The AI Money Loop Is Closing at the Top of the Market

Share

I have spent a good deal of 2026 writing about dependency, mostly aimed at founders who build their products on someone else's model. The worry has always been concrete: pricing you do not control, deprecation on someone else's schedule, an architecture that quietly bends itself around one vendor. I still think all of that is true. But the past week gave me the strongest evidence yet that the people we tend to think of as the sellers, the firms at the very top of the market, are themselves trapped in the same problem, and on a scale that makes my usual examples look small. The AI economy has started looking like a closed loop, and the biggest company in it is stuck at the center.

The occasion was Anthropic's journey toward going public. The headline numbers are the kind that make markets go quiet. Reports describe an offering that could raise in the neighborhood of sixty billion dollars, at a valuation that bankers have floated north of a trillion. On paper it reads as the arrival of the first pure AI giant, a company finally ready to cash in on the gold rush it helped start.

Then you read past the top line, and the arithmetic gets stranger. Alongside the public offering, about twenty six billion dollars of nonvoting shares were sold in a private placement at the IPO price, and the buyers were a list that reads less like financiers and more like a parts catalog: Nvidia, Google, Microsoft, AMD, Broadcom, Micron, and TSMC. These are not, by and large, whoever-they-are money funds. They are the firms that make and run the hardware and the clouds this company depends on for everything it does.

That in itself would be a footnote. The part I keep turning over is the compute commitments. In its public filing, the company disclosed roughly nine hundred billion dollars in purchase commitments and leases, substantially all of it tied to compute and infrastructure. Reports by several outlets characterize around four fifths of those commitments as noncancellable, or payable regardless of usage. That last phrase is the one that matters. It means this company has promised to buy enormous amounts of compute whether it needs it or not, with no real way out.

The individual agreements make the point concrete. There is an arrangement to spend more than a hundred billion dollars with Google over the next several years regardless of usage, another around a hundred and ten billion with Amazon, and tens of billions more with Microsoft, again on terms that do not depend on how much compute the company actually consumes. Add it together and the firm that is supposed to be the most powerful seller of intelligence in the world has promised its suppliers roughly fifteen times the money it is now trying to raise.

I have written before that dependence on any single provider is a strategic risk, that optionality you cannot exercise is not really optionality. All of that is mild compared to what these contracts describe. Here is a company whose customers depend on it, which everyone in the industry already understood, and which in turn has committed to pay its own suppliers more than it has ever raised, in good quarters and bad, whether the chips sit idle or not. The dependency I have been warning founders about is not just present at the top of the market. It is most severe there.

And there is a quiet irony in it. The firm that occupies the privileged seat in this economy, the one selling intelligence to everyone else while even its rivals pay it for a model, holds the least leverage in its own most important deal. Leverage in a negotiation tends to go to whoever can walk away. This company cannot. Its suppliers get paid no matter what happens, and when you get paid no matter what happens, you have won the negotiation before it starts. Value, over time, drifts to whoever does not have to be there.

I am not suggesting anyone involved is being naive. The people running these negotiations know exactly what they are signing, and they sign anyway on purpose. Compute is genuinely the constraint that decides who reaches the frontier first, and in a race where the winner takes an outsized share of the industry, paying for certainty has a rationale. When the whole future of the company is at stake, an extra ten billion dollars of guaranteed compute is a cheap price for not being the one caught short. The suppliers know this, which is precisely why they can demand the terms they do.

The shape of it is not new, either. Capital equipment makers financing or underwriting their largest customers is an old story, as old as farmers borrowing from the railroad and shops buying machinery on terms set by the machine seller. The picks and shovels logic has always had this second half to it: whoever owns the thing everyone needs ends up with an unusual amount of control over the people who think they are the ones buying. What is new here is not the mechanism but the tightness and the scale. The same handful of names appears on every side of the ledger at once.

That is the strange part worth sitting with. Google and Nvidia each write checks for ten billion dollars to buy shares in a company, while at nearly the same moment that same company writes them contracts for far larger sums for the chips and cloud capacity it cannot live without. Microsoft appears as an investor, a supplier, and in some quarters a competitor all in the same financial statement. The money does not really leave the room. It goes from the model maker to its suppliers, and a slice of it comes back through share purchases, and then goes out again as compute, around and around among a small number of familiar balances.

I want to be careful here, because it is easy to overstate this into a scandal it is not. There is no fraud in any of it, as far as the filings show, and the commitments purchase real capability that buys real competitive position. A good deal of the underlying usage is probably real too, driven by genuine customer demand for increasingly capable models. If the products keep getting better and the market keeps growing, a lot of this promised spend will be spending the company wanted to do anyway. It would be wrong to assume that any arrangement this large has to be hollow.

But it would be equally wrong to miss what the structure does permanently. It hands the durable leverage to the suppliers, because their revenue is guaranteed whether the model maker thrives or struggles. And it means that part of the growth the market is now pricing is growth happening inside a party of a few. When a substantial slice of a giant company's most important business is conducted with its own shareholders, who are also its suppliers, who are also in some cases its competitors, the tidy picture of independent giants freely competing starts to fray.

I keep coming back to the definition of a moat. A moat is something that protects you from having to compete on someone else's terms. The companies in this loop are building real things, but the loop itself is a form of mutual protection, a way for a small group of very large firms to make sure they keep winning and keep each other financed. Which is fine for them. The question it raises for everyone outside the room is what independence means anymore in this industry, when even the biggest independent company turns out to be bound to its suppliers more tightly than almost any customer and vendor pair I can think of.

There is a lesson in it for founders, though it is not an instruction and I will not dress it up as one. It is simply an observation about where the power actually sits. We tend to picture the AI value chain as a clean hierarchy, the model makers on top, everyone else below. The reality of the past week is that the firm at the very top is the one least able to leave its own table. The companies that look like they are being sold to are, in the end, the ones collecting a guaranteed premium. Power lives with whoever is paid no matter what happens next.

None of this tells me the bubble is about to pop, and I am not predicting collapse. These are serious businesses with serious revenue and serious customers, and the frontier race is not a mirage. But when the biggest company in an industry has to promise its suppliers more money than it has raised, and those same suppliers hold a stake in its outcome, the industry has stopped resembling a market and started resembling an agreement among a few large parties about how to keep winning for a long time.

The AI gold rush has become, at its top, a closed system of a handful of companies circling one another, each one investing in the others, each one selling to the others, each one depending on the others to keep the machine running. Everyone in that room believes they are buying insurance against the day one of them stumbles. The quieter truth is that only the people whose checks arrive regardless of how the race ends are collecting a premium on it. Everyone else in the loop is just paying one.