AI stopped paying for itself with profits and started paying with debt. Ask the fiber-optic bondholders how that ends.

Torsten Slok, the chief economist at Apollo, said something this week that is worth writing on the back of your hand. The profits powering the AI boom, he argued, are being funded by investors rather than earned from customers — and the capital going into it is now growing about twice as fast as the money that went into the mid-2000s housing bubble. Set the housing comparison aside; everyone reaches for it and it flatters no one's thinking. The load-bearing clause is the first one. Funded by investors, not customers. That is not a statement about whether artificial intelligence is real. It is a statement about who is paying for it while we wait to find out, and it points at the part of this boom almost nobody is watching, because it does not trade on a screen you check hourly.

Let me separate the two things that always get welded together at this stage of a mania, because the welding is how people lose money in both directions. The first thing is the technology, and the data centres, and the demand. Those are largely real. Chips are being installed, models are being trained, and a genuine amount of that compute is being rented by customers who use it. I am not here to tell you the picture is fake. The second thing is the trade built on the first — the prices, and, more than the prices this time, the financing. And the financing has quietly changed character in a way that should get far more attention than another debate about Nvidia's multiple.

The tell is not the price. It's the funding source.

For most of this build-out, the hyperscalers paid for their ambition the boring way: out of the enormous operating cash flow their existing businesses throw off. That is the healthiest way to fund a bet, because if the bet sours you have cut your own profits and no one else's. What changed over the last year is the mix. By Morgan Stanley's estimate, AI-related debt issuance will approach 570 billion dollars this year, more than double last year's, with something like 236 billion already sold by the end of May — roughly four times the prior year's pace. Big-tech capital spending is now running at close to a hundred percent of operating cash flow, against a ten-year average nearer forty. Read that shift slowly. The build-out has stopped being paid for out of profits and started being paid for out of borrowing. Even Moody's, not an institution given to shouting, warned in late July that the scale of the spending threatens the credit quality of the companies doing it.

This matters because equity and credit are two different instruments for reading the same story, and they fail differently. The stock market prices disappointment: when a bet underperforms, the shares fall, believers take a loss, and life goes on. The credit market prices something colder — default. It does not care whether AI is transformative. It cares whether the specific borrower generates enough cash, on time, to make a coupon payment, and if the answer is in doubt it does not mark the bond down twenty percent and shrug. It asks for its money back. When Slok says the returns are being funded by investors, the sharpest version of his point is that a growing share of those investors are creditors, and creditors are the ones who turn the lights off.

The stock market prices disappointment. The credit market prices default. Only one of them can ask for its money back. — On why the funding source is the story

Where it shows first

If you want to see where a credit story surfaces first, do not look at Microsoft. Look at the pure-play AI cloud companies — the 'neoclouds,' the ones that do nothing but buy graphics chips on borrowed money and rent them out. Late last month the market gave a preview. The credit-default swaps on CoreWeave, the largest of them, spiked to around 855 basis points — a level that, on a standard model, implies something close to a fifty-percent chance of default within five years. Its shares and those of its peer Nebius fell hard the same week, extending declines that ran to a third or more over the month. This is a company that has had negative free cash flow since 2022, whose interest expense alone doubled to 536 million dollars in a single quarter, and which is rated deep in junk. None of that means CoreWeave is going to fail. It means the people who lend to it are now demanding to be paid as if it might, and that is new information the equity headline missed.

Here I have to be honest about my own record, because a columnist who only remembers her hits is running a marketing operation, not an argument. I have called this kind of thing early before, and early is a way of being wrong — you can be right about the destination and lose your shirt on the timing, and I have the scar to prove it. I am not calling a top. Slok is not calling a top; his own phrase is that rising borrowing costs could make the capex cycle 'self-throttle,' which is a slowdown, not a crash. The point is narrower and more useful than a prediction. It is that the composition of the money has changed, from equity that absorbs losses quietly to debt that enforces them loudly, and that the enforcement mechanism has already started clearing its throat.

A bad company, a bad price, and a bad structure

My usual discipline is to separate a bad company from a bad price, because conflating them is how people misjudge every cycle. This one needs a third box: a bad structure. Consider how the neocloud borrowing is actually built. CoreWeave's landmark facility — 8.5 billion dollars, arranged this spring, celebrated as the first investment-grade financing backed by GPUs — earned that rating on a detail worth reading twice: it rests on Meta's contractual obligation to pay, not on CoreWeave's own credit. The high rating is Meta's balance sheet wearing a borrower's name. Then look at the next facility down the line, a 3.1 billion-dollar deal a couple of months later, backed by weaker customer contracts and rated squarely in junk. That is the tell. As a lender you take the best counterparties first; when the new financings step down in quality, it means the best contracts are already pledged and what is left is thinner. The collateral underneath all of it is graphics chips — an asset that, unlike a fiber cable in the ground, can lose something like half its resale value in three years as faster chips arrive. You are lending against a depreciating asset, secured by demand you are taking partly on faith. And the insiders are reading the same page: CoreWeave's founders sold hundreds of millions of dollars of stock last quarter — 736 and 447 million by two of them — alongside a 407 million-dollar sale by a director of Nvidia, the very company selling them the chips. Legal, pre-scheduled, and still worth noticing who is trimming while the story is loudest.

It is not only the neoclouds. Oracle, which has bet its next decade on renting AI compute, has watched its own credit-default swaps climb past 215 basis points from about 145 at the end of last year, and one of its long bonds now yields close to eight percent. The company reported a remarkable 638 billion dollars of contracted future revenue — a genuinely enormous, genuinely real-looking number — of which roughly half is owed by a single customer, OpenAI, which does not yet turn a profit and only a small fraction of which is due to convert to cash next year. A 638-billion-dollar backlog is exactly the sort of round, magnificent figure I have learned to distrust, not because it is fake but because its quality — who owes it, when, and how surely — is doing all the work the headline hides.

The precedent everyone has agreed to forget

Now for the memory, because my only real edge is remembering what the room has agreed to forget. We have seen a real, world-changing technology financed with mountains of debt against physical collateral, on demand projections that turned out to be part true and part myth. It was called the telecom build-out, and it ran from roughly 1998 to 2002. The internet was real; the traffic was real; the fiber was really laid. The demand forecast that justified the borrowing — WorldCom's famous claim that internet traffic was doubling every hundred days — was not real, and it did not have to be entirely false to be lethal. It only had to be less true than the coupons required. When it broke, WorldCom filed the largest bankruptcy of its era with some 35 billion dollars in liabilities, Global Crossing went down with it, around two dozen telecom companies failed, and the bondholders — the people who thought they were the safe ones, senior to the equity — recovered on the order of twenty cents on the dollar. As much as ninety-five percent of the fiber that had been laid sat 'dark,' unlit, waiting for demand that arrived years later and enriched whoever bought the assets out of bankruptcy, not the people who financed them.

The parallel is not that AI is fake; it wasn't fake then either. The parallel is the structure: a real technology, a genuine demand curve, and a financing stack that assumed the demand would arrive faster than it did, secured against an asset — fiber then, chips now — that was worth a great deal if the timing held and very little if it didn't. As Robert Schiffman of Bloomberg Intelligence put it, the market is starting to ask how much debt capacity these companies really have, and the answer increasingly depends on showing that AI can be monetised rather than merely built. That is the WorldCom question in a 2026 suit.

There is an honest rebuttal, and I will make it myself so you don't have to. The hyperscalers are not the overbuilt telecom start-ups of 1999; they are among the most profitable enterprises in history, and they can fund a great deal of this out of cash for a long time. True. But that defence is precisely the line between the two halves of this trade. The companies that can self-fund are not the ones borrowing at 855 basis points. The neoclouds — the levered, single-purpose GPU landlords — are the ones sitting where the CLECs sat, and they are where the credit market is pointing. A great technology and a solvent borrower are different claims, and this cycle has learned to let the obvious truth of the first smuggle in the unexamined hope of the second.

So here is where I would leave it, before the round numbers get any rounder. The question worth holding is not the one the stock market keeps re-litigating — is AI real, is it a bubble, is this time different. The technology is real and this time is a little different, which is what makes it dangerous, because it is always a little different. The question the credit market is quietly asking is older and ruder: when the compute has been built and the demand shows up on its own schedule rather than the coupon's, who gets paid back, and in what order? Equity busts burn the believers. Credit busts send the bill to everyone — the pension fund, the insurer, the senior lender who was told the chips were collateral. Slok's line deserves its place on the back of your hand. If the returns are being funded by investors and not customers, then the boom is running on borrowed money and borrowed time, and only one of those two is patient.