Arm Builds Its First Chip in 35 Years — and Meta Is First in Line

Arm has never made a chip. For 35 years, the U.K. company designed processor architectures and licensed them to Apple, Nvidia, Amazon, and nearly every other major semiconductor company on earth. That model ends now. At an event in San Francisco on March 24, CEO Rene Haas unveiled the Arm AGI CPU — the company’s first piece of physical silicon.

The specs are built for AI inference at scale. The AGI CPU packs 136 Arm Neoverse V3 cores, runs at 3.7 GHz, and is manufactured on TSMC’s 3-nanometer node. Up to 64 chips fit in a single air-cooled rack, delivering more than 8,000 cores at 36 kW. Arm claims that’s twice the performance-per-watt of an equivalent x86 rack. Liquid-cooled configurations scale further — past 45,000 cores per rack.

Meta is the launch customer and was a co-development partner throughout the design process. The company has committed to multiple generations of the chip and has earmarked up to $135 billion in capital expenditures for 2026. Seven other customers are signed: OpenAI, Cloudflare, SAP, and others not yet named publicly.

The strategic logic is straightforward. Agentic AI runs on CPUs. GPU clusters handle training and some inference, but coordinating multiple agents across long tasks — the kind of workload increasingly common in production — demands general-purpose compute at high density. Haas said he expects CPU demand to grow fourfold around agentic workloads. Arm’s existing licensees, who are already building chips on Arm’s architecture, now have a competitor from the IP source itself.

The revenue projections are aggressive. Haas told investors the AGI CPU will generate $15 billion annually by 2031, contributing to $25 billion in total annual revenue and $9 earnings per share. For context, Arm generated roughly $4 billion in revenue in 2025. Arm’s stock jumped 6% the evening of the announcement and another 16% the following day after the 2031 guidance landed.

Building silicon required real infrastructure. Arm spent $71 million over 18 months constructing three new lab facilities at its Austin, Texas campus, growing the team there to over 1,000 engineers. Manufacturing is at TSMC in Taiwan, in line with every other fabless AI chipmaker.

The pricing strategy is deliberate: Arm’s CFO confirmed a roughly 50% gross profit margin, with pricing positioned as “competitively attractive” for companies that can’t afford to build custom silicon at the scale Meta and Google can. That’s a specific market — the tier just below hyperscaler — and it’s large.

Arm’s move changes the competitive map for AI infrastructure. Licensees building Arm-based chips now compete with the architecture’s creator. That tension will define the company’s next decade.