ARM built its first chip in thirty-five years. Intel bought back the fab it sold two years ago. Nvidia took equity in its fourth supply-chain partner. Every layer of the AI chip stack is vertically integrating — because the IP licensing model dies when inference value exceeds license fees.
For thirty-five years, ARM was Switzerland.
Founded in 1990 as a joint venture between Acorn, Apple, and VLSI Technology, the company settled on the purest possible business model: design the architecture, license it to everyone, compete with no one. Qualcomm, Apple, MediaTek, Samsung — they all paid ARM a royalty for the right to build chips on its instruction set. ARM never touched silicon. It never had to. The licensing fees were good enough.
On March 24, ARM launched its own chip.
The AGI CPU — ARM General Infrastructure, a deliberate naming choice — is a 136-core processor fabricated by TSMC on its most advanced 3nm process. Meta co-developed it from the design phase. Seven other customers signed on, including OpenAI, Cloudflare, and SAP. ARM's stock surged sixteen percent, its largest single-day gain since going public.
The thirty-five-year neutrality is over. ARM is now competing with its own licensees.
The Pattern
ARM's move is not an isolated event. It is the latest in a sequence that has been accelerating for months, and it follows a single principle: when the value of inference exceeds what you collect in fees, vertical integration becomes inevitable.
Nvidia has taken equity stakes in four companies across the AI supply chain since January. Two billion dollars in CoreWeave for cloud compute capacity. Two billion in Nebius for geographic expansion. An undisclosed but significant investment in Mira Murati's Thinking Machines Lab, paired with a gigawatt of next-generation chips. And on March 31, two billion more in Marvell — for custom XPUs, NVLink Fusion interconnects, and joint silicon photonics research replacing copper with light.
Jensen Huang is not just defining the stack. He is buying it.
On April 1, Intel repurchased its forty-nine percent stake in an Irish chip factory from Apollo Global Management for $14.2 billion. Intel had sold that stake for $11.2 billion in 2024, when it was cash-strapped and desperate. Apollo made roughly three billion dollars in under two years. Intel's stock jumped nine percent on the buyback news — the market rewarding the admission that selling your manufacturing was a mistake.
The fire-sale buyback tells the story more clearly than any earnings call. Private equity firms are extracting billions in rent from semiconductor companies that cannot afford to be tenants in their own fabs. The physical layer is non-negotiable. Everyone who sold capacity is buying it back at a premium.
The Architecture Layer
The journal has covered nearly every layer of the AI chip stack. The memory layer — SK Hynix and Samsung's HBM dominance. The fabrication layer — TSMC's irreplaceable position. The GPU layer — Nvidia's pricing power. The custom silicon layer — Broadcom building ASICs for hyperscalers. The networking layer — the fiber optic buildout.
ARM occupies the layer underneath all of them: the instruction set architecture. Every mobile phone, most data center CPUs, and an increasing share of AI inference hardware runs on ARM's designs. The architecture layer is the substrate on which chips are built — and ARM just decided that licensing the substrate is no longer enough.
The economics are straightforward. ARM collected four billion dollars in revenue last fiscal year — almost entirely from licensing fees and per-chip royalties averaging pennies per unit. A single hyperscaler deploying ARM-based inference chips across its fleet generates billions in inference revenue per quarter. The gap between what ARM collects and what its architecture enables has been widening for years. The AGI CPU is the moment that gap became intolerable.
Two Directions
On April 2, IBM and ARM announced a strategic collaboration to bring ARM-based software into IBM mainframe environments. The partnership targets virtualization technologies that let ARM applications run inside Z-series mainframes — the enterprise computing platforms that still process the majority of the world's financial transactions.
This is vertical integration from the opposite direction. ARM is not only going down-stack by fabricating its own chips. It is going up-stack by entering enterprise infrastructure. The mainframe — the most locked-in, highest-margin computing environment in existence — is absorbing ARM architecture.
Two-directional integration confirms what the licensing fees already revealed. When every layer of the stack is trying to own more of the stack, the pure licensing model is no longer viable. Switzerland works when the countries around it are content to trade. It stops working when each of them decides to build its own army.
The Principle
Every IP licensing model in technology has followed the same arc. A company creates something foundational and licenses it broadly, capturing a fraction of the value its creation enables. For a time, this is rational — the licensees build the market, expand the ecosystem, and the licensor collects rent without risk.
Then the value of what the licensees build exceeds the value of the license. The fraction becomes intolerable. The licensor looks at the margin its customers earn and recognizes that the gap is not a business model — it is a subsidy.
ARM watched its licensees build trillion-dollar businesses on its instruction set while it collected single-digit billions in royalties. Nvidia watched cloud providers resell its GPUs at multiples of what they paid. Intel watched Apollo profit three billion dollars by holding a factory stake for two years. Each responded the same way: integrate vertically, own more of the stack, stop subsidizing the layer above.
The AI inference market is projected to consume two-thirds of all AI compute by the end of this year. Every chip company, every cloud provider, every hyperscaler is racing to own as much of the inference path as possible. The licensing model — the idea that you can design the foundation and let others capture the value — is dying at every layer simultaneously.
The stack is not a supply chain. It is becoming a series of competing vertically integrated companies, each trying to own from architecture to application. The question is not whether this consolidation continues. It is which layers have enough gravitational pull to hold the others in orbit, and which get absorbed.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)