The most important company in AI doesn't make a model or a chip. It makes the one machine that makes the chip.

We spend an enormous amount of breath arguing about who owns artificial intelligence, and we almost always name the wrong companies. We argue about the labs — OpenAI, Anthropic, Google — as if the models were the asset. We argue about Nvidia as if the chip were the asset. On Wednesday a Dutch company most people cannot pronounce reported its second quarter, raised its full-year guidance for the second time in six months, and mentioned in passing that it already holds orders stretching into 2028. ASML did not hold a keynote. It did not promise anyone artificial general intelligence. It just booked the future of computing and moved on. That is what power looks like when it stops performing and starts invoicing.

Here is the thesis, stated plainly so you can argue with it: the most concentrated, least contestable power in the entire AI economy does not sit with anyone who makes a model, and does not even sit with the company that designs the chips. It sits with the single firm on earth that makes the machine the chips are printed on. There is exactly one. Strip away the narration — the demos, the manifestos, the trillion-dollar valuations — and follow the dependency down to the floor, and the floor is a factory in Veldhoven. The compute is the company, I have argued before. I want to go one layer deeper than I did then, because that is where the argument actually bottoms out.

Follow the dependency all the way down

Do it in order, because the order is the whole point. A frontier model is trained on Nvidia chips. Everyone knows this, and it is why Nvidia became the most valuable company on earth. But Nvidia does not manufacture anything; it designs. Its chips are made by TSMC. So the power, people then say, really sits with TSMC — the foundry that turns designs into silicon. Closer. But TSMC cannot etch the features on an advanced chip without extreme-ultraviolet lithography, the only technology precise enough to draw circuits a few nanometers wide. And EUV lithography machines are made by one company. Not one company that leads a field of several. One company, full stop. ASML is the sole supplier of EUV on the planet, and its chief executive said again this week what he has said before: supply cannot meet demand, and will not through this year.

So run the chain back up. No ASML machine, no TSMC leading-edge node. No leading-edge node, no Nvidia chip worth having. No Nvidia chip, no frontier model. Every layer above rents its position from the layer below, and the descent terminates at a single vendor with a several-year backlog. This is not a metaphor. It is a supply chain that has narrowed, at its most critical joint, to a single point — and a single point of dependence is the most valuable real estate in any industry, because it is the one position no amount of capital above it can route around on demand.

Everyone else in this industry performs power. ASML just books it.

The order book is the argument

I want to be precise about why this quarter, specifically, is the tell. ASML raised its 2026 revenue guidance to €43–€45 billion, up from the €36–€40 billion it had guided only a quarter earlier — a second upward revision inside six months, which is not the behavior of a company reading a soft market. Its chief executive said it is close to holding every EUV order it needs for 2027 and has already taken a large volume of orders for 2028. It is expanding EUV output by roughly 30% for next year and weighing another 30% on top for the year after. Memory revenue, it said, should grow around 75% this year. Read those as financial facts and they are a good quarter. Read them as a power question and they are something else entirely: the entire semiconductor industry has pre-committed its production schedule, years out, to one supplier's capacity plan.

That is the part worth sitting with. A backlog into 2028 means the most sophisticated, best-informed buyers in technology — Intel, TSMC, Samsung, the memory makers — have looked at the same landscape you have and decided the single most important thing they can do with their capital is get in line, early, for a machine only one company sells. They are not hedging. They are not waiting for a second source, because there isn't one. When the people with the most information and the most money queue years in advance for a single vendor's output, that queue is not a market signal. It is a measurement of leverage, and it is pointing at Veldhoven.

Power that never has to narrate

The tech industry runs on narration — the keynote, the mission statement, the founder who tells you the future so persuasively that you forget to ask what he actually controls. I have spent a career trying to separate the people who hold the microphone from the people who hold the power, and the two are almost never the same. Nvidia's Jensen Huang narrates brilliantly; it is part of why the company trades where it does. The labs narrate constantly, because narration is how you raise the next round and recruit the next researcher when the underlying asset — the model — leaks, gets cloned, and depreciates in months.

ASML does the opposite, and the opposite is more instructive. It does not need to tell you a story about intelligence, because it is not selling a story; it is selling the one thing in this whole edifice that cannot be cloned by Friday. Its leverage is so structural that its quarterly earnings call — order intake, capacity, lead times — does more to set the actual trajectory of AI than any product launch this year. The clearest sign of real power is that you can be boring and still be indispensable. ASML is boring in the way a load-bearing wall is boring. Everyone else in this industry performs power. ASML just books it, and lets the order book do the talking.

The strongest case against me

Let me make the counter-argument as well as I can, because it nearly moves me, and a column that won't state the case against itself isn't worth reading. The strong version goes like this. Monopolies at the bleeding edge are more fragile than they look. ASML's next-generation High-NA machine — the one that just printed its first high-volume logic at Intel this week — costs around $400 million a unit, roughly twice the previous generation, and more than a billion dollars by the time you have installed and staffed it. TSMC itself has said, in as many words, that the machine is too expensive. Only Intel is shipping product on it. If the tool becomes so costly that customers stretch their existing machines and slow their adoption, the monopoly's pricing power has a ceiling after all. Add to that the obvious geopolitical fact: China is pouring state money into building a domestic EUV capability precisely because depending on one Dutch chokepoint is intolerable to Beijing. A monopoly with a superpower actively funding its replacement is a monopoly on a clock.

All of that is true, and I was, for a while, more persuaded by it than I am now — I used to think the concentration story topped out at Nvidia and that the layer beneath was too capital-intensive and too slow to be where the real leverage lived. Watching the order book extend into 2028 is what changed my mind. Because here is the answer to the counter-case: expensive is not the same as replaceable, and someday is not the same as now. High-NA being costly slows how fast the frontier moves; it does nothing to introduce a second seller. China being able to make an EUV machine eventually — and "eventually" is doing years and multiple nodes of work in that sentence — is not the same as there being two suppliers today, and today is when the orders for 2028 are being signed. A clock on a monopoly is real. But a monopoly that can raise prices, raise guidance, and fill its book years out while its only rival is still trying to build one working machine is a monopoly enjoying the long, profitable middle of that clock, not the end of it.

Who actually decides

So strip it back to the question this column always asks: when the mission statements and the valuations and the keynotes are cleared away, who actually decides how fast the frontier of computing can move? Not the labs, who wait for chips. Not Nvidia, which waits for TSMC. Not TSMC, which waits for a delivery date from a single company in the Netherlands that is, by its own account, sold out. The pace of the most consequential technology of our era is gated, physically, by the production schedule of one firm that answers to no customer because it has no competitor. That is not a market. That is a checkpoint, and everyone else is in the queue.

If that is right — and the backlog into 2028 is the hardest evidence I know of that it is — then the most urgent industrial-policy question of the decade is not which country wins the model race, or even the chip race. It is whether anyone, any rival or any government, can build a credible second source of EUV before the first one's leverage hardens into something permanent. Everyone is busy narrating a future about intelligence. The people who understand power are doing something quieter and far more revealing: they are standing in line in Veldhoven, paying in advance, hoping the machine ships on time. Follow the compute far enough and it stops being about who is smartest. It comes down to who owns the one machine — and right now the honest answer is a company that never had to raise its voice to own it.