I won a Pulitzer for explaining the Great Depression. The AI spending boom terrifies me

When ChatGPT was first released in late 2022, few economists anticipated what an enormous impact it would have. Large language models have been adopted on a mass scale faster than virtually any other major technological innovation of the last century. This rapid adoption has triggered an epic capital investment boom, as technology companies have raced to expand spending on data centers, specialized chips, cooling systems, and the power generation capacity the technology demands.

The five major technology companies—Amazon, Microsoft, Alphabet, Meta, and Oracle—along with the two leading independent model builders, OpenAI and Anthropic, and a host of other start-ups are projected to spend anywhere from $800 billion to $1 trillion a year for the next few years on AI infrastructure. These are staggering numbers amounting to 2%–3% of GDP, made all the more striking by how few hands they are concentrated in.

I have seen numbers like these before. They did not end well.

The Railroad That Broke America

Large as these investments are, they are not entirely without precedent when measured against the broader economy. History offers several comparable episodes of private capital investment booms. The buildout of the fiber-optic network from 1995 to 2000, which provided the backbone of the modern internet, saw IT investment climb from 3% of GDP to nearly 5% at its peak. The investment surge of the Roaring Twenties—driven by the mass adoption of the automobile, the expansion of the electricity grid, and a construction boom in housing—was another. Finally, and most significantly, there was the railroad boom of the late 19th century— the subject of my new book, 1873, and the parallel that keeps me up at night.

The fear is that, if the past is anything to go by, the AI boom will follow a similar arc to these other technology-driven infrastructure booms: a flood of speculative capital will flow in, leading to massive overinvestment, asset price bubbles and ultimately a crash as euphoria collides with a disappointing reality. That is not a prediction drawn from ideology. It is a pattern I have spent years documenting in the archives.

The most instructive parallel to today’s AI boom may be the railroad construction boom of the 1870s and 1880s. Like AI, railroads were a transformative new technology that promised to reshape the country’s geography and remake its economy. Between the end of the Civil War in 1865 and 1872, the U.S. railroad network doubled from 35,000 to 70,000 miles. Railroad construction was voracious in its appetite for capital, accounting on average for roughly 2.5% of GDP throughout the 1870s and 1880s.

Because equity markets were then quite small, more than three-quarters of the railroad construction boom was financed by issuing bonds. This made railroad companies especially vulnerable during periodic revenue downturns, when many found themselves unable to service their debts and were forced into bankruptcy. As a result, the pace of construction was anything but steady. There were three distinct booms and busts in railroad construction between 1869 and 1896, the most dramatic of which came in the early 1870s, when spending peaked at $400 million a year—nearly 5% of GDP.

Will the AI investment boom mimic that of the railroads? Until recently, most investors took comfort in the fact that, unlike the 19th-century railroads, much of the AI capital investment boom was being financed out of the massive profits earned by big tech companies such as Microsoft, Alphabet, Meta, and Amazon. The expectation was that startup AI companies such as OpenAI and Anthropic—which have comparatively modest revenues—would have no problem raising capital through  IPOs or from deep-pocketed sovereign wealth funds of the Middle East. Easy access to capital was thought to preclude the sort of boom-bust cycle that characterized the railroad era.

Jay Cooke’s Trillion-Dollar Mistake

The 1870s, however, offer a cautionary tale. After the completion of the first transcontinental railroad in 1869, the single biggest railroad project of the early 1870s became the race to build a second one. Among the most ambitious was the Northern Pacific: a long-dormant plan to run a line from Lake Superior to Puget Sound. Its sponsor was Jay Cooke, the most prominent banker in the country and one of its wealthiest men. Cooke committed to raising $100 million for the Northern Pacific—the equivalent, relative to the size of the global economy, of over $1 trillion today. If anyone could have pulled it off, it was Cooke. He had made his fortune by selling $2 billion of government bonds for the Union cause directly to the public during the Civil War, a feat that had made him a national figure. He was exceptionally well connected: a personal friend of President Ulysses Grant, his name had been floated as a potential Secretary of the Treasury when Grant first took office in 1869.

Nevertheless, by 1873, Cooke found himself struggling to deliver. The underlying finances of the U.S. railroad industry had begun to look increasingly precarious. There had been ruinous duplication of lines, and much of the construction boom had been concentrated in thinly populated states like Michigan, Illinois, Indiana, and Wisconsin—regions unlikely to generate the passenger and freight traffic needed for profitability. By the fall of 1872, barely 100 of the 350 railroad companies were paying dividends.

When Easy Money Ran Out

Most importantly, as Cooke embarked on his effort to raise capital, financial conditions in the world’s capital markets changed dramatically. Europe had long been a primary source of financing for American railroad investment. In 1870, however, the unexpected outbreak of war between France and Prussia—the third and fourth largest economies in the world—resulted in two years of financial turmoil on European capital markets that cooled the appetite for American securities. The ill repute of the U.S. railroad industry, increasingly beset by political scandal, compounded the problem: big European investors such as the Rothschilds, who then ran the largest and best-capitalized bank in the world, would only underwrite bonds issued by U.S. railroads if they carried a U.S. government guarantee—a request that would never have received Congressional approval.

By the beginning of 1873, Cooke had managed to raise only $14 million of the $100 million he had promised, with less than a third of the planned track having been laid. Finding it increasingly difficult to persuade the public to buy Northern Pacific bonds, and unwilling to abandon the project, he began—as a last resort—to pump funds from his own bank into the railroad to bridge its financing needs. Using money from depositors who had the right to withdraw their funds on demand to finance long-term, illiquid investments with an uncertain and distant payoff was a highly risky strategy—and one with echoes of what is currently creating turmoil among private credit funds today.

On September 18, 1873, Cooke’s principal creditors—the largest banks in New York—pulled the plug, announcing they would no longer lend to Jay Cooke and Co. The fatally overextended bank was forced to close its doors. The collapse of this one institution had a devastating impact on investor psychology. The New York Stock Exchange suspended trading for ten days. Investor expectations were drastically revised downward, and doubts spread about the viability of railroad investment broadly. Hard-hit investors dumped their railroad bonds. Within two years, a third of all railroad bonds in the country had stopped paying interest, and construction on new lines ground to a halt.

The Danger of Competing for History

Could something like this derail the AI investment boom? The companies involved in building out AI infrastructure, committed as they are to spending vast sums on capital expenditure, now face a not-too-dissimilar situation to the railroad companies of the early 1870s. In particular But in the last six months two things have changed dramatically over the last twelve months.

First, estimates of what companies are willing to spend on AI infrastructure keep being revised upward. As in so many preceding booms, even though tech companies may be aware that collectively they are overinvesting, each individual company—believing that the rewards for being one of the winners are enormous—has a powerful incentive to go for broke. Competitive pressures are therefore forcing even the giant hyper-scalers to seek alternative sources of outside capital in the form of debt to supplement their internal funds. More and more data centers are being financed by companies issuing debt, tapping private credit funds, or through special purpose vehicles backed by securitized debt and mortgage bonds.

The net effect has been to make the whole funding structure supporting AI more vulnerable in a downturn.

Second, in much the same way as the Franco-Prussian War of 1870 dramatically changed conditions on international capital markets, the war in the Middle East has upended the world’s equity and credit markets. Fears that European economies will be squeezed by rising inflation and slowing growth have caused the risk premium on long-term interest rates around the world to spike. Meanwhile the supply of capital from the Middle East is likely to slow, creating a global liquidity squeeze.

The Line You Can’t See Until You’ve Crossed It

Six months ago, the expectation was that the AI capital investment boom would be easily funded. That path now looks less certain than it once did, and it is not inconceivable that one of the original AI startups could run out of money before meeting its gigantic investment goals. Even some of the giant tech companies, such as Oracle, are facing a squeeze on their finances and may have to drastically scale back their ambitions. The AI capital investment boom, once thought to be immune from the ups and downs of capital markets, has become hostage to their vagaries. In these circumstances, it would not take much to create a crisis of confidence among investors comparable in its psychological impact to 1873.

Long-term investors should take comfort in one thing: 1873 was not the end of the railroad boom. Even though it set back the development of the U.S. railroad network by half a decade, the animal spirits of investors eventually revived. There were two more railroad mini-booms, in 1881 and again in 1893, both followed by busts. But the railroads survived and became a central vehicle for driving the economic growth of the United States for half a century. Like the railroads, the AI capital investment boom—even if it suffers a temporary setback—is not going away. It will be with us for decades.

Perhaps the most important lesson of the railroad boom — and of the Depression, and of every capital mania I have studied — is that the line between visionary investment and ruinous excess is invisible until after you have crossed it. We may have already crossed it.

The opinions expressed in Fortune.com commentary pieces are solely the views of their authors and do not necessarily reflect the opinions and beliefs of Fortune.