Arm closed FY26 with $1.49B in Q4 revenue and shipped the AGI CPU with $2B+ in committed customer demand across FY27 and FY28. The supply chain cannot keep up. That is the headline number, and it is not about Arm's quarterly financials. It is about Arm crossing a structural threshold: the company that licenses ISA and IP to chip designers has now become a chip designer competing with its own licensees.
The AGI CPU is purpose-built for agentic AI inference: many simultaneous short tasks at low latency, not one large batch job at high throughput. It is a different optimization target than a Graviton or an M-series core, and Arm is not offering it as a licensable design. It ships as Arm silicon, sold direct. The $2B+ demand figure comes from hyperscaler commitments; the supply constraint is TSMC capacity and packaging, not wafer yield or design readiness.
The business model shift is the part with structural consequences. Arm's royalty model prices compute at dollars per design win, averaged across billions of low-end cores. A hyperscaler running millions of AGI CPU instances at full utilization is worth a fundamentally different number on an ASP-per-unit basis than a royalty on a mid-range smartphone SoC. If Arm prices the AGI CPU as a product rather than as IP, every Arm licensee running their own data center silicon program is now a direct competitor, not just a downstream customer.
The next 18 months will clarify whether this is a one-time silicon product or the start of a compute platform business. AWS Graviton, Google Axion, and Microsoft Cobalt are all Arm-based and all designed in-house. If the AGI CPU is faster and more efficient for agentic inference than what those teams can build, hyperscalers will face a build-vs-buy decision on their own infrastructure for the first time. That conversation was not possible three years ago.