Arm just did something it spent decades avoiding. The company that quietly powered most of the world’s chips from behind the scenes has now stepped out and built its own processor. Instead of only selling designs and collecting royalties, Arm is now putting its own silicon into the market, starting with a new 3-nanometer AGI CPU aimed straight at data centers. That puts Arm into direct competition with some of the same companies that built products on its architecture. Chief executive Rene Haas unveiled the new in-house processor on Tuesday at an event in San Francisco. The company calls it the AGI CPU, a data center chip aimed at AI workloads. Arm is no longer only selling the foundation. It is now selling the chip itself. Meta joins Arm and expands its AI infrastructure Meta is the first customer. The company is building multiple gigawatts of AI data centers and plans to spend as much as $135 billion on capital expenditures this year. In February, Meta also secured a large supply of chips from Nvidia and Advanced Micro Devices. The new Arm deal gives it one more processor source as its AI buildout keeps expanding. Paul Saab, a Meta software engineer who has worked on the Arm chip project since 2023, told reporters that, “In today’s world, you really only have a couple of players.” Paul also said the deal “allows a lot more flexibility in our software stack and in our supply chain.” The companies did not disclose the terms. CPU demand rises as agentic AI strains compute The launch comes as CPUs are getting more attention in AI. Nvidia, which leads the market for AI graphics processors, had before said that CPUs are “becoming the bottleneck” as agentic AI changes compute needs. Futurum Group called it a “quiet supply crisis” and predicted CPU market growth could move ahead of GPU growth by 2028. The split between the two chip types is simple. GPUs are useful for training and running AI models because thousands of cores can do many operations at once. CPUs handle a smaller number of stronger cores built for general tasks that run in sequence. Agentic AI needs a lot of that general compute work because huge amounts of data move across multiple agents. At Nvidia’s GTC conference last week, chief executive Jensen Huang revealed an entire rack filled only with Vera CPUs. At the Arm event on Tuesday, Huang appeared in a recorded statement congratulating the company on the launch. Leaders from Google, Amazon, Microsoft, Oracle, Broadcom, Micron, Samsung, SK Hynix, and Marvell also appeared. Arm told reporters that around 50 partners had shown support before the launch. Mohamed Awad, Arm’s head of cloud AI, said, “It’s a $1 trillion market, and what we’re seeing over and over again is actually our partners coming out and understanding and realizing this is actually great for the industry.” Arm expands testing and relies on TSMC To reach this point, Arm spent $71 million and about 18 months building three new lab rooms at its campus in Austin, Texas. A team that once was small has grown to more than 1,000 people. Engineers use those labs to bring up the chips after they leave the factory and put them through repeated rounds of testing. Like most fabless AI chip companies, Arm is using Taiwan Semiconductor Manufacturing Company to manufacture the processor.The AGI CPU is made on TSMC’s 3-nanometer node, and all production is in Taiwan for now. TSMC is preparing a 3nm fab in Arizona. Awad said Arm would “love to manufacture here,” but said the final choice depends on what customers want. Arm is still best known for powering mobile chips in almost every smartphone.It entered data center chips in 2018 with Neoverse.Amazon pushed that platform into the mainstream with Graviton, and Google and Microsoft also base AI chips on Arm designs. Moorhead said, “If Arm didn’t exist, then all of those companies who have their own processors wouldn’t be able to create their own.” Most server chips still use x86 designs from Intel and AMD, which Moorhead called “tried and true.” Awad said the team “ruthlessly optimized” the AGI CPU for artificial general intelligence, which is why it carries that name.One air-cooled rack can hold up to 64 CPUs, or about 8,700 cores.He said, “You can get two times the performance-per-watt than you can from an x86 rack.” Paul said wattage is “a very scarce resource,” adding that better performance per watt leaves more power for other parts of the data center. The smartest crypto minds already read our newsletter. Want in? Join them .
