen 6 cores into a single rack, using up to 128 of AMD’s 6th-gen EPYC “Venice” processors with 256 cores each
→ An enterprise configuration adds up to 16 Instinct MI350P GPUs for inference, rated at 74 petaflops of FP4 combined. The GPU nodes run independently, with no scale-up GPU fabric between them
→ The rack supports up to 102 Tbps of programmable networking, with liquid or air cooling options
→ A built-in token router can send an agent’s model calls to frontier models in the cloud or to open-source models running inside the rack
→ AMD splits the CPU nodes by agent type: dense 256-core servers for personal agents, balanced servers for enterprise agents, and high-frequency 96-core parts for technical agents
→ Timing: first partner-built configurations are targeted for November 2026, with broader availability in the first half of 2027. Supermicro is among the partners named
Why it matters: an AI agent spends much of its time running code, calling tools, and browsing, not generating tokens. Those are CPU jobs. As agent counts grow, the CPU side of the data center becomes its own infrastructure problem. In a Futurum analysis, AMD sized the server CPU market above $200 billion by 2030.
There are caveats. Analyst coverage notes that no agent-specific benchmarks have been published yet, and Helion is a blueprint, not a shipping product. AMD says it targets enterprise agentic workloads and inference, not frontier-scale training, so it isn’t a direct challenge to Nvidia’s GPU racks. But the open OCP approach lowers switching costs for customers wary of lock-in, and that’s a strategic angle on its own.
The bigger signal is that the agent era is starting to get its own hardware category, and the winners may be decided by who solves the unglamorous parts: orchestration, routing, and sandboxed execution at scale.
Do you think the next big infrastructure bottleneck for AI is GPUs, or everything around them?
#AMD #AgenticAI #AIInfrastructure #DataCenters #OpenComputeProject #TechNews
— 𝔖𝔞𝔫𝔡𝔢𝔢𝔭 ℜ𝔞𝔦𝔷𝔞