The credit cycle is turning, or so the headlines blare, in block capitals.
After a year in which developers borrowed on a scale the market had never absorbed, banks and bond investors are getting choosier. Yields on weaker deals have climbed. The smaller neoclouds are being pushed to the back of the queue.
That is what a functioning market does when one sector hoovers up most of the available capital. The more interesting question is what all this debt is betting on, and whether that bet is as solid as the volume suggests.
Rational repricing
Investment-grade hyperscalers still borrow cheaply. Weaker, non-investment-grade developers pay up or get turned away.
Riley Thompson of Mitsubishi HC Capital America reckons the market has real appetite for perhaps 20 neoclouds, not 50. That is not a crisis, it is triage. Capital is flowing to the borrowers most likely to repay, which is the system working, not failing.
The circular money problem
The worry is not the price of the debt but the nature of the demand. A growing share of AI spending is self-referential.
Nvidia has committed to invest up to $100bn in OpenAI, which will spend much of it on Nvidia chips.
OpenAI's roughly $300bn arrangement with Oracle sends money to Oracle, which buys more Nvidia chips. Nvidia is also reported to be weighing guarantees that would let OpenAI-linked developers raise debt on better terms. Analysts put the interlocking commitments at somewhere between $1tn and $1.5tn.
Critics draw a direct line to the late-1990s telecom bust, when vendor financing inflated apparent demand until real usage failed to show up. The charge is that cash can leave one balance sheet as an investment and return to another as revenue, flattering demand that has yet to be tested against paying customers outside the circle.
Where the stress is already showing
There are early cracks. About $18bn of loans tied to an Oracle-leased campus in New Mexico, part of its OpenAI agreement, has been quoted by banks at 89 to 91 cents on the dollar after attempts to syndicate the debt stalled. Prime Data Centers delayed a planned bond.
Moody's reckons roughly $1.2tn of data-centre liabilities sit off balance sheets, much of it tied to projects still under construction. JPMorgan has estimated that $4.1tn of AI-related debt could be issued by 2030. Even the wider market is feeling it, with the surge in tech borrowing cited as one reason government bond yields have pushed higher.
Why it probably isn't a crash yet
The bust case is not proven. Demand for frontier compute is real and, for now, outstrips supply.
The Meta-linked junk deal that priced at 8.25% drew orders more than four times its size, so appetite has not vanished. It has just got pricier. The hyperscalers funding most of the build-out carry investment-grade ratings and strong cash flows. KBRA's head of infrastructure finance says higher rates may shape future deals without denting borrower demand much.
For most analysts this is a mapped vulnerability rather than an unfolding cascade: a structure that would be tested by a shock to power, demand or asset values, not one collapsing on its own.
What to watch
The number that matters is not the yield on the next bond. It is whether the capacity being built gets filled by revenue from customers outside the circle of AI companies financing each other.
If that external demand shows up, today's harder money is a footnote and the weak projects fall away quietly. If it does not, the same debt that looks like prudent risk-pricing now will start to look like the early, orderly phase of something larger.
An AlixPartners survey found that 68% of data-centre executives expect distress to rise over the next 18 months. The people closest to the numbers are not assuming the comfortable outcome.
