AI's Next Battle: Arm CEO Says CPU Demand Is Off the Charts

For the past two years, GPUs—especially NVIDIA’s—have dominated AI infrastructure discussions, from large‑model training to inference clusters, with the industry equating compute power to GPU availability.

Arm CEO Rene Haas now argues that this narrative is changing: the market for advanced AI CPUs is "off the charts." He notes that, as of early 2024, each 1 GW of data‑center capacity will require about 120 million CPU cores, up from 30 million the previous year.

Haas explains that the driver is the emergence of Agentic AI. In Agent workflows, more cores enable a dedicated Agent to run a virtual task, making high‑core‑count CPUs essential. Arm’s research identified agents as the key factor propelling CPU demand, because these workloads cannot be efficiently handled by GPUs alone.

Reflecting this shift, Arm is moving from a pure licensing model to offering complete compute platforms. Its newly announced AGI CPU, built on TSMC’s 3 nm process, packs up to 136 Neoverse V3 cores, 300 W TDP, DDR5‑8800 memory, 96 PCIe Gen6 lanes, and CXL 3.0 support, and claims more than double the rack‑level performance of comparable x86 systems.

Haas also acknowledges growing competition: NVIDIA unveiled the Vera CPU for agents at GTC 2026, Intel introduced Xeon 6+ (based on its 18A process) at Computex 2026, and AMD is emphasizing rack‑level planning for Agentic AI. All three aim to address the same shift toward CPU‑centric AI workloads.

The interview highlights a broader industry consensus that AI infrastructure can no longer be understood solely through GPUs. As AI evolves from chatbots to agents, CPUs become the core for scheduling, execution, and orchestration, making the ability to support more agents the next competitive frontier.

Arm maintains its licensing business while offering the AGI CPU to customers who lack chip‑design capabilities, positioning the move as a massive market expansion opportunity.