Arm is becoming the chipmaker it never wanted to be

ARM HOLDINGS HAS spent 35 years perfecting one of the most lucrative business models in technology: designing processors, licensing the blueprints, and collecting royalties on every chip made with its architecture. More than 350 billion such chips have shipped — inside iPhones, Nvidia servers, Amazon data centers, Tesla robots, and roughly everything else with a transistor. The model has produced gross margins of 98%, a figure that would make a SaaS executive weep with envy. On Tuesday in San Francisco, CEO Rene Haas held up a piece of silicon and announced, in effect, that the old model is no longer enough. Arm is now a chipmaker.

The new product is called the Arm AGI CPU — a nod to artificial general intelligence that is more aspirational branding than technical descriptor. Built on TSMC's 3-nanometer process with up to 136 Neoverse V3 cores, it is designed to orchestrate fleets of AI agents in hyperscale data centers: the autonomous software programs that book flights, write code, and manage infrastructure with minimal human oversight. Meta is the lead partner and co-development customer. OpenAI, Cloudflare, SAP, Cerebras, and SK Telecom have also committed to buying it. Haas projected the chip would generate $15 billion in annual revenue by 2031, lifting total company revenue to $25 billion and earnings per share to $9 — roughly six times the $4 billion Arm brought in last fiscal year. The stock popped 6% in after-hours trading. Creative Strategies forecasts the data center CPU market will grow from $25 billion this year to $60 billion by 2030; factor in CPUs purpose-built for agentic AI and that estimate approaches $100 billion.

The pitch, in short, is that agentic AI is about to create a new category of compute demand so large that even a fraction of it would transform Arm's revenue profile. Haas predicted a fourfold increase in CPU demand from agentic workloads alone, then added — with the kind of understatement that doubles as salesmanship — that the company may be under-calling that number.

Yet the strategic elegance of the move conceals a structural tension that Arm will have to manage for years. The company has long been called the Switzerland of semiconductors — a neutral platform whose intellectual property underpins the custom chips of Apple, Nvidia, Amazon, Google, and Qualcomm alike. That neutrality was the source of its power. Every hyperscaler could build on Arm precisely because Arm never competed with them for silicon revenue. Now it does.

Arm insists there is no friction. Haas said the company consulted customers before entering the market and received no pushback. Jensen Huang, James Hamilton of Amazon, and Google's Amin Vahdat all appeared in pre-recorded testimonials praising the AGI CPU — albeit without committing to purchase it. The diplomatic choreography was careful, bordering on theatrical. Meta's head of infrastructure, Santosh Janardhan, took the stage in person to endorse the chip, citing his company's 3.5 billion daily users and the rapid expansion of its AI clusters from 128 GPUs in 2023 to tens of thousands today. OpenAI's Kevin Weil appeared too, noting that the most common refrain inside his company is a simple one: more compute.

But as Ben Bajarin of Creative Strategies notes, the relationship dynamics shift as Arm's ambitions expand. The AGI CPU is currently a narrow product — a relatively small-core-count chip tailored to agentic AI workloads, designed for companies that cannot afford to design custom silicon themselves. Mohamed Awad, who leads Arm's cloud AI business, framed it explicitly as expanding the market rather than cannibalizing it. Over time, however, Arm plans to release new chip designs at 12- to 18-month intervals. If those designs broaden into general-purpose data center CPUs, the polite coexistence becomes real competition — not just with AMD and Intel, but with Nvidia's Grace and Vera CPUs, Amazon's Graviton, Google's Axion, and Microsoft's Cobalt, all of which are built on Arm's own architecture. BNP Paribas analysts have already flagged the question of how Arm will manage the inherent conflict of selling silicon to customers who also license its IP.

The more immediate risk is financial. Arm's IP licensing business is among the highest-margin operations in all of technology — 98% gross margins with minimal capital expenditure. You design a thing once, license it a thousand times, and collect a toll on every chip that ships. Manufacturing silicon is an entirely different beast. Haas acknowledged the AGI CPU would carry operating margins above 30%, which sounds respectable until you compare it to the 96-point spread Arm currently enjoys. Every dollar of chip revenue that displaces a dollar of royalty revenue compresses the blended margin. Wall Street, which has valued Arm at a $141 billion market cap on the strength of that near-perfect profitability, will be watching the mix closely.

Arm spent $71 million and about 18 months building three new lab rooms at its Austin campus, where a team that has grown to over 1,000 engineers puts chips through multiple rounds of post-fabrication testing. The company manufactures entirely at TSMC's plants in Taiwan for now, though Awad hinted at eventual U.S. production once TSMC's Arizona fabs come online. This is a permanent, capital-intensive commitment — not an experiment. Analysts at HSBC described the launch as transformational; others flagged that the stock's 25% rally this year already prices in considerable optimism.

The bet is that the total addressable market for Arm-branded silicon dwarfs what the company could ever collect in royalties alone. That arithmetic may well be correct. Arm currently earns a royalty of roughly 2-5% on each data center chip built with its designs; selling the chip directly captures the full revenue. A single Arm-based CPU — which chip analyst Patrick Moorhead estimates will cost in the thousands of dollars — generates orders of magnitude more revenue per unit than a royalty payment. If agentic AI drives the kind of CPU demand Haas is projecting, even a modest market share could dwarf the royalty business entirely.

The competitive context is formidable but also favorable. Intel's data center CPU business hit a 13-year revenue low in recent quarters. AMD, under Lisa Su, has surged to roughly 40% server market share — up from near zero in 2018 — but remains locked in an x86 architecture that Arm's efficiency claims directly challenge. Arm says its AGI CPU delivers twice the performance-per-watt of comparable x86 chips, a claim that matters enormously when data center operators like Meta are building five-gigawatt campuses where every watt is, as one Meta executive put it, "a very scarce resource." The company also noted that the chip does not fall under current U.S. export controls, leaving the door open for eventual sales into China — though no Chinese customers have been announced.

For three decades, Arm profited by being indispensable and invisible — the architecture inside everything, the name on nothing. The AGI CPU puts the name on the box. If Haas is right about $15 billion in chip revenue, Arm will have reinvented itself as one of the most formidable semiconductor companies on earth. If the margins compress and the customers bristle, it may look back fondly on the days when 98% gross margins and quiet royalty checks were enough. ■

For more, join 75,000 subscribers getting tech's favorite brief here