On the chart, the two lines separate. One is the credit spread on the AI hyperscalers, the handful of companies that own most of the computing capacity in the world, and it has widened to its highest level of the year. The other is the share price of the same companies, which has fallen but not by as much. Goldman Sachs published the comparison this week and drew the conclusion the gap implies: either the bond market is too pessimistic about these businesses, or the stocks have roughly 15 percent further to fall.
The 15 percent is not a forecast with a date attached. It is what the equities did the last time spreads on this group were as wide as they are now, a historical analogy offered as a measure of the distance between two markets that are supposed to be pricing the same cash flows.
The credit side of the chart has been moving for weeks. Credit default swaps tied to the largest AI borrowers reached records in October, and the debt of the second tier has been repriced harder. CoreWeave’s 2031 notes, couponed at 9.75 percent, have traded at yields above 15 percent, a level the corporate market usually reserves for companies already negotiating with their lenders. Oracle’s five-year swaps widened to a record 288 basis points on October 15, according to market data.
Equity investors have watched this with the patience of people who own the residual. That patience is rational in one sense and dangerous in another. A bondholder is paid a fixed coupon and cares about one thing, which is whether the money comes back; a widening spread is therefore a direct statement about survival odds, made by people who lose first and recover least. A shareholder owns everything left over, which means a shareholder can be right about the long run and wiped out before it arrives.
There is a mechanical reason credit tends to move first. The high-yield market is smaller, more concentrated, and repriced by a few hundred accounts that read indentures for a living. Equities are repriced by everyone, including people who trade an index without reading a balance sheet. When the two disagree about the same set of contracts, the disagreement usually resolves in the direction the specialists were pointing, though not always on their timetable.
The fundamental case on the equity side is about growth rates rather than levels. Capital spending by the cloud operators has been growing at a rate that cannot be sustained, and the year-over-year growth rate is expected to peak in the third quarter before decelerating. A decelerating growth rate is not a decline. But equity valuations built on a second derivative do not survive its reversal comfortably, and the suppliers of equipment feel it before the operators do.
Then there is depreciation, which is the part of the argument that arrives late. A fleet of accelerators is capitalised and written down over a handful of years. While the fleet is growing, new machines dilute the cost of old ones and the charge looks manageable. When growth slows, the same charge lands on a smaller revenue base, and the useful life assumed at purchase becomes the number that decides whether reported profit exists. Companies that extended useful lives to flatter earnings will have to defend the assumption in front of auditors and analysts who have started asking.
Put together, the two branches of the Goldman argument have different consequences for different companies. The largest platforms generate cash from search, advertising, software and retail. For them, a wider spread is a statement about the marginal project, not about solvency, and a 15 percent drawdown in the shares would be a valuation event. For the companies whose only asset is a building full of rented silicon, the same spread is a statement about refinancing, and a 15 percent drawdown would be the least of it.
That asymmetry is what makes the chart useful rather than merely alarming. It does not say the sector is a fraud or that the debt is unpayable. It says the equity market is charging a price for AI computing capacity that assumes the financing keeps working, while the credit market is charging a price that assumes it might not, and one of them is wrong.
The resolution will probably come through disclosure rather than through a single dramatic session. Guidance for the fourth quarter, the terms on the next bond issue, the willingness of banks to extend construction loans, the tone of the next set of lease disclosures from the neoclouds: these are the places where the gap closes. Credit markets close by a queue that stops moving, and queues are visible before they stop.
For now the more likely reading is the bleaker one, for a reason that has nothing to do with AI. When equity and credit diverge this far, the usual outcome is that equities converge toward credit, because credit investors have less room to be wrong and therefore less incentive to be hopeful. The 15 percent is the size of the gap. Whether it closes by the bond market relaxing or by the stock market falling is the question the chart cannot answer, and the one that decides who pays for it.
