AMD's 32,768-Core AI Rack Ships in November, Minus Power Specs

What exactly is in AMD's Helion Agentic AI Rack?

It is a modular rack built to run AI agents, not to train giant models. The headline configuration combines up to 32,768 Zen 6 CPU cores with 16 Instinct MI350P accelerators and 74 petaflops of FP4 performance, tied together with up to 102 Tbps of networking.

The numbers are big enough to make you read them twice. AMD's Helion Agentic AI Rack is a rack-scale system aimed squarely at the agentic AI boom, where software agents run tasks continuously rather than waiting for a human to type a prompt.

The architecture splits work across three node types. CPU sandbox nodes handle agent execution, Instinct MI350P accelerators take care of inference, and general-purpose servers carry databases, storage and technical computing. An integrated token router directs model requests across frontier and open-source models, and the whole thing is built on Open Compute Project ORv3 44OU hardware with programmable networking.

• Up to 32,768 Zen 6 x86 CPU cores per rack
• Up to 16 Instinct MI350P accelerators
• 74 petaflops of FP4 compute
• Up to 102 Tbps networking bandwidth
• Integrated token router for agent model requests

Three racks for three kinds of agents

AMD is not selling one rack to everyone. It groups agent workloads into personal assistants, enterprise applications and technical computing, and tunes core density, clock speed and power draw for each.

The personal-agent configuration leans on 256-core Venice processors running at 400W, prioritising density and energy efficiency over raw speed. That makes sense: a personal assistant agent handling calendar, email and search tasks needs lots of small jobs done cheaply, not a single monster thread.

The enterprise version uses the same 256-core Venice chips but at 600W, combining CPUs and accelerators in one integrated rack. Technical computing goes the other way, with 96-core high-frequency Venice processors for workloads where single-thread speed matters more than core count. AMD also offers a CPU-only hyperscale configuration.

That flexibility is the pitch. Operators can mix core density, performance and power consumption to match what their customers actually run. The catch is that a rack full of mixed server types is harder to plan for than a row of identical boxes.

Why your data center may need a rethink

Most facilities were built for standard cloud servers, not racks pulling enormous power through dense CPU and GPU configurations. AMD's silence on power and cooling is the part operators will notice first.

A rack that can hold 32,768 cores and 16 accelerators is not something you slide into an empty space and forget. Power delivery, cooling capacity and network infrastructure all have to line up, and AMD has not published complete facility upgrade costs or cooling requirements.

The networking side adds another wrinkle. Supporting 102 Tbps and routing model requests across multiple AI systems may require compatible switches and software that existing facilities do not have. Data centers that cannot support the chosen mix of high-core-count and high-frequency servers will need to modify their infrastructure.

This is not unique to AMD. Nvidia's rack-scale systems have raised the same questions about power density and liquid cooling. But AMD is asking operators to commit to a new class of hardware without a full spec sheet, which makes capacity planning a guessing game until November.

Availability and pricing

AMD says the Helion Agentic AI Rack will be available in November 2026. Pricing has not been announced, and neither have the power and cooling figures that would let buyers estimate total cost of ownership. Expect pricing to be quoted per configuration rather than as a single list price.

For US, UK and EU buyers, the practical question is whether existing colocation contracts can support the power density. Operators in dense metro markets may find that the rack itself is the easy part and the facility upgrade is the real budget line.

AMD has built a convincing case that agentic AI needs its own hardware shape, and the Helion rack is the most concrete answer yet. The missing power and cooling numbers will decide whether data centers see it as an upgrade path or a renovation bill.