Weak Links Start to Show in A.I. Borrowing

Trillions of dollars in borrowed money have powered the A.I. boom. But now there is growing uncertainty around just how easily more money can be obtained — and how much of it will be paid back.

Those questions burst into the forefront this week when investors rapidly lost confidence in one debt-laden data center company, CoreWeave. Prices of its bonds — the company has accrued $72 billion in liabilities to operate 51 data centers — crashed, and investors placed a more than 60 percent likelihood CoreWeave would fail to meet its obligations.

Investors have sold off bonds for other A.I. giants as well, including Larry Ellison’s company, Oracle, though the price declines are not as steep as CoreWeave’s.

At the same time, Wall Street investors have been bidding up the price of some complicated financial instruments that insure corporate debt against default, such as those insuring SoftBank, the Japanese technology colossus that has borrowed heavily to bet its future on a partnership with OpenAI.

The recent market moves could spell losses for investors that own the companies’ existing bonds and could mean the companies will be forced to pay higher interest rates when they borrow again, raising the costs of their already wildly expensive A.I. projects.

The broader question is whether companies facing debt trouble like CoreWeave, Oracle and SoftBank are the industry’s weak links or are harbingers of deeper problems in the A.I. build-out.

The spread, or extra premium that investors demand to put their money into risky A.I. debt investments, has shot up more than 50 percent over the past six months, according to Barclays, the investment bank.

“The market is on a knife’s edge everywhere,” said Alexandra Coleman, founder of Azraq, a data center risk researcher.

A lot is riding on the continuation of the borrowing binge. Every week lately, economists, banks and private equity firms project higher and higher amounts of debt that they say will be needed to finance the artificial intelligence systems that their boosters assert will run the world in a few years.

Virtually every major mutual fund company, insurer and private equity firm has been buying up the A.I. debt, to say nothing of the scores of individual investors who have poured money into skyrocketing A.I. companies.

Not every data center developer worries bond investors equally. The larger so-called hyperscalers, such as Meta and Alphabet, which continue to borrow heavily, have been forced by investors to pay only modestly more in interest in recent debt deals. Investors have also sold out of some these companies’ existing debt, though the drop in bond prices has been less steep than for CoreWeave.

Some of this can be attributed to the Federal Reserve’s recent decision to begin raising interest rates in an effort to control inflation, which is being caused, in part, by the data center boom.

David Munves, head of credit strategy at the researcher AIR Platforms, said there was a “collective realization” this week that the risks of the artificial intelligence surge were greater than anticipated.

“The market’s waking up to the fact that there is potentially a mismatch between the money they have to commit today and borrow today and the ability to generate revenues down the road,” he added.

This week, the yield on CoreWeave’s bond eclipsed 15 percent, well into “junk bond” territory, and the company’s stock dropped 24 percent, including a 12 percent fall on Friday alone.

CoreWeave’s business model defies easy description. Its financial arrangements have been compared by some investors and analysts to that of the troubled office-space provider WeWork.

First, CoreWeave buys Nvidia and other A.I. chips. Then, as with WeWork, which leases space inside office buildings owned by others, CoreWeave contracts with data center owners to use the structures for A.I. computing for 15 years or more. CoreWeave then subleases the space to customers, including Meta, Microsoft and OpenAI, for a shorter time frame.

The rub is that CoreWeave is on the hook for costs on both ends: It must pay to assemble and operate the facilities before any tenant is in place, and it is committed to the space for longer than its customers may stick around.

Several investors said they were worried about whether hyperscalers such as Meta and Microsoft would even need to lease space from CoreWeave in the future. Alternately, they’re wary of whether the data centers could become dated or even obsolete as their existing contracts expire.

A CoreWeave spokeswoman, Lia Davis, said the company disagreed with the WeWork comparison. In a statement, she said that “the complexity of our financing reflects discipline, which is a point of pride for us.” CoreWeave’s chief executive, in a note to employees on Wednesday, called the recent market moves “irrational.”

In a March presentation to potential lenders, CoreWeave said it would earn enough from a five-year lease to cover its related expenses for twice that time.

“We are confident in our business model, our customers, the extraordinary quality of our execution and the relentless discipline around the financial engine we have built,” CoreWeave’s chief executive, Michael Intrator, told employees in his note.

Another worry: the increasing number of data centers that are being delayed by permitting issues, and in some instances falling through.

Oracle has faced slowdowns on multiple projects. Last year, the company easily borrowed $38 billion from a group of banks to build, with the help of project developers, a pair of data centers, one in Texas and another on 2,000 acres of farmland in Port Washington, Wis.

Oracle planned to pay back the borrowed money on the Texas facility helped by profits from the latter complex, called Project Lighthouse, which it expected to be operational by 2028.

Since then, Project Lighthouse has run into local hurdles, including from state power regulators who have pushed back their time horizon to approve the construction of a power grid critical for the data center to operate.

Aterio, a data center researcher, said it was looking less likely that Project Lighthouse’s four data centers would start operating before 2029.

That’s just a year before Oracle’s debt comes due. Some of the debt has traded around 90 cents on the dollar or below that, according to two traders, a steep drop for a relatively new loan.

An Oracle spokeswoman declined to comment.

Oracle faces similar permitting issues at a data center project in New Mexico. The project’s developer, the investment firm Blue Owl, has started construction on the facility, but Oracle is struggling to get state permits to provide enough power.

Some of the $18 billion in debt tied to the project has recently sold at steep discounts, roughly 80 cents on the dollar, according to two people familiar with the trades.

SoftBank has borrowed heavily to make one of the biggest bets — by dollar value — on OpenAI. The Japanese conglomerate, led by Masayoshi Son, wrote its first $500 million check in 2024 and a year later invested $30 billion more.

To get that money, Mr. Son sold off most of SoftBank’s holdings of Nvidia and T-Mobile and took out loans against his stake in the chip designer Arm Holdings.

This year, Mr. Son invested again. On Oct. 1, SoftBank completed another $30 billion investment in OpenAI. Before doing so, in late September, the company raised a record $11 billion in debt at interest rates of nearly 10 percent.

In response to questions about SoftBank’s increasing debt, Alex Clavel, chief executive of the SoftBank Vision Funds, said the amount of their debt as a percentage of SoftBank Group’s assets hadn’t changed.

“The assets we’ve owned have increased significantly,” Mr. Clavel said in an interview.

The cost of insuring against losses on SoftBank’s debt has spiked. It now costs more than $40,000 to insure against $1 million of SoftBank debt. At the beginning of this year, the cost was less than $27,000.

Some investors see a potential silver lining to the troubles. John Lloyd, the global head of multisector and corporate credit at the asset manager Janus Henderson, said slowing down the financing for projects “makes existing projects and existing A.I. infrastructure even more valuable.”