Whose Demand Is the $250 Billion Really Buying?

The circular-financing panic over this week's biggest AI deal is right about the money and wrong about the demand. Nvidia is in talks to guarantee roughly $250 billion of debt so that OpenAI, its largest customer, can lease a vast data centre and buy still more of Nvidia's chips. The loop is genuinely fragile. But the demand underneath it is real, already here, and running short of supply, which changes what the deal means for anyone running a business.

Give the skeptics their strongest case, because it is strong. The reported arrangement would have a chipmaker underwrite its buyer's borrowing: Nvidia's credit standing behind roughly $250 billion of OpenAI debt, so that OpenAI can lease a new data centre and buy still more Nvidia hardware. Reporting from Bloomberg and CNBC on July 26 and 27 put the full campus near half a trillion dollars. It is the clearest instance yet of a supplier funding the demand for its own product.

This is not a one-off. Nvidia has announced more than half a trillion dollars of such deals this year, taking stakes in or lending to firms that then spend the money on its chips, while OpenAI loses around $14 billion in 2026. The precedent critics reach for is the telecom build-out of the late 1990s, when equipment makers like Lucent lent customers the cash to buy switching gear and fibre before those customers had the revenue to justify it. When demand fell short in 2001, the buyers went bankrupt and the lenders were left holding the debt. The financing today has the same architecture, and interlocked balance sheets magnify a fall rather than cushion it. That risk is real, and it deserves the attention it is getting.

Here is the variable the analogy quietly drops. The carriers of the 1990s laid fibre for traffic that did not yet exist, and much of it sat dark for years. The AI build-out is chasing demand that already arrived. Roughly a billion people use ChatGPT every month, more than nine in ten of the largest US companies use it, and OpenAI's revenue has climbed to about $25 billion a year while the company still cannot secure compute fast enough to serve them. AI's bottleneck today is supply: there is more demand than hardware to meet it.

Even the doubters have noticed. Derek Thompson, who spent much of 2025 arguing that AI was surely a bubble, wrote this summer that for now it surely is not, after corporate spending on coding agents grew so fast that one major lab's servers buckled under the load. When the same careful observer argues both sides inside a year, that is information: the yes-or-no bubble question is the wrong one to build a business around. The sharper question is whose demand the money is actually buying.

Two different demand signals are being folded into one word. The first is end-user pull: the billion monthly users and the corporate seats, buyers paying for output because it does work they value. The second is capital-spending demand: a supplier guaranteeing its customer's debt so the customer keeps buying chips, orders a vendor is manufacturing for itself. Most commentary blends the two into 'AI demand' and then argues about whether the total is a bubble. An operator has to pull them apart, because only one of them says anything about their own business. End-user pull is a fact to build on. Supplier-funded capital spending is a headline to discount.

This is not only an American story, and the contrast abroad is instructive. The same week the Nvidia talks surfaced, the European Union opened its AI Gigafactories programme, up to €10 billion of public and national money meant to unlock more than €30 billion in total for a network of large compute sites across the bloc. Europe is financing its build-out through the state and industry rather than through a supplier lending to its own customer. The demand thesis is identical on both continents: compute is scarce and worth a fortune to secure. What differs is who carries the risk when the bill comes due.

So the operator's job is not to guess the timing of a burst. We argued in You're on the Other Side of the AI Bubble that a company renting AI by the month sits across the loop from the investors at risk, and that a correction hands it cheaper capability rather than a catastrophe. That still holds. This week sharpens it with one update. If real demand is outrunning compute, an operator's nearer risk is frontier capacity that stays scarce, rationed, and priced at a premium for longer than the cheap-and-abundant story promised. Plan for two tiers. Run the bulk of the work on commodity models whose price keeps falling, and treat guaranteed access to the frontier as the contested input it is becoming. The people who should be uncomfortable are the ones whose plan assumes frontier AI will be cheap and readily available next quarter, and the ones reading capital-spending headlines as proof the market is hot for their product. That heat may belong to the chip supplier.

The consensus has the risk half-right and the lesson half-drawn. The financing is circular and fragile, and a chipmaker standing behind its customer's debt is close to the move that broke the telecom equipment makers a generation ago, so anyone with money inside the loop should read the warnings closely. But the fibre in 2000 was laid for customers who never came, and the compute in 2026 is being built for a billion people already waiting. That difference is what matters here. For an operator, the deal is not a cue to retreat from AI, nor to time a crash you hold no position in. It is a cue to separate the demand that is yours from the demand a supplier is buying on credit, to build on the first, and to plan for a frontier that stays expensive precisely because too many real customers want it at once.