You're on the Other Side of the AI Bubble

The consensus that AI is a circular-financing bubble about to pop is right about the money and wrong about you. The financing is a loop: chipmakers invest in the model labs that buy their chips through cloud providers the same chipmakers backstop. The loop is fragile and a correction may come. But an operator buying AI by the month sits on the far side of it from the investors at risk, and a bust hands value to that operator.

Give the bears their strongest case, because it is strong. Nvidia took roughly a 7% stake in the cloud provider CoreWeave and agreed to buy more than $6 billion of its capacity, while CoreWeave borrows against contracts like that to buy more of Nvidia's chips. Repeat that shape across the same small set of chipmakers, model labs and clouds and you reach Bloomberg's tally of the interlocking commitments: more than a trillion dollars, a large share of it the same capital going in circles.

OpenAI sits at the loop's center. It is on track to lose around $14 billion in 2026 and is reportedly preparing a public listing for late this year at a valuation bankers have floated between $850 billion and a trillion dollars. The bear case writes itself: heavy losses, a price tag staked on a future that has not arrived, and demand partly manufactured by the sellers financing their own buyers.

The analogy the skeptics reach for is the late-1990s US telecom build-out, when carriers borrowed more than $500 billion to lay fibre-optic cable on demand projections that never showed up, and the sector collapsed by 2002. The investment firm GMO has said today's loop "resembles" that engineering, and the comparison is fair. Vendor financing that makes demand look organic when part of it is the seller funding the buyer is exactly how that bubble inflated. If you hold equity in any name inside the loop, the people shouting bubble are doing you a service.

Here is what the warning leaves out. That argument is about asset prices, conducted among the people who own the assets. The operator paying by the seat or by usage owns none of them. You are the demand the loop is straining to manufacture, which means every dollar of OpenAI's loss is partly a subsidy paid to you: frontier capability sold below what it costs to make, because the producers are buying market share with their investors' patience. A correction ends that patience. It does not reach into your business and remove the capability you have already learned to run.

The telecom bust is the precedent that matters, and it runs against the way the bears use it. The fibre did not vanish when the companies that laid it went bankrupt. Four years after the crash about 85% of it still sat dark, and the price of bandwidth fell by roughly 90%. That glut, bought for almost nothing, became the physical backbone of the broadband, streaming and cloud era. The loss fell on the financiers; the businesses that won the 2000s were the ones using bandwidth that had become nearly free, and the bondholders were not in the room.

The comparison has a real weakness, and skating past it would be the same motivated reasoning the bears are accused of. Fibre in the ground lasted decades; the chips at the center of this build-out lose most of their value in about three years and need constant, costly replacement to stay current. A data centre is not dark fibre waiting to be lit cheaply later. So the durable residue of an AI correction lies elsewhere: in the model capability already trained and frozen into weights, which does not un-happen, and in the long fall in the price of using it. A capable lightweight model now serves output at around $0.25 per million tokens of input, the small units of text these models are billed by, and the strongest models held their prices flat through 2026 even as their capability climbed. That curve has bent one way for three years, through boom and pullback alike, and a bust only speeds it up as stranded compute is sold to whoever can use it.

The loop is not only American, which matters if you are weighing whether to build on it from outside the United States. In France, Mistral is funding its expansion the same debt-heavy way, raising $830 million from European banks for a Paris data centre and committing to about a billion euros of capital spending in 2026, with a sovereign-compute site in Sweden behind it. The label is European autonomy; the mechanics are the same leveraged bet on demand arriving. The lesson carries: an operator in Madrid, São Paulo or Berlin who rents metered access holds no risk on any of these balance sheets. They use the capability and let other people's shareholders fund the build.

So an operator's response to "the bubble is about to pop" should be close to the reverse of caution. The real exposure is internal: a business whose only advantage is access to AI, a thin layer of prompts over a model anyone can call, priced as if the underlying capability will stay scarce and dear. A glut destroys exactly that premium. We argued in defensibility in the AI era that models commoditise while judgment, trust and relationships compound, and a compute glut is the fast-forward button on the first half of that sentence. The operators who can hold their ground use cheap, abundant AI as an input to something a price collapse cannot manufacture. The ones who should be uncomfortable sell access itself, and a glut turns their supplier into their direct competitor.

The consensus has the finance right and the conclusion backwards. The build-out is funded by a fragile loop, and some of the names inside it will not survive a correction; that is an investor's problem, and the people naming it are doing investors a favour. The operator stands on the other side of the loop, buying frontier capability that is currently sold below cost and getting cheaper every quarter. A bust will not raise that price. The telecom crash says it lowers it, and leaves the capability behind for whoever keeps building. The only operators a correction genuinely threatens are the ones whose plan quietly depends on AI staying expensive. If that sentence made you uncomfortable, you have just located the part of your business that was never really yours.