Arm's new chip is a $100 billion bet — and a quiet declaration of war on its own customers.
The company that built its business licensing instruction set architectures to every chipmaker on the planet shipped its first complete CPU this week. The Arm AGI CPU, announced at the company's annual developer conference, is a 136-core slab of TSMC 3nm silicon tuned specifically for the agentic AI workloads that are supposed to fill hyperscale data centers over the next decade. Meta is the lead partner. OpenAI asked for the same thing three years ago. And two of Arm's biggest historical customers — Apple and Qualcomm — were notably absent from the 50-plus companies endorsing the launch.
The product is real and the specs are serious: 300-watt thermal envelope, 6 gigabytes per second of memory bandwidth per core at sub-100-nanosecond latency, PCIe Gen6 with 96 lanes. According to Arm's launch announcement, the chip delivers more than twice the performance per rack compared to x86 CPUs, which at AI data center scale translates to what Arm claims is up to $10 billion in capital expenditure savings per gigawatt of capacity. The math isArm's to prove, but the physics are plausible.
The business logic is harder to argue with. Arm currently earns roughly $50 per chip in IP royalties. On the AGI CPU, it estimates $500 per chip in gross profit — a tenfold step change. Per EE Times, the company projects the overall Arm CPU market will reach around $100 billion by 2030, and the AGI CPU is its play to own a meaningful slice of the infrastructure that trains and runs the AI systems that every hyperscaler is racing to deploy. CEO Rene Haas told attendees at the event that the chip alone could generate $15 billion in annual revenue by 2031, lifting total company revenue to $25 billion with earnings per share of $9.
That is not licensing economics. That is semiconductor economics.
The tension is baked into the product's name. AGI — artificial general intelligence — is what every hyperscaler says it is building, and every hyperscaler now needs more CPU than current projections account for. Arm's own data shows data centers expected to require more than four times the current CPU capacity per gigawatt as agentic AI scales. That is a problem Arm's old customers — Intel, AMD, the custom silicon teams at Apple and Google — are all racing to solve. Arm is now trying to solve it alongside them, and for some of them, against them.
Apple designs its own chips for iPhones and Macs. Qualcomm does the same for Android devices and Windows laptops. Both rely on Arm's instruction set. Neither was on the endorsement list for a product that is explicitly designed to go into the same data centers their chips might otherwise populate. Qualcomm declined to comment. Apple did not respond to a request for comment. Nvidia, which competes with no one it does not also partner with, showed up anyway.
Intel's data center chief Kevork Kechichian was more direct. Speaking to The Register, he said Intel does not see much demand for Xeon 6 processors in agentic use cases — that the chip is most popular in networking applications like packet processing. Intel's counter-program is Clearwater Forest, its next-generation data center silicon, which Intel argues addresses the same workloads without requiring a new architectural category. Whether the market agrees with Intel's skepticism or Arm's framing will be settled by procurement decisions that are happening now, in private, at every major cloud provider.
The AGI CPU has one thing going for it that previous Arm server chips did not: a committed customer with scale. Meta has been building its own AI infrastructure — the MTIA accelerators that the AGI CPU is explicitly designed to complement — and has been asking Arm for finished silicon, not just IP, for three years. Per The Next Platform, Meta and OpenAI were the first to make that request. Meta's involvement is not a reference design. It is a co-development partnership with real deployment volume behind it.
Whether that is enough to overcome the fundamental conflict depends on how much of the data center CPU market Arm can capture before its licensees decide the conflict of interest is no longer theoretical. Apple and Qualcomm staying away from the launch is a signal. So is Nvidia showing up. The company that built its business on not competing with its customers just started.
† Consider adding: "[source-reported; not independently verified]" at end of sentence.
†† Consider adding: "[source-reported; not independently verified]" after "earnings per share of $9" or attributing more explicitly to company guidance/forecasts.
††† Consider adding: "[source-reported; not independently verified]" after the sentence.
