Arm ships an AGI CPU

Most of the chips inside an artificial intelligence server are not doing the flashy part. Graphics processors do the matrix math, but central processing units handle memory, networking, storage, security checks, and the constant traffic-cop work that keeps thousands of jobs moving. (arm.com) Arm has spent decades selling blueprints for that kind of processor instead of selling the finished chip. Apple, Amazon Web Services, and Qualcomm license Arm designs, then build their own silicon around them. (arm.com) Now Arm has crossed that line itself. On March 24, 2026, it announced the Arm AGI central processing unit, which it calls its first production silicon product and its first Arm-designed data-center central processing unit. (arm.com) The chip is aimed at what Arm calls agentic artificial intelligence, which means software agents that keep running, calling tools, and passing work back and forth instead of answering one prompt and stopping. Arm says those workloads need sustained performance across thousands of parallel tasks, not just short bursts. (arm.com) Arm says it built the chip with Meta as the lead partner. Meta said the new processor will join its in-house silicon portfolio, which already includes the Meta Training and Inference Accelerator family it uses for recommendation systems and generative artificial intelligence workloads. (about.fb.com 1) (about.fb.com 2) The technical pitch is density. Arm says a standard 36 kilowatt air-cooled rack can hold up to 8,160 cores of this design, because each chip is rated at 300 watts and the system was planned around fitting more useful compute into the same power and cooling envelope. (arm.com) The chip itself uses up to 136 Arm Neoverse V3 cores. Arm’s reference server brief also lists 6 gigabytes per second of memory bandwidth per core and memory latency below 100 nanoseconds, which are the plumbing numbers that decide how often a processor sits idle waiting for data. (arm.com) That is where the “2x per rack” claim comes from. Arm says the AGI system can deliver more than twice the performance per rack of comparable x86 deployments, which means the comparison is not one chip against one chip but one full cabinet of servers against another full cabinet of servers. (arm.com 1) (arm.com 2) Arm is also trying to remove the usual adoption delay by shipping a reference server at the same time. Its 1 open unit dual-node design pairs two AGI chips in a standards-based chassis, so original design manufacturers and cloud builders can start from a working template instead of designing a motherboard from scratch. (arm.com) The ecosystem list is almost as important as the chip. Arm says customers and manufacturers are already committed for production, and its launch page includes support quotes from Lenovo, Cloudflare, OpenAI, and Positron alongside Meta. (arm.com 1) (arm.com 2) That makes this less like a normal product launch and more like a change in Arm’s business model. The company still sells intellectual property cores and compute subsystems, but it is now also selling finished silicon into the data center, which puts it in more direct competition with the companies that used to be only its customers. (arm.com) If Arm’s numbers hold up in real deployments, the immediate effect is simple: more central processing unit work can stay inside the same power budget, the same cooling budget, and the same rack footprint. In a market where Meta, Microsoft, Amazon, and OpenAI are all constrained by megawatts as much as by chip supply, that is a very concrete thing to sell. (arm.com) (about.fb.com)