The Chipmakers Won Q2's AI Race, Not the Model Labs
The biggest surprise of Q2 2026 wasn't a model. It was a margin.
For three months the story looked like a heavyweight fight between labs. Anthropic passed OpenAI at a $965 billion valuation, then filed to go public. OpenAI and SpaceX filed too — three of the largest IPOs in history, within weeks of each other. Everyone argued about which lab would win.
We spent the quarter building The State of AI · Q2 2026, a force-by-force map of where the industry actually moved. And the most durable answer to "who's winning" turned out to be a company almost nobody was arguing about: Micron.
A company that already cashed the check
On its June quarter, Micron's revenue quadrupled to roughly $42 billion and its gross margin crossed 81% — up from 27% a year earlier. Its entire 2026 supply of high-bandwidth memory, the specialized chips every AI accelerator needs, is sold out under multi-year contracts. The CEO said the company can fill only half to two-thirds of the demand.
That is not a company hoping AI pays off. It's a company that already booked the profit.
Four toll booths
Micron isn't the exception — it's the pattern. Strip away which lab wins the model war and one fact survives: every one of them buys compute from the same tiny set of suppliers, and that chain has only four real chokepoints.
- Lithography — ASML. The only company on earth that makes the EUV machines that etch advanced chips. Chipmakers route roughly a quarter of their capital spending straight to it.
- Foundry — TSMC. Where nearly every frontier AI chip, Nvidia's included, is fabricated. HPC and AI are now 61% of its revenue, growing 40% year over year at 66% margins.
- Memory — Micron and SK Hynix. The HBM oligopoly. SK Hynix posted a 72% operating margin; its high-bandwidth-memory orders exceed three years of capacity.
- Accelerator — Nvidia. Still the default. A record $75 billion data-center quarter, up 92% year over year. "Cloud GPUs are sold out," said Jensen Huang.
Four companies. Each a near-monopoly. Each sold out. Each printing the best margins of its life.
The bubble debate is happening on the wrong floor
This is where the AI-bubble argument goes wrong — and there are good skeptics making it. Ray Dalio called a bubble. A live scoreboard tracks the industry's bottom line and keeps coming up "no." On one Friday in early June, Nvidia fell 6% and slipped below a $5 trillion valuation, dragging the whole complex down with it.
But the question the skeptics ask — will the returns ever show up? — already has an answer. They showed up. Just not where everyone was looking. The labs are still promising profitability. The companies that sell them the picks and shovels have already collected it.
Whoever wins the model war — Anthropic, OpenAI, a Chinese lab, or none of them — still has to write the same four checks first. The surest way to own the AI boom may turn out to be owning the toll booths, not the cars.
The squeeze is leaving the data center
And it's no longer contained. With HBM and DRAM sold out years ahead, the price of the memory in your next phone, laptop and car is already climbing. The AI build-out is quietly becoming a tax everyone pays — whether you ever touch a chatbot or not.
That's one of six forces we traced through the quarter. The full map — every claim linked to its source, a reader poll in each section — is here: The State of AI · Q2 2026.
— Alexis