AI Debt Now Trades on Whose Name Is on the Paper

On October 14, chip stocks shed roughly a trillion dollars in a single day, and the news that set it off was not a bad quarter or a failed product. It was two reports about one private company's revenue. The Financial Times put OpenAI's revenue near $50 billion. Axios put it at $70 billion. Neither number came from the company, and neither could be checked against a filing, so the gap itself — twenty billion dollars, on the private firm at the center of the boom — became the day's news. Broadcom's chief executive Hock Tan likened the market to a casino, saying investors are placing bets without reliable data. ^[1][2] A revenue discrepancy at one private company sounds like gossip, not a credit event. It became one because of what the neoclouds are. A neocloud is a landlord for chips: it borrows to buy graphics processors and rents them out by the month. The loans are repaid from rent, and the rent is paid by the AI labs. If a lab's revenue cannot be pinned down within twenty billion dollars, the rent it can keep paying is unverifiable too. Goldman Sachs put the October repricing in plain terms.

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

HSBC named the revenue side outright.

Debt market woes along with moderating [large language model] revenues are exacerbating structural concerns for neoclouds. — Abhishek Shukla

The doubt was on file long before the market priced it. In November 2025, the Financial Times reviewed leaked documents showing OpenAI's inference costs — the expense of running its models to answer users — climbing from $3.8 billion in 2024 to $8.65 billion in the first nine months of 2025, spending more to serve its models than it earned. Sam Altman was asked to square the numbers with his public claims.

I am not sure how to reconcile that statement with the documents I have viewed — Sam Altman

^[3] Since the October rout, the market has sorted AI paper into two piles. In one pile, borrowing still arrives at record demand. A year earlier, Amazon had sold $15 billion in bonds and drawn roughly $80 billion in orders, including a forty-year note; that window stayed open for cash-rich names through the entire cycle. ^[4] In the other pile sit the leveraged landlords. CoreWeave's bonds trade below 80 cents on the dollar, with yields near 15 percent, and its stock fell more than 15 percent on October 14 alone. Its own guidance shows why: the company told investors to expect $860 million to $940 million of interest expense against $200 million to $260 million of adjusted operating income. It pays its lenders roughly four times what it earns operating the business, even with a $104 billion contracted backlog and sold-out capacity. ^[2][5][6] The company said the market had it wrong.

We are confident in our backlog which more than supports all of our debt obligations. — CoreWeave

That sentence went out the same month the bonds hit distress. ^[2] The strangest paper in the welcomed pile shows how far the market will go to avoid trusting a tenant's cash flow. Broadcom is arranging an up-to-$100 billion financing not as a plain corporate loan but as a special-purpose vehicle: a separate legal shell that buys Broadcom's own chips and leases them to end tenants. Broadcom is partially guaranteeing the vehicle's $60 billion to $70 billion senior tranche, the layer of debt repaid first. Analysts call the structure "phantom leverage" — the chipmaker lending its own balance sheet to its own customers, because the customers' cash flow is what no one can verify. The guarantee was not free. Negotiating it drove up the price of insuring Broadcom's own debt against default — its credit-default swaps — and lifted its bond yields, as investors weighed how much risk the chipmaker had just bolted onto its own name. ^[7] None of this is a verdict on whether the chips rent. Capacity is sold out and CoreWeave's contracted backlog tops $100 billion. ^[6] Two weeks ago Nvidia's chief executive Jensen Huang publicly vouched for CoreWeave's financial health — Nvidia is its second-largest shareholder — and the stock fell another 11.9 percent anyway. ^[8] The disagreement is not about demand. In the same week, Goldman's credit analysts flagged the repricing of lower-rated AI borrowers while the firm's equity desk upgraded CoreWeave to Buy, calling the correction a buying opportunity. ^[9][10] Goldman's analyst covering AI token demand, Bobby Molavi, explained why two desks can disagree on identical facts. Demand is growing, and both camps read the same numbers.

one thing is clear for now...the demand is growing. — Bobby Molavi

His figures describe usage, not a trading instrument — 72 percent of compute units are being utilized while only 27 percent of the money is being captured. The monetization gap, not the demand, is the fault line neither side can settle. ^[11] The repricing has so far touched only the debt in plain sight, and the first refinancing dates do not arrive until 2027. Moody's counts a wave of private AI-loan maturities between 2027 and 2029, and flags two weak points: circular financing, where chipmakers and tech firms lend to one another, and payment-in-kind interest, which lets borrowers defer cash payments by folding them into the principal. ^[12] S&P Global's Mark Habib said the interlocking structures between hyperscalers and smaller AI firms may be hiding debt no one has priced.

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

^[13] That hidden debt is the part of this repricing still ahead. October set new prices for the loans in plain sight. The ones Habib cannot see come due in a wave between 2027 and 2029, and the first of them will meet a lender within the next year. ^[12]