TL;DR · 30-second read
The Short Version
Banks lent enormous sums to build the giant computer warehouses that power artificial intelligence. Now some are struggling to pass that debt on to other investors.
One group has spent more than six months trying to sell $38 billion in loans for sites in Texas and Wisconsin rented to the software company Oracle. Some lenders will now accept less than full price.
Why it matters: the artificial intelligence boom runs on borrowed money. Banks also cap how much they will risk on one industry, and some are nearing those caps. That can make the next project harder to fund.
Startup Fortune reported on October 11, citing the Financial Times and Bloomberg, that JPMorgan, Morgan Stanley, SMBC and MUFG have been exploring significant risk transfers on AI-linked loans as they approach internal limits on exposure to the sector. These deals let a bank keep a loan on its books while shifting the riskiest slice to outside investors.
Separately, a group of banks including JPMorgan and MUFG has reportedly spent more than six months trying to distribute $38 billion in construction debt for a Texas and Wisconsin data center project leased to Oracle. Some lenders are now reportedly willing to sell pieces to non-bank buyers at a discount.
Executive Summary
Underwriting a loan and holding it are different businesses. Banks arrange large financings expecting to sell much of the debt to other investors, which frees room for the next deal. The $38 billion Oracle-leased project suggests that cycle has slowed for at least one of the largest AI data center financings.
The wider credit backdrop has also tightened. Oracle’s five-year credit default swaps, a market price for insuring against a company’s default, reached about 151 basis points in early October, or roughly 1.51% of the insured amount per year. Order books for bonds from Alphabet, Amazon, Meta, Microsoft and Oracle reportedly fell from about five times the amount on offer in February to under two times by July.
The central question is whether investors outside the banking system will absorb the risk of financing the AI buildout, and at what price. Current evidence points to continued demand on more selective terms, not to a closed market.
Why a Slow Loan Sale Becomes a Capacity Problem
The mechanism behind this story is balance sheet capacity. When a bank underwrites a large loan and sells it down, its exposure falls and it can finance another project. When the debt stays on its books, that room remains occupied. Internal limits on exposure to a single sector then turn a slow sale into a constraint on new lending.
The two reported developments fit together. More than six months into the effort to distribute the $38 billion, some lenders are reportedly prepared to accept a discount, while JPMorgan, Morgan Stanley, SMBC and MUFG are exploring risk transfers as they near those limits. JPMorgan and MUFG appear in both accounts. A discount is a real cost, and paying it to move debt suggests that freeing capacity has become worth more than waiting for a better price.
The people most exposed are those who need the next loan: developers, data center operators and smaller AI infrastructure companies. The evidence does not yet show that any project has been delayed or denied funding. It shows pressure on the channel that funds these projects, which is an earlier and narrower finding, but one that matters to anyone with a construction schedule tied to debt financing.
Concentration Is Not the Same as Lost Confidence
It is tempting to read these moves as lenders betting against their own loans. The evidence supports a narrower conclusion. Selling down is what underwriters ordinarily do, and an exposure limit measures how much of one risk a bank will hold, not whether a given loan is likely to default. A bank can like a loan and still hold too much of the sector.
Concentration carries its own concern, however. Special purpose vehicles, separate legal entities that hold project debt, have reportedly been used to spread about $120 billion of AI infrastructure spending, with Oracle, Meta, xAI and CoreWeave all using the structure to keep debt off their own balance sheets. Those vehicles can move debt away from the tech companies, but the banks that financed them can still end up holding the risk. Keeping debt off one balance sheet does not remove it from the financial system.
Scale is part of the challenge. Matthew Moniot, co-head of credit risk sharing at Man Group, told the Financial Times the deal sizes are “out of scale to anything we’ve thought about, ever.” That comment speaks to size more than credit quality, but size alone matters if the pool of risk-transfer buyers was built for smaller transactions.
Investors Are Still Buying, on Stricter Terms
Falling order books need careful reading. Coverage of under two times still means bids exceeded the bonds on offer, and a lower ratio can reflect heavier supply as well as softer appetite. Even so, the drop from about five times is a sharp change in a matter of months, and these bonds now reportedly carry a premium to similarly rated non-AI corporate debt.
Investors are also separating tenants by credit strength, with capital reportedly rotating toward data centers leased to investment-grade hyperscalers, the largest cloud operators, such as Microsoft and Google. Bond market data compiled by CCIR puts hyperscaler-leased paper at 5.70% to 7.875% and CoreWeave-leased paper at 7.00% to 9.875%. The gap is 1.30 percentage points at the low end and 2.00 at the high end, which on an illustrative $1 billion of debt would mean $13 million to $20 million more in annual interest. Maturity and deal structure can also affect such ranges, so the gap is better read as directional than as a pure price of tenant credit.
That pricing is a sign of a functioning market as well as a stressed one. Weaker credits can still find a price, but it may change project economics, and smaller companies raising debt next quarter would be competing for the same lenders as Oracle without its balance sheet or customer base.
Oracle and Treasury Yields Raise the Stakes
Oracle sits at the center of this test. S&P cut its rating in July to BBB-, one notch above junk, and the company is reportedly targeting more than $40 billion in additional debt and equity this fiscal year, on top of the $18 billion in bonds it sold in a single day last September. It is also suing Wisconsin’s Public Service Commission over a rule that could require more than $7 billion in collateral for its Port Washington campus. Its default insurance costs are at levels last seen in 2009.
The base cost of money has risen as well. The 10-year Treasury yield hit 5.304% this month, its highest since 2002, after a $39 billion auction cleared at 5.30%, the highest auction yield since 2000. The Federal Reserve’s September minutes named AI-related corporate borrowing as a contributor to the move, while Capital Economics argues that inflation and government deficits are the bigger drivers. Whichever explanation carries more weight, data center borrowers now price their debt from a higher starting point.
Background
AI data centers are typically financed with large construction loans, often arranged by a group of banks and tied to a long-term lease with a major tenant. Many deals use special purpose vehicles, so the debt sits with a separate entity rather than on the tenant’s balance sheet. Banks normally expect to sell part of this debt to other investors after arranging it.
Six months ago, banks were competing to underwrite data center loans tied to Oracle, Meta and other AI buildout participants. Since then, Oracle has been cut to one notch above junk, order books for large tech issuers have thinned, and long-term Treasury yields have reached their highest levels in more than two decades. Source: Banks are now trying to offload the AI data center debt they just underwrote, Startup Fortune’s report on banks seeking to distribute and transfer the risk of AI data center loans.Sources