Arm CEO: AI Agents Driving Surge in CPU Demand
Miles Bennett
Arm CEO Rene Haas says each gigawatt of data-center capacity will eventually need 120 million CPU cores, up from roughly 30 million in 2025 — AI-agent workloads are shifting the competitive focus from GPUs to CPUs.
Why are CPU core counts suddenly not enough?
Haas puts a number on it: every 1 GW of data-center capacity will ultimately require 120 million CPU cores — about 4× the roughly 30 million needed in 2025.
This means → the CPU gap is not incremental growth; it is a step-change in scale.
The signal appeared roughly eighteen months ago. Customers designing Arm-based system-on-chips went from thinking 128 cores were sufficient to asking for 160, then 192. In plain terms = customer orders ran ahead of Arm's own forecasts.
Why can't GPUs handle AI-agent workloads?
AI agents — programs that act autonomously — do not just compute. They must orchestrate, schedule, and execute many virtual tasks simultaneously. That workload cannot simply be offloaded to GPUs.
This means → GPUs excel at parallel math, but the real bottleneck for agents is system scheduling, tool invocation, and concurrent workflow management — CPU territory.
This reflects a broader shift: AI is moving from chatbots to autonomous agents, and the binding constraint on an AI factory is no longer raw compute but orchestration capacity.
Arm is selling its own CPU now — does that threaten its customers?
Arm launched the AGI CPU in late March 2026 in San Francisco, stepping directly into the AI data-center CPU market.
Haas stressed this does not change Arm's core IP-licensing model. He pointed to Amazon: it will keep developing its in-house Graviton processors, and whether it also buys Arm's AGI CPU is Amazon's call.
In plain terms = companies without chip-design capability can buy an Arm CPU off the shelf; those with in-house silicon teams can stick with licensing — or do both. Arm adds a revenue line without closing its customers' self-design path.
What will future data centers look like?
Haas expects future data centers to run heterogeneous deployments: Arm CPU racks, Amazon Graviton, and Nvidia Vera coexisting in a single facility.
Nvidia's Vera CPU platform packs 256 Vera CPUs into one liquid-cooled rack; Arm's AGI CPU rack is built on the Open Compute Project (OCP) standard and fits naturally into mixed liquid- and air-cooled environments.
This means → the next phase of AI-infrastructure competition may pivot from "who has the most GPUs" to "who can support the largest fleet of AI agents" — and CPU orchestration will be the new core variable.
Content is for reference only, not financial advice.