The 2031 notes yield close to 16 percent, the shares fell 25 percent last week, and the question circulating on credit desks was the one nobody had answered with a figure: how much of this sits on bank balance sheets. JPMorgan’s analysts put out an answer on Tuesday: the largest American banks and Japan’s Mitsubishi UFJ carry exposure of 1 billion to 3 billion dollars, and most European and Asian lenders hold less than 500 million.
The estimate came from Kian Abouhossein’s team, which described the exposure as well secured. It is the first quantification by a bank’s research department of what the AI cloud company’s debt problem could cost the institutions that arranged it, and it arrived with a caveat that matters as much as the number: the math rests on simplified assumptions because disclosure is thin.
The structure explains why the analysts are calmer than the bond market. Much of CoreWeave’s borrowing sits in committed delayed-draw term loans, of which about 24 billion dollars has been committed, 15 billion drawn and 9 billion still available. The money is released as projects hit construction targets, in many cases through special-purpose vehicles that hold the assets and the debt together. A bank in that arrangement is not marking a loss every time a bond trades; it is holding a commitment that gets funded on a schedule.
That reframing moves the risk from price to timing. If the undrawn 9 billion dollars is called while the high-yield market is shut, the lenders have to fund it from their own balance sheets at a moment when they would rather not. JPMorgan expects CoreWeave’s free cash flow to stay negative until 2030, which means the company will return to the capital markets repeatedly before it generates cash to service what it has already borrowed.
The trigger for last week’s selloff was a comment rather than a number. A co-founder said the company was exploring ways to raise money in the investment-grade market, where insurers and pension funds buy. To holders of its junk bonds, that reads as an intention to borrow more cheaply against the same assets, ranking new lenders ahead of them. Venture investor Chamath Palihapitiya then posted a chart of the 2032 bond on X, and a repricing became a stampede.
The repricing was violent by any standard. The 9.75 percent notes due in 2031 traded near a 16 percent yield to maturity, and the 9.625 percent notes due in 2032 near 14.5 percent, far above the coupons at which they were sold. The cost of insuring the debt against default rose past 1,100 basis points at one point. CoreWeave carries roughly 50 billion dollars of debt and obligations, most of it raised against contracts with customers whose own credit is what the lenders are really underwriting.
Not everyone reads the bond prices as a verdict on the company. Goldman Sachs analysts argued that the market has picked the wrong benchmark, because unsecured notes do not reflect what new borrowing actually costs when it is secured by specific assets and contracts. That is technically correct and politically awkward: the unsecured market is where the price is discovered, and it is the price that determines whether the next deal gets done.
The spillover has been wider than one borrower. Five-year credit default swaps on Oracle widened to a record 288 basis points on October 15, and SoftBank’s reached 424, according to market data. Both are financing data center capacity at a scale that makes them sensitive to the same question. Morgan Stanley estimates that hyperscalers and AI cloud providers will spend 1.58 trillion dollars in 2027, up from about 980 billion this year, and that number is the demand forecast behind nearly every loan in the sector.
The banks’ own behavior is the tell. JPMorgan and MUFG have spent months trying to place the last portion of a 38 billion dollar loan tied to Oracle’s data center buildout, and the syndication has been slow. Lenders that were eager eighteen months ago are now negotiating covenants and holding positions larger than they intended.
For the regulated institutions, the distinction that matters is between credit risk and liquidity risk. A secured loan against a building full of graphics processors, with a customer contract attached, is a decent asset if the processors hold value and the customer pays. The hardware’s residual value depends on how quickly Nvidia renders it obsolete, and the contract depends on a business that is itself burning cash. Neither assumption gets tested in a rising market.
The 1 to 3 billion dollar figure also has a boundary problem. It is an exposure number rather than a loss estimate, and it excludes the second-order commitments of regional banks, insurers, private credit funds and the vendors extending trade credit. The largest banks are the visible part of the financing stack, not the whole of it. When a market closes, the lenders who are least able to absorb a loss are usually the ones who were last to leave.
Abouhossein’s conclusion is that the exposure is manageable, and at the level of an individual bank’s capital that is probably right. A 3 billion dollar exposure against a balance sheet measured in trillions does not threaten solvency. What it threatens is appetite. Every bank that has to explain a CoreWeave position to its credit committee will price the next AI data center loan a little wider, and the cost of that caution lands on the borrowers who need the money most.
The yield on the 2031 bond is the market’s estimate of the risk. The 9 billion dollars of undrawn commitments is the part that has not been priced, and the schedule on which it gets called will decide whether this month’s scare was a repricing or the beginning of one.
