It seems like only yesterday. So can yesterday’s crisis help us understand today’s threats? Perhaps. Nearly 20 years have passed since the global financial crisis of 2007-9, which was fueled by reckless subprime mortgage lending in the United States. Now there are rumblings that the artificial intelligence boom could similarly end in tears. How worried should we be?
A financial crisis is worse than an ordinary recession. It’s when the plumbing of finance fails, and credit stops flowing even to borrowers who did nothing wrong. The Great Depression of the 1930s was an extreme financial crisis, in which 9,000 banks closed and the money supply dropped by a third.
Policymakers sometimes comfort themselves with the notion that financial crises, like earthquakes, are impossible to foresee. It “was very difficult,” Alan Greenspan said when he was chair of the Federal Reserve in 2002, “to definitively identify a bubble until after the fact — that is, when its bursting confirmed its existence.”
In fact, though, financial crises are at least partly predictable, scholars from the United States and Denmark wrote in the Journal of Finance in 2022. They studied 42 countries from 1950 to 2016 and found that the danger signal is rapid growth in borrowing plus rapid growth in asset prices in the same sector over three years. When both conditions hold, the chance of a crisis in the next three years is about 40 percent, compared with roughly 7 percent in normal times.
The housing bubble that preceded the global financial crisis checked both boxes: rapid growth in both borrowing and asset prices. Housing isn’t the risk it was 20 years ago, and banks are safer. But some of the behaviors are reappearing today in the artificial intelligence investing boom. “The lights are flashing orange,” Samuel Hanson, an author of the Journal of Finance article who is an economist at Harvard Business School, said in an interview. The level of warning should give no one comfort, he said, likening the economy to a tinder-dry forest. “People will ask, ‘Well, is it imminent?’ Of course not. That’s an absurd question, and that’s not how we should think about managing risk.”
A warning that borrowing could be reaching a dangerous point came on Oct. 13 when the International Monetary Fund issued a report on financial stability. So far, soaring A.I. investment has been “an important source of support for the global economy and financial markets,” it said. That’s exactly why an A.I. stall would be dangerous. The report went on to warn of “rising leverage in the A.I. ecosystem,” concluding that “the consequences of an A.I. correction could be global.”
Some A.I.-related stocks and bonds sold off sharply the very week the I.M.F. issued its warning. Shares of CoreWeave, a U.S. operator of A.I. data centers, fell 25 percent that week and an additional 9 percent through Thursday, leaving them down more than half from their 52-week high.
Stijn Van Nieuwerburgh of Columbia Business School estimates that U.S. investment in data centers, chips, networking and power will total $10.3 trillion from 2025 to 2032. This would come to an average of 3.6 percent of gross domestic product a year, which would be more as a share of the economy, according to his calculations, than past investment surges in canals (0.7 percent of GDP from 1836-41), railroads (2.2 percent from 1870-90), electrification (0.5 percent from 1905-25), highways (1.1 percent from 1956-73), or telecommunications (1.1 percent from 1996-2003).
This kind of spending makes sense if the world really does want lots of A.I. Everything changes if demand falls short of expectations, which remains a real possibility. Gartner, a research firm, estimated on Oct. 14 that if current investment plans stay on track, the owners of A.I. data centers will need to take in $8.5 trillion in revenue through 2030 to cover their spending, financing costs and a 12 percent return. The firm forecasts that total A.I. revenue attributable to data centers over that period will come to no more than $3.6 trillion, and that less than half of that will go to companies that own the infrastructure.
“The burden is on the A.I. businesses to continue to deliver high, high growth rates,” Torsten Slok, chief economist and partner at Apollo Global Management, an alternative asset manager, said in an interview. “If not, we could have that dreaded ‘aha’ moment.”
The financing of the A.I. build-out has also gotten more opaque over time, calling to mind the famously convoluted structures at the core of the 2008 crisis. At first, the tech giants paid for A.I. investment out of their own abundant cash flows. But as expenses have risen, the burden has been increasingly shouldered by outside investors and lenders. This has sped expansion, but it has also piled up debt in places that don’t show up on the giants’ balance sheets. Mr. Van Nieuwerburgh calls it “a more layered and potentially opaque allocation of risk.”
Similarly, in the lead-up to the last financial crisis, much of the apparent demand for subprime mortgage debt was the industry buying from itself. It was circular financing — and it’s not difficult to draw parallels between that moment and now. For instance, from 2021 to 2025, 55 percent of the investment dollars received by A.I. firms came from other A.I. firms, according to a report this month by the Bank for International Settlements.
A notable example here is the relationship of Nvidia, which makes A.I. chips, with CoreWeave. Nvidia sells its chips to CoreWeave, but it also owns part of CoreWeave. Meanwhile, CoreWeave has long-term contracts with major tech firms, which gave its lenders the confidence to lend it more money so it can grow (and buy more chips from Nvidia). In addition, Nvidia pays CoreWeave for “access to any residual unsold cloud computing capacity.”
The interconnectedness works as long as demand for A.I. is strong. But “these linkages,” the International Monetary Fund wrote, without naming names, “could amplify shocks if A.I. monetization or productivity gains disappoint.”
Richard Bookstaber, a risk manager, wrote a book titled “A Demon of Our Own Design” that was published in April 2007 and warned, presciently, that complex financial products and heavy borrowing had made the financial system brittle. Months later, the financial crisis began to unfold with the collapse of a pair of Bear Stearns hedge funds that were loaded with subprime mortgage securities. Mr. Bookstaber, who is now the chief risk officer for UC Investments, a unit of the University of California with $236 billion under management, says he’s more worried today than he was on the eve of the last crisis. “It’s a period of major risk, and people just don’t seem to get it,” he said.
Paul Kedrosky, an investor, writer, and researcher, agrees. “We’re so far out over our skis, we’re out over someone else’s skis,” he said.
Right now, the A.I. build out itself is inflationary. Constructing data centers is soaking up electricity, copper and workers, while the demand for loans is helping drive up interest rates, including mortgage rates, and crowding out other types of business investment.
In the long run, though, A.I. is likely to displace people from jobs and push down wages, which is “profoundly deflationary,” Mr. Kedrosky said. Deflation, if it occurs, would raise the risk of a crisis. Why? Because debts stay fixed while wages fall, leading to defaults on mortgages, car loans and other forms of debt.
And then there’s the stock market. A.I.-related companies account for nearly half of the S&P 500’s total value. The strength in A.I. has masked weakness in market stalwarts like Home Depot, McDonald’s, and Nike. Mr. Bookstaber said an A.I. crash could send the overall market down 40 percent and a subset of the market down 70 percent. This may sound alarmist, but it’s less than what happened when the dot-com bubble burst in 2000. While a stock market meltdown isn’t as systemically dangerous as a wave of defaults, a big drop would make people feel poorer. Also worth noting: Hyperscalers, including Alphabet and Meta, announced this year they’re selling shares to finance their growth. That strategy depends on their stock prices staying high. Bank lending to private equity and credit rose 26 percent this year, according to the I.M.F. report. “History teaches us to worry when any form of credit, especially a new one, grows this fast,” said Mr. Hanson, the Harvard Business School economist. Lending by private credit firms is not inherently risky, he said, but, “You’re just worried more about unknown unknowns.” Higher interest rates are hurting, too. A.I. stock prices are based on profits expected years from now, and the higher the rate, the less those distant profits are worth today. A.I. doesn’t have to fail for high rates to do damage. It just has to succeed more slowly than its lenders and shareholders are counting on — making it one more reason that there might be more peril in U.S. markets than meets the eye.
The post There May Be More Peril in the U.S. Markets Than Meets the Eye appeared first on New York Times.
