Arm Made Its First Chip in 35 Years. Meta Bought It.

For 35 years, Arm Holdings played a particular role in the technology world: it drew the blueprints. Apple, NVIDIA, Qualcomm, Amazon — they all built chips using Arm's instruction sets, paid Arm a royalty, and shipped the hardware. Arm stayed out of the silicon business entirely. That was the deal, and it worked.

On Tuesday, Arm CEO Rene Haas walked onto a stage in San Francisco and tore the deal up.

At the company's Arm Everywhere conference, Haas unveiled the Arm AGI CPU — a 136-core, 3nm data center processor that Arm designed, built, and intends to sell as finished silicon. The first customer is Meta. And the company projects the chip will generate $15 billion in revenue by 2031.

The Company That Always Said No

Arm was founded in 1990 as a joint venture between Acorn Computers, Apple, and VLSI Technology. From the start, the model was licensing: Arm would design processor architectures and sell the right to use them. It would never own a fab, never compete with its own customers, never ship a chip with its name on the box.

That restraint built a dynasty. Today, Arm architecture runs in roughly 99% of the world's smartphones. It powers Apple Silicon, Amazon's Graviton, NVIDIA's Grace CPU. Arm collected royalties from all of them. The company's IPO in September 2023 valued it at $54.5 billion.

The licensing model was elegant precisely because it was hands-off. Arm's customers were also its partners. It had no incentive — or so the theory went — to ever compete with the people writing it checks.

Then the hyperscalers started building their own chips. Amazon made Graviton. Google made Axion. Microsoft made Cobalt. These were all Arm-based designs, all displacing the x86 processors from Intel and AMD in cloud data centers. Arm was collecting its royalty cut, but watching billions of dollars in silicon revenue flow to its customers instead.

The AI boom changed the math entirely.

What Meta Asked For in 2023

Mohamed Awad, Arm's Executive Vice President of Cloud AI, has said the project began three years ago at Meta's request. Hyperscalers wanted more than blueprints. They wanted Arm to hand them a production-ready chip, not just the IP to design one themselves.

Arm took the ask seriously. The company spent $71 million building three new chip lab rooms at its Austin, Texas campus. A team that started small grew to over 1,000 engineers.

Meta software engineer Paul Saab has been involved in the project since its start in 2023. Meta's head of infrastructure, Santosh Janardhan, was direct about why. "Delivering AI experiences at global scale demands a robust and adaptable portfolio of custom silicon solutions, purpose-built to accelerate AI workloads and optimize performance across Meta's platforms," he said. "We worked alongside Arm to develop the Arm AGI CPU to deploy an efficient compute platform that significantly improves our data center performance density and supports a multi-generation roadmap for our evolving AI systems."

The co-development went deep enough that Meta will deploy the chip alongside its own custom MTIA AI accelerators. Meta also plans to open-source board and rack designs through the Open Compute Project — a move that could accelerate adoption across the rest of the industry.

What the Chip Actually Does

The AGI CPU is not a GPU. It does not train AI models. Arm is positioning it specifically for AI inference — and more precisely, for the emerging class of agentic AI workloads.

Traditional chatbots were relatively simple: a user sends a message, a model generates tokens, the exchange ends. Agentic AI systems are different. They run multi-step tasks continuously, orchestrate other tools and models, execute code, and loop back on themselves. These workflows are CPU-intensive in ways that GPUs were not built for.

Awad described a structural shift in data center economics: agentic AI requires roughly 120 million CPU cores per gigawatt of power, compared to 30 million for traditional configurations. "We think that the CPU is going to be fundamental to ultimately achieving AGI," he said.

The chip itself reflects those priorities. Its two 68-core chiplets are connected by a die-to-die interconnect. Each core runs at up to 3.7 GHz with no simultaneous multithreading — a clean, high-frequency architecture designed for the kind of serial reasoning that agents require. Memory latency sits below 100 nanoseconds. The Intel Xeon 6980P, a 128-core competitor running at 500 watts, burns 3.9 watts per core. The AGI CPU runs at 2.2 watts per core.

In rack configuration, Arm claims more than 2x the performance per rack compared to current x86 platforms.

The Timeline From Idea to Silicon

$15 Billion by 2031, and Haas Thinks That's Conservative

Haas put a very specific number on the opportunity during Tuesday's event. The AGI CPU should generate $15 billion in revenue by 2031 — against a total addressable market of $100 billion for data center CPUs in agentic AI by 2030.

Then he pushed further. "We may be under-calling that number," Haas said. "I think the demand is higher than we think it is."

Arm's full-company revenue target for 2031 is $25 billion, with earnings per share of $9.

Wall Street responded immediately. Arm shares surged 16% on Wednesday. Evercore ISI analyst Mark Lipacis raised his price target from $170 to $227 — a 34% jump — and described Arm as "a key beneficiary of agentic AI." Barclays analyst Tom O'Malley lifted his target from $165 to $200, noting the chip "plays into Arm's strength in energy efficiency." Raymond James upgraded the stock to Outperform with a $166 target.

Citi analysts did not mince words. The AGI CPU, they said, represents "the most significant shift in the company's history."

Among 25 analysts tracked, 20 hold Buy ratings on Arm. The average price target sits at $170.86 — before Tuesday's announcements are fully reflected in consensus models.

The Name Carries a Claim

The branding choice was deliberate and, to some, provocative.

AGI — artificial general intelligence — is the term the research community uses for hypothetical AI that can perform any intellectual task a human can. It has been the stated long-term goal of OpenAI, DeepMind, and Anthropic. It has never been achieved.

Arm is calling its processor the AGI CPU.

Awad addressed the name head-on: "We think that the CPU is going to be fundamental to ultimately achieving AGI." The company is not claiming the chip produces AGI. It is claiming the chip is part of the infrastructure that will enable it.

The Register captured the tension with its subheadline: "Turns out artificial general intelligence was a CPU this whole time." The framing was sardonic, but the underlying point was fair. Arm has attached the most loaded term in AI to a server processor at exactly the moment the term has maximum marketing power.

Some observers warned that attaching "AGI" to commercial products risks eroding the meaning of the term entirely — comparing it to the drift of words like "meme," which now means something far narrower than its coinage. The chip's actual function is AI inference, not general intelligence. But Arm's bet is that the distinction will matter less than the association.

Why Arm's Old Customers May Have Reason to Worry

The conflict-of-interest question was never hard to see.

Arm built its business on a promise of neutrality: it would design architectures and license them equally, never favoring one customer over another, never competing in the markets its customers operated in. The AGI CPU breaks that promise structurally.

Amazon, Google, and Microsoft all build their own Arm-based data center chips. They are now, in the data center CPU segment, competing with the company that sells them their instruction sets.

The Register noted that Arm will need to manage accusations of giving its own chip preferential access to new architecture features before they are made available to licensees. Arm has not publicly addressed how it intends to handle that tension.

There is also the question of what comes next. Haas teased additional mystery products at the Arm Everywhere event, hinting at a $1 trillion total addressable market across compute categories. The AGI CPU may be the first product in what becomes a much broader silicon portfolio.

For companies that have spent years building their own chips on Arm's blueprints, that expansion is worth watching carefully.

The Bottom Line

Arm spent 35 years as the most important company in chip design that nobody sold chips. That era ended Tuesday in San Francisco.

The AGI CPU is technically impressive — better power efficiency than Intel, purpose-built for the workloads hyperscalers are racing to deploy, backed by Meta's scale and credibility. The $15 billion revenue projection is ambitious, but the demand signal from OpenAI, Cloudflare, and SAP alongside Meta suggests real customer conviction.

The harder question is what Arm becomes next. The licensing business that generated Arm's IPO valuation was built on trust — the trust of customers who believed Arm would never compete with them. That model is now in tension with the silicon business Arm just entered. Both cannot grow indefinitely without one eventually constraining the other.

Rene Haas has already said the demand for the AGI CPU may be higher than Arm's own forecasts. If he is right, the company will be pulled further into the market its customers built on its blueprints.

That is a very different company from the one that existed last week.

Sources

• Arm expands compute platform to silicon products in historic company first — Arm Newsroom (March 24, 2026)
• Announcing Arm AGI CPU: The silicon foundation for the agentic AI cloud era — Arm Newsroom (March 24, 2026)
• Arm releases first in-house chip, with Meta as debut customer — CNBC (March 24, 2026)
• Arm jumps 16% as company expects revenue windfall from new chip — CNBC (March 25, 2026)
• Arm stock pops 6% as CEO Haas issues $15 billion revenue expectation for new chip — CNBC (March 24, 2026)
• Arm is releasing the first in-house chip in its 35-year history — TechCrunch (March 24, 2026)
• Arm rolls its own 136-core AGI CPU to chase AI hype train — The Register (March 24, 2026)
• Arm comes full circle with homegrown, AI-tuned server CPU — The Next Platform (March 25, 2026)
• Meta Partners With Arm to Develop New Class of Data Center Silicon — Meta Newsroom (March 24, 2026)
• Arm moves beyond IP with AGI CPU silicon — Tom's Hardware (March 24, 2026)
• Arm Holdings (ARM) Soars 16% on Debut of AGI CPU With $15B Revenue Goal — Parameter (March 25, 2026)