The AI Buildout Started Moving Its Debt Out of Sight

Start with one company's own numbers, because they hold the whole month on a single page. CoreWeave — the cloud outfit that rents out Nvidia chips — closed its second quarter sitting on a $104 billion backlog of signed business. The same filing carried its guidance for the next quarter: interest expense of $860 million to $940 million, against adjusted operating income of $200 million to $260 million. The cost of servicing the debt it took on to buy those chips runs three to four times what the chips currently earn. ^[1] The credit market read that page the week of October 14 and drew its conclusion. CoreWeave's bonds fell below 80 cents on the dollar, a level traders call distressed, with yields above 15% — a price that assumes lenders won't get paid back whole. The cost to insure against its default, priced through instruments called credit default swaps, widened past 11% a week later, the kind of number attached to a company already in trouble. The move wasn't confined to one borrower. Meta's default-insurance costs hit a record the same week even as its free cash flow went negative, and the neoclouds — the rented-chip middlemen like CoreWeave, Nebius, and IREN — each shed more than 30% in a month. SK Hynix pulled an IPO. ^[2][3] Goldman Sachs gave the selloff its own explanation: ^[2]

We think the widening reflects a reassessment of the economics of the AI buildout, particularly for lower-rated borrowers that are more exposed to higher-for-longer rates and potential energy-cost pressures. — Goldman Sachs Private Wealth Management

That phrasing matters. Nobody is saying the compute itself has stopped selling. In fact the opposite case is easy to make on the same week's record. Crusoe's chief executive says rental prices for older Nvidia A100 and H100 chips are higher now than when they were first deployed — used hardware, appreciating. ^[4]

We're seeing continued upward mobility in terms of the kind of price per megawatt. — Marc Boroditsky

Nebius is raising its per-megawatt prices with 2027 and 2028 capacity already expected to sell out. ^[4] Even Jensen Huang, Nvidia's chief executive, conceded the wall in plain terms the same week he defended his order book — first calling neocloud demand "off the charts," then admitting the constraint is money, not buyers. ^[3]

demand for the neoclouds is off the charts. — Jensen Huang

The demand of the capital is much greater than the availability of the capital. — Jensen Huang

So the equity analysts are still fighting about which half to believe, and the fight is genuinely split. Thirty-two of forty-five analysts rating CoreWeave still say buy. ^[5] HSBC's Abhishek Shukla is not among them; he cut the stock and put the math bluntly. ^[5]

Not all neoclouds are created equal, and we view CoreWeave as being toward the top of the quality rankings in the ecosystem. — Gabriela Borges

Goldman's Gabriela Borges goes the other way, arguing CoreWeave's data-center leases are not really corporate debt at all but "asset-backed project financings" — buildings that stand behind the loans. ^[5]

The more the company invests with this return profile, the more the value destruction. — Abhishek Shukla

Here is the thing worth noticing, and it is the reason to set the analyst fight down. The debt market has already stopped arguing. It priced the risk and moved on. And the industry's response — visible only when you line up four deals that no single story assembles — has been to make the debt less visible, one instrument at a time. The first is the biggest. SpaceX is seeking $40 billion to buy Nvidia chips, routed through a special purpose vehicle — a separate legal shell that keeps the borrowing off the parent company's balance sheet — managed by an Apollo unit called Atlas SP. The stated purpose is explicit: keep the liability off SpaceX's books and protect its investment-grade credit rating. The precedent shows exactly what this structure is engineered to slip past. An earlier, smaller attempt — a $20 billion financing for the xAI venture — tried to hide the debt through a sale-leaseback, and the auditors rejected it, walking $13 billion back onto SpaceX's books as recorded debt. ^[6] The second is smaller but no less telling. Stonebriar Commercial Finance sold roughly $869 million in asset-backed securities — bonds whose repayments come from a pool of underlying loans — with AI chip loans making up about 15% of the pool. It is one of the first broadly syndicated equipment-financing deals to use GPU loans as collateral, turning a bet on whether a used chip holds value into a bond a pension fund could buy. ^[7] The third is insurance. Nvidia is in early talks, through the broker Howden Re, to have insurers protect lenders if borrowers default and the pledged chips can't be resold for the full value of the debt. The chip vendor, having sold the machines, would now be arranging for someone else to absorb the risk that the machines aren't worth what they were lent against. ^[8] And the fourth is the mildest, which is what makes it land. Nebius, a neocloud competing with CoreWeave, now markets its lighter balance sheet as a selling point — borrowing less as a feature. Its chief revenue officer put it as competition on leverage, not capacity. ^[4]

we're not carrying the kind of debt that they are. — Marc Boroditsky

The sameness accumulates on its own, without anyone needing to say these firms coordinated. And one quiet detail is almost too neat: the economist arguing that AI compute now costs more at the margin than it returns to users is Apollo's chief economist — working at the firm whose Atlas SP unit is arranging SpaceX's $40 billion. ^[9] Make of that what you will. What all four deals share is where the risk ends up. Not on the company's income statement, and not where the equity headlines look. S&P Global's Mark Habib flagged the interlocking financing between hyperscalers and their suppliers and named the consequence directly. ^[9]

It is safe to assume that there is more (debt) out there that we are not seeing. — Mark Habibi

And not everyone wants the paper. UniSuper, a A$180 billion Australian pension fund, is declining direct data-center investments, reasoning that the technology's users, not its infrastructure builders, will collect the payoff. ^[10] So where does the month leave things? The demand is real and still growing; the money behind it is stretched thin and getting thinner. And the market's answer has been to hand the risk to insurers and bond buyers — right down to insuring lenders against the possibility that a pledged chip, seized in a default, won't cover the loan it secured. ^[8] That is not a bet that the buildout stops. It is a bet, hedged and re-sold in smaller pieces, that the buildout's bill can be moved somewhere the quarterly headlines don't reach.