Arm is building its own chips for the first time in 35 years

What Happened

For 35 years, Arm made its money by designing the blueprints that other companies — Apple, Qualcomm, Samsung — used to build chips. Arm collected royalties. It never sold silicon. That arrangement ends now.

The AGI CPU is a 136-core processor built on TSMC's 3-nanometer manufacturing process, the same advanced node used in the latest consumer smartphones. Arm says it delivers twice the performance per watt of comparable x86 server chips. Up to 64 of these CPUs can fit in a single air-cooled rack, yielding roughly 8,700 cores. Volume production is expected in the second half of 2026. Meta is the launch customer; OpenAI, Cloudflare, and SAP are among the seven others already signed.

Why It Matters

The business logic is clear: AI workloads are driving massive demand for data center hardware, and Arm wants a larger share of that revenue than royalties alone can capture. The company spent $71 million and 18 months building a dedicated chip lab in Austin, Texas, growing a team there to over 1,000 people. Arm projects the move will add billions in annual revenue.

The strategic risk is harder to dismiss. Arm's chip-design customers — the same companies paying it royalties — now face a competitor who knows their architecture better than anyone. Qualcomm and Apple were not named as AGI CPU launch partners. Whether they stay comfortable licensing from a company now selling chips directly is the open question hanging over this announcement.

The chip's name signals intent as clearly as the hardware. Arm called it the AGI CPU — placing the product at the center of the infrastructure buildout for agentic AI, the category of AI systems that take actions autonomously. Whether that name reflects engineering reality or marketing ambition, it tells you where Arm intends to compete.

Key Takeaways

• Arm entering chip sales is a structural shift, not a product launch. It changes the company's relationship with its entire customer base.
• Eight AI companies — including Meta and OpenAI — have already committed. If the performance claims hold at scale, Arm's timing is good.