CPU demand resurges in AI infrastructure due to requirements for control logic, networking, and pre-post-processing alongside GPUs.
Rene Haas, chief executive of Arm Holdings, unveiled the chip company's latest CPU design at a San Francisco waterfront pavilion Tuesday, marking a significant moment in the semiconductor industry's response to artificial intelligence. "People thought CPUs were dead," Haas told the crowd, but as AI rapidly develops, "You need more and more CPUs. Lots of them." After years of being an afterthought in AI computing, central processing units—the core engines of most computers from laptops to data-center servers—have become essential to the infrastructure supporting advanced AI applications.
Arm is undertaking a dramatic business transformation, shifting from its traditional model of licensing CPU architecture to competitors like Nvidia and Qualcomm to become a chip designer and seller itself. The company's new processor is marketed as the "world's most efficient agentic CPU," with twice the efficiency of equivalent competitor processors. Meta Platforms has signed on as the lead partner and first major customer, with other customers including OpenAI, Cerebras, SAP, Cloudflare, SK Telecom and Rebellions. The chips will be fabricated by Taiwan Semiconductor Manufacturing Co., with first orders shipping in the months ahead.
The expansion reflects Arm's confidence in the CPU market's growth trajectory. Haas stated the company expects annual revenue to quintuple to around $25 billion over the next half-decade, significantly exceeding analyst predictions from FactSet. Arm projects the total addressable market for data-center CPUs will grow to about $100 billion a year over the next half-decade. Following the announcements, Arm's shares rose more than 16% on Wednesday.
The surge in CPU demand stems directly from the explosive growth of agentic AI, driven by coding tools such as Claude Code and OpenAI's Codex, as well as platforms like OpenClaw, which enables users to create virtual personal assistants capable of real-world tasks. Thomas Sohmers, chief technology officer and co-founder of Positron AI, explained the computational requirements: "If people have multiple OpenClaws running all day, you need a lot of dedicated CPUs. A few months ago, people just weren't running this many agents. This is truly new computational demand." Haas added that agents operating autonomously can generate dramatically more computational work than humans: "When you go to an agentic world, where a human can generate 100 tokens in a minute or two, an agent can do 15 times that number. That's all CPU work."
Industry leaders confirmed the scale of this shift. Santosh Janardhan, Meta's head of infrastructure, described his company's goal to "deliver personal superintelligence to billions of people," adding that success requires "silicon. A lot of silicon" and noting "There is absolutely no sign of this slowing down." Kevin Weil, OpenAI's vice president for science, emphasized that powering AI has become a matter of "system performance" rather than relying on individual processors, making CPUs crucial. "I need more compute is the coin of the realm," Weil said. "We have more demand from customers, more things we want to do, than frankly the industry can keep up with."