Debt Now Comprises A Third Of AI Boom Financing, Exacerbating Economic Risk

There was worry enough when big tech companies were spending vast sums on AI with little evidence to suggest revenues and profits would ever be sufficient to generate a return on investment. 

But the risk has become substantially greater now that so many companies are using debt financing - borrowed funds - which comprises more than a third of all dollars being spent on the AI boom. The impact of debt on top of an equity collapse would be catastrophic for the financial markets and the broader global economy. JL

Cade Metz reports in the New York Times:

In the gallop toward bigger AI projects, outfits, far from household names, not nearly as wealthy but eager to get in on the A.I. boom, have started to build their own giant data centers. And are borrowing tens of billions of dollars to do it. There are growing concerns that these companies are shouldering risks they may not be able to handle. So are much larger companies working with OpenAI. The debt used to fund data centers could exceed $1 trillion by 2028, more than a third of all dollars spent on these facilities. If A.I. does not pull in as much revenue as expected, the debt-laden companies could be left holding the bag. “People are lending an awful lot of money to companies based on speculative returns.” The shift to debt financing is reminiscent of the dot-com boom.

For years, the tech industry’s giants, which make tens of billions of dollars in annual profits, usually built new data centers with their own money. Just last year, Google expanded an already massive computing facility in Oklahoma, while Amazon went to work on a new data center in Indiana that will eventually use enough electricity to power over a million homes.

But a new set of free spenders is emerging in the gallop toward bigger and bigger artificial intelligence projects. Smaller outfits — far from household names, not nearly as wealthy but eager to get in on the A.I. boom — have started to build their own giant data centers. And they are borrowing tens of billions of dollars to do it.

In September, Meta agreed to buy $14.3 billion in computing power from CoreWeave, a New Jersey company that went public this year. CoreWeave has told financial analysts that for every $5 billion in computing power it plans to sell to customers over the next four years through new data centers, it must borrow $2.85 billion. 

The same month, Microsoft made a similar deal with Nebius, a start-up based in Amsterdam, for $19.4 billion. To help build its data centers, Nebius recently sold $3.16 billion in what are called convertible notes, which can be converted into company shares years down the road but begin as debt. Now, there are growing concerns that these smaller companies are shouldering risks they may not be able to handle, entwining themselves in relationships that financial analysts say are worryingly opaque. So are a handful of much larger companies that are working with OpenAI, the San Francisco company that launched the A.I. boom with its ChatGPT chatbot three years ago.

The debt used to fund data centers could exceed $1 trillion by 2028, or more than a third of all dollars spent on these facilities, according to analysts at Morgan Stanley. If A.I. technologies do not pull in as much revenue as expected over the next several years, the debt-laden companies could be left holding the bag for the rest of the industry.

“People are lending an awful lot of money to companies based on speculative returns,” said Gil Luria, head of technology research at the technology analyst D.A. Davidson.

The shift to debt financing is reminiscent of the dot-com boom in the late 1990s, when many companies racked up debt as they raced to lay the fiber-optic cables that would become today’s high-speed internet. When the bubble burst, companies like WorldCom, Global Crossing and Lucent went bankrupt or had to sell themselves to larger rivals. Contributing to this mounting pile of debt are projects led by OpenAI. Even though the company is pulling in billions of dollars in annual revenue, its chief executive, Sam Altman, has said it will not be profitable until 2029. But OpenAI and several partners, including the software giant Oracle and the Japanese conglomerate SoftBank, plan to spend more than $400 billion building data centers in Texas, New Mexico, Ohio and Wisconsin. It is unclear how much debt the company and its partners are taking on to pay for all of that.

Days after OpenAI announced its new data centers, Oracle said it would take on an additional $18 billion in debt. Analysts at KeyBanc Capital Markets estimate that Oracle will have to borrow $25 billion a year over the next four years to build these facilities.

And in some cases, Oracle is paying for only part of the data center. At OpenAI’s first data center in Abilene, Texas, Oracle is paying for the computer hardware inside, while a company called Crusoe is responsible for erecting the building, the cooling equipment and other infrastructure.

For its part of the project, Crusoe borrowed $15 billion through a partnership involving Blue Owl Capital, a private credit lender that is helping to fund several data centers across the country and will own the one in Abilene. SoftBank and OpenAI intend to pay for new facilities in Ohio and Texas in part by borrowing money, according to two sources familiar with the arrangement who spoke on the condition of anonymity.

In recent weeks, OpenAI made a series of unorthodox deals to raise additional money. First, it sold a stake in the company to the chipmaker Nvidia for $100 billion. Then AMD, an Nvidia rival, agreed to essentially give OpenAI tens of millions of AMD shares, which could be worth tens of billions of dollars.

OpenAI has also agreed to buy computer chips from those companies as it builds its data centers. If OpenAI does not need those chips, Nvidia and AMD can pull out of the deals, but OpenAI and others could still be on the hook for the debt they have taken on. “The risk is that companies will buy a bunch of computer chips for A.I. and they won’t have the revenues to pay for them,” said Andrew Odlyzko, a mathematician and historian who specializes in the technological manias of the recent and distant past. “Bubbles take a long time to build. But they burst very quickly.”

As companies borrow money to build data centers, their collateral is sometimes the computer chips they will install in these giant facilities; if the companies default on their loans, their lenders own the chips. But computer chips, like cars, are worth less over time.

The risks could extend beyond the tech industry, financial experts say. The data center debt is held by a wide array of financial institutions, including traditional banks, private credit lenders like Blue Owl and Ares Management, and the many companies that are funding new computing facilities. “Who is holding all this debt? It is not like it is just being held by big banks. It is all over the place,” said Paul Kedrosky, a research fellow at M.I.T.’s Institute for the Digital Economy who has closely tracked spending on new data centers for A.I. “A lot of little pieces of debt are spread throughout the economy.”

The opacity of many of these deals has made that hard for financial analysts to figure out, said Jeremy Kress, an associate professor of business law at the University of Michigan who specializes in systemic risk in the economy and financial stability. If the loans are highly leveraged — meaning that the lenders have taken on their own debt to make the loans — the risks go up.

“Leverage in the system is what drives risk,” Mr. Kress said. “And it is hard to know how much leverage is in the system.”