Sponsored by edgeful

For over three decades, Arm has been the quiet architect behind the world's devices. It designed the blueprints, licensed them out, and collected a modest royalty every time someone else turned those blueprints into silicon. It was elegant. It was profitable. And as of Tuesday night in San Francisco, it's no longer enough.

At the "Arm Everywhere" event, CEO Rene Haas walked onto a stage and did something the company has never done: he unveiled Arm's own chip. Not a reference design. Not a subsystem. A full, production-ready, dual-chiplet data center CPU called the Arm AGI CPU, fabbed on TSMC's 3nm process, co-developed with Meta, and pointed directly at x86's last great stronghold.

The timing is no accident. The semiconductor world has spent the last two years obsessing over GPUs, pouring billions into accelerators for training massive AI models. But something shifted in recent months. The rise of agentic AI, systems that don't just answer questions but act on your behalf, chaining together searches, tool calls, code execution, and database queries in rapid succession, has quietly rewritten the math on what data centers actually need.

AI agents don't sleep. They submit requests faster than any human, and they generate roughly 15x the token volume per query compared to a simple chatbot interaction. Every one of those orchestration tasks (scheduling, routing, memory management, workflow coordination) lands on a CPU. Haas put the number plainly: each gigawatt of data center capacity in an agentic world needs 120 million CPU cores, up from 30 million just last year. A fourfold increase.

CPUs, it turns out, aren't fading into the background. They're becoming the central nervous system of the AI data center.

The AGI CPU itself is a serious piece of engineering. Up to 136 Neoverse V3 cores per chip, 300 watts of thermal headroom, PCIe Gen 6 connectivity, and CXL 3 memory expansion. Arm claims more than 2x the performance per rack compared to incumbent x86 platforms, a number that, if it holds up in production, could shift how hyperscalers think about density and cooling.

The customer list is notable: Meta as the anchor partner, with OpenAI, Cerebras, Cloudflare, SAP, SK Telecom, and others signed on for deployment. OEMs including Lenovo, Supermicro, and Quanta are building systems around it. Early hardware is available now, with broad shipments expected in the second half of the year.

What makes this particularly interesting is Arm's inherent advantage in power efficiency, the same DNA that put its architecture into every smartphone on Earth. In a world where data center operators are scrambling for power capacity and governments are scrutinizing energy consumption, a CPU that delivers more compute per watt matters more than it used to.

The risk here is…

Arm's entire business model has been built on neutrality. It licenses to everyone (Nvidia, Qualcomm, Apple, Amazon, Samsung) and profits from the ecosystem's collective success. The moment Arm starts selling its own chips, it becomes a competitor to the very customers who pay its royalties. That's a tension that doesn't resolve easily, no matter how many reassurances management offers.

There's also the margin question. Arm's IP licensing business runs at a staggering 98%+ gross margin. The chip business will operate somewhere in the 40-50% range. The gross profit dollars should be accretive (Arm is projecting up to $15 billion in chip revenue by fiscal 2031) but the blended margin profile will look very different. Investors accustomed to Arm's asset-light model will need to recalibrate how they think about this company.

And then there's execution. Arm has never manufactured and sold a full chip at scale. The logistics of supply chain management, yield optimization, customer support for physical silicon, and competing with Intel and AMD's entrenched x86 ecosystems are hard problems that don't get solved by having great IP alone.

Arm is pointing this as a $100 billion total addressable market by 2030, and the company's long-term financial targets are ambitious: $25 billion in total revenue by fiscal 2031, with EPS north of $9. Even if you take a meaningful discount to those numbers, and there are reasonable arguments for doing so, the growth trajectory is hard to ignore.

What's perhaps most interesting is a comment Haas made about Arm eventually addressing a $1 trillion TAM. The implication is that CPUs are just the first chapter. If Arm can prove it can design and sell competitive data center silicon, AI accelerators are the logical next step. That would put Arm in direct competition with Nvidia, AMD, and Broadcom, a significant expansion from a company that, until yesterday, had never sold a chip.

The safe play would have been to stay in the licensing lane, collect growing royalties, and let others handle the silicon. Arm chose the harder path. Whether that turns out to be the right call will take years to sort out, but the logic behind the bet is coherent, and the early customer traction suggests the industry is at least willing to give them a shot.