Few people doubt that AI is among the most transformative technologies to emerge in the past century. By many measures, it is spreading faster and more widely than any innovation since the dawn of the Industrial Revolution. Steam engines came into common use in the nineteenth century and electricity in the early twentieth, but their full potential was not realized before the completion of railways, highways, and power grids decades later. Computers for business use emerged in the 1950s, but they gave no visible lift to the economy before the Internet connected them in the 1990s.

Commercial artificial intelligence, by contrast, appeared almost overnight on screens worldwide, at a moment when the overwhelming majority of humans already had access to a computer or smartphone. Less than four years after the release of ChatGPT, roughly 30 percent of Americans are now using AI, five times the share of Americans who were using smartphones, the Internet, or electricity at comparable stages of development. Revenue for AI companies is three times what previous digital technologies, including the Internet, cloud computing, and early mobile apps, generated in their early years.

Policymakers, investors, and business leaders are counting on AI to be a miracle cure for everything from runaway public debt to the costs associated with aging populations. Among investors, even some of the world’s most successful contrarians—who make their fortunes betting against conventional wisdom—are embracing the idea that AI could yield bigger benefits more rapidly than any past technological advance.

The euphoria is not universal, however. An ever-lonelier minority of skeptics argue that AI’s rise may seem unusually fast but is in fact following a centuries-old script. Major technology revolutions have always triggered dramatic boom-bust cycles. People get overexcited about the latest breakthrough and begin investing indiscriminately in the companies driving the innovation. As a team of researchers from Texas Tech University, Texas A&M University, and the Vlerick Business School in Belgium found in 2018, all of the most revolutionary technological innovations of the last two centuries inflated a major market bubble—a rapid increase in stock prices, followed by a precipitous decline.

The word “bubble” tends to conjure an idea as thin as air. But as with all previous tech bubbles, the innovations of the current AI boom are real, the hype born of genuine promise. That promise, however, is exactly what is inflating the bubble. Once the mania begins to crack, as it inevitably will, the shock will ripple back through financial markets into a global economy increasingly dependent on AI for growth. The damage will be felt worldwide, not only in the countries leading AI’s development but in the many now battling to supply them with inputs, as well. The weakest players will be wiped out, and the survivors will consolidate their lead. And from the rubble, a more stable AI ecosystem will emerge.

The AI competition is already reshaping the economic world order. At the top of this order stand the United States, where the foundational AI model, ChatGPT, reached consumers in late 2022, and China, where DeepSeek made its debut around a year later. Since then, the race for supremacy has come to drive both countries’ economies. In the United States, growth has been fueled primarily by the hundreds of billions of dollars of Big Tech investment in AI projects and by consumer spending energized by gains in a stock market dominated by AI. The top ten stocks in the benchmark S&P 500 are all big AI players and together account for nearly 40 percent of the index and around half of its returns this year.

China is spending far less on AI but would be lost without it. After falling in the last decade, net exports—mainly of AI-related goods—now account for a third of China’s growth. The domestic economy is otherwise moribund, hampered by bad debts and a shrinking labor force. Beijing is also heavily supporting the export of its homegrown AI models, which are catching on because they are cheaper than American frontier models but perform nearly as well.

The AI rage now extends beyond Washington and Beijing. Investors are plowing funds into companies that supply American tech giants with the computing power, semiconductors, and other building blocks they need, fueling stock market booms worldwide. In Asia, for example, three major semiconductor companies have reaped the benefits: TSMC in Taiwan and SK Hynix and Samsung in South Korea, which comprise an even larger share of their local stock indexes than big American tech firms account for in the U.S. index. In these relatively small economies, the global frenzy for AI has generated even greater shocks than in the United States. The total value of South Korea’s market exploded from $1.5 trillion at the start of 2025 to over $5 trillion this summer. Individual investors saw bigger gains in stock wealth over 18 months than they had in a lifetime. Chip exports have driven real growth in Taiwan toward double digits in some recent quarters, a pace rarely approached in any economy.

Other countries at this level of the so-called AI stack of computing infrastructure have benefited similarly from the global market tilt toward builders of AI hardware, or “picks and shovels.” High tech now accounts for as much as half of all economic growth in Israel, owing in part to the country’s expertise in the high-speed networks needed to train AI. In Japan, firms that specialize in AI-related equipment, materials, robots, and semiconductors have become market darlings; after Kioxia, a pioneer in memory chips, became a major supplier to U.S. tech companies, its share price spiked to more than 45 times what it had been a year earlier, briefly becoming Japan’s most valuable stock.

All five of these leading tech powers—China, Japan, South Korea, Taiwan, and the United States—spend, on average, much more than other countries on R & D and technology as a percentage of GDP. But they are far from the only countries racing to secure footholds in the AI stack.

Below them, secondary suppliers are forging ties to the leaders. U.S. and Chinese AI giants are spending billions on data centers in Singapore and Malaysia. Firms are investing heavily in AI server production in Mexico, which is emerging as an important exporter of AI-related hardware, as well. At least half a dozen tech companies have launched AI projects in Vietnam, including a major R & D center and training facilities to develop the country’s talent pool.

In turn, the worldwide AI construction project is expected to provide years of steady demand for the raw materials required to wire and store energy for data centers: nickel from Indonesia, lithium from Australia, copper from Chile, Peru, and the Democratic Republic of Congo. And the large cloud-service providers known as hyperscalers are looking far and wide for that energy. Amazon, Google, Meta, and Microsoft accounted for half of all clean energy deals signed across the globe in 2025.

Private spending on data center construction is up by roughly 50 percent from a year ago.

Not all countries have joined the bonanza. Those outside the stack struggle to attract investment and credit as the overwhelming focus on AI sucks money away from their industries. Much of Europe finds itself excluded. So do countries with emerging economies such as India and the Philippines, doubly shunned because their thriving back-office service sectors are particularly at risk of displacement by artificial intelligence.

It’s hard to think of any period in history when the markets were so obsessed with—and so dependent on the success of—a single sector. The Internet era saw money spill out of the booming sector and into other industries. Today, AI, and little else, is hot. Private spending on data center construction is up by roughly 50 percent from a year ago, while spending on all other building projects in the United States is down by five percent.

So far, the lift from AI has been strong enough to shield the global economy from the geopolitical shocks of the last two years, including U.S. President Donald Trump’s tariffs and the U.S.-Israeli war against Iran. Gains in advanced manufacturing and stock market–driven consumer spending continue to power growth in the United States, and AI-related exports are propping up countries further down the stack. Tech exports from Mexico and Thailand, for example, are rising fast enough to offset slack domestic demand in both economies.

The question is whether AI can sustain the world economy or is a bubble waiting to pop. A bubble can be characterized as a good idea taken too far. And there is mounting evidence that markets have begun to push the tantalizing idea of AI beyond its limits.

Going back centuries, pioneering innovations have spawned remarkably consistent patterns of human behavior. Firms spend heavily on developing a new technology in the hope of dominating a burgeoning industry. Investors follow, with professionals leading the way and amateurs jumping in later. The firms borrow to build; investors borrow to buy more stock in those firms. Trading reaches manic intensity. Prices rise beyond a point that potential profits from the innovation can support. As the peak nears, price increases both accelerate and become more volatile as doubt begins to gnaw at the edges of the frenzy.

This basic dynamic has driven bubbles in everything from tulips to housing since at least the 1600s. In short, the object of desire becomes overvalued, overowned, overinvested, and overleveraged—what I have elsewhere called the “Four O’s” of a bubble.

Historical evidence suggests that the most impressive innovations generate the biggest bubbles. The scale of the bubble also depends on how far and fast word of the hot new thing travels. In the 1870s, a series of bubbles, including one in railroad stocks, moved quickly between Europe and the United States via the new transatlantic telegraph. In the 1920s, electricity and electric gadgets were promoted over the airwaves on one of those gadgets, the radio. Today, hype for AI is spreading on smartphones and social media, digital products that did not yet exist when the Internet bubble popped, in 2000.

Taken together, these phenomena should raise concerns over just how far the mania will go. If AI is indeed the most transformative and most rapidly diffused new technology ever, the bubble it creates could be the most dramatic yet.

In the major bubbles of the past half century, prices have typically risen tenfold, adjusted for inflation, over a period of 15 years. U.S. tech stocks crossed that threshold in 2025. The AI bulls argue that unlike the dot-coms of the Internet boom, today’s marquee tech firms are generating strong earnings and will continue to do so, and their escalating stock prices are therefore justified.

But the excitement about tech bubbles is always stirred by genuine progress, not mere marketing. Although the dot-com era, for instance, is linked in the popular imagination with profitless startups such as Pets.com, these were minor players. The giants of the late 1990s, including Qualcomm, Oracle, and Cisco, logged earnings on par with Amazon, Google, and Microsoft today. The tech sector posted the same annual rate of earnings growth in the late 1990s as it has in the current decade: 22 percent.

The assumption that a boom can’t be a bubble if corporate earnings are strong is belied by history. It is in fact not uncommon in a bubble for earnings to be artificially inflated as well. This twin risk was only partly evident in the dot-com period, but it was clear when the craze for housing and commodities took off in the early 2000s.

What is more, earnings may not be as solid as they appear. They are propped up in part by “circular funding,” in which a hyperscaler invests in a frontier AI lab such as OpenAI or Anthropic. The lab then uses the money in part to buy services from—and thus pump up the earnings of—the hyperscaler. Hyperscalers also revalue their stake in an AI lab and report the increase as “other income,” a category that has accounted for a quarter of their earnings growth this year.

AI labs are among the many U.S. companies that have been choosing to remain private for as long as possible to avoid public scrutiny of their earnings. Now, many of these firms are planning to go public, but they are far less profitable than their predecessors in the late 1990s were. The two biggest names in the pipeline for an initial public offering of stock are Anthropic and OpenAI. To gain customers and market share, both companies are charging their subscribers less than it costs to train and run their AI models. Analysts have already estimated each company’s value at well over a trillion dollars, even with no profits, based on rapid revenue growth alone.

Americans, meanwhile, are investing in stocks as they never have before. The United States is the only country in which families hold more wealth in stocks than in their homes. The typical U.S. household holds more than 30 percent of its wealth in stocks, surpassing the prior peak of 27 percent in 2000. Since much of this money is held in “passive” funds that simply track the S&P 500 index, the bulk of it goes into the AI stocks that dominate the index.

Professional money managers aim to beat the market by being different, but they are finding it difficult to buck the crowd. Their holdings look increasingly like those of the retail investors they once dismissed as “dumb money.” Nearly a third of the stocks held by amateur traders are also held by hedge funds and other quintessentially aggressive pros, a record share.

Over the last five years, the number of shares traded each day in the United States has surged by 50 percent to around 18 billion. Retail investors have surpassed the pros as the most active traders and are aggressively “buying the dips”; in 2026, trading on down days has surged to more than three times the average for all days. Foreigners are betting big on American AI, too; their holdings of U.S. stocks have mushroomed from under $2 trillion to more than $20 trillion since the dot-com bust, with most of that increase coming since the launch of ChatGPT, in November 2022.

AI is thus already overowned by all players in the market. The zeal among retail investors has spread to other AI-driven markets, including South Korea, which in the past has been a bellwether of global stock market movements. Sharp drops in big semiconductor stocks pushed the Korean index down this summer, before locals once again started “buying the dips.” This kind of stop-and-go volatility often signals that a market rally is in its last stages.

Much of the debate about the economics of AI focuses on whether the enormous sums that technology firms are investing in building infrastructure can generate commensurate returns in the foreseeable future. Tech investment has more than quadrupled in the last three years, a boom unmatched in scale or speed by the rise of railways, canals, or the Internet. Tech spending on new plants and equipment hit a record 6.0 percent of GDP at the height of the Internet bubble. Today, it stands at 6.3 percent and is projected to keep increasing.

One way this splurge could pay for itself is by raising productivity, or output per worker, and energizing the entire economy. But at a similar stage of the Internet boom, growth in the key measure of “total factor productivity,” which aims to capture how much output workers are generating from machines, was surging. Now it is barely rising.

Even if productivity does ultimately rise, most AI firms may not reap the rewards. The economist William Nordhaus has shown that only two percent of the value created by innovations of the past, including railroads, airlines, and the Internet, went to the companies that built these networks. The rest went to the consumers and businesses that used the inventions. Most of the 50,000 startups of the dot-com boom were wiped out in the eventual bust, destroying wealth for their investors; around ten survive today as tech giants.

AI faces the prospect of a similar crunch in a wider field. With 70,000 AI-related startups in the United States and abroad, as the tech strategist Jeff McMillan recently put it, “the coming massacre of AI companies and the coming growth of the AI economy are the same event.” But caution fades before a bubble reaches peak size. AI spending is spreading worldwide, with roughly 40 percent of all outlays by U.S. hyperscalers going into projects abroad. Cumulative global AI-related investment is expected to more than double again, to $6 trillion by 2030.

In the biggest bubbles, investment reaches a tipping point after which the cost of the bust in failed businesses, bad debts, and job losses outweighs the benefits of the boom. Even companies at the vanguard of the new technology end up investing too heavily to earn a return. The Bank for International Settlements, a consortium of the world’s major central banks, warned in July that the hyperscalers’ spending race was on track to blow past this “socially optimal level” of investment.

The AI investment boom rests on hidden subsidies and debts, which makes it more fragile and harder to measure. A few years ago, the top five hyperscalers were still generating large piles of surplus cash flow; today, only Microsoft does. Four of them have seen an increase in net debt, issuing bonds as they run low on cash and have to borrow to fund their buildouts.

As of July, U.S. tech firms still had a relatively small reported net debt burden of $250 billion. But these visible figures do not include projects the hyperscalers are financing through complex leases and other off-the-books deals, which could add around $3 trillion to their borrowing over the next decade.

More support is coming from the public sector. During the late 1990s, the U.S. Federal Reserve kept monetary policy tighter than it does now, and the U.S. government ran a surplus. The money that financed the Internet boom came almost exclusively from the private sector.

A boom can be a bubble even if corporate earnings are strong.

Everything changed over the next decade, particularly after the 2008 global financial crisis. To revive and sustain growth, the Fed markedly loosened monetary policy—and kept it loose even when good times returned. Today, the Fed is holding its basic lending rate a full point below what standard academic guidelines would advise, given today’s low unemployment and solid growth. Inflation has exceeded the Fed target of two percent for more than five years running.

Meanwhile, the U.S. government deficit has doubled over the last decade to more than six percent of GDP, in effect transferring billions in borrowed money to households and corporations. According to the Kalecki Levy equation, which tracks the flow of government largess into corporate profits, the budget deficit has accounted for more than half of U.S. corporate profits this decade—twice the level of the dot-com era. Without sky-high deficits, AI earnings would not look nearly as strong as they do. And if AI earnings are not what they appear, the industry’s ability to manage a fast-rising debt burden is likely to be shaky, as well.

It’s not unusual for bubbles to hide behind murky accounting, but the scope of these hidden subsidies is new. So are the sources. The torrent of money pouring out of the government, the central bank, and the private sector will make the AI bubble much bigger than any of its predecessors.

Debates over how the current AI mania might end tend to focus on scenarios in which AI firms cut back on investment, report surprisingly weak earnings, or appear to concede one way or another that AI is not going to pay off richly or soon enough to justify current stock prices. Yet bubbles do not burst entirely of their own accord; investment crazes don’t wane because fundamentals suddenly change. Earnings can remain strong as bubbles inflate, and in a downturn, stock prices typically begin to fall several months before earnings do.

The one consistent signal comes from the financiers. Bubbles pop when liquidity tightens, meaning higher rates or stiffer regulation make money harder to borrow, and private creditors grow less willing to lend at any rate. As money dries up, firms and investors are left with less money to invest. This can happen in myriad ways.

Over the past century, central bank rate hikes have preceded the popping of every significant bubble. Now, however, it is governments, not corporations, who shoulder the largest burden of rising debt, and the most important players to watch are no longer central bankers. They are the “bond vigilantes,” investors who periodically punish free-spending governments by selling their bonds, pushing long-term interest rates to painfully high levels. Of late, vigilantes have been rearing their heads in the United States, Europe, and Japan. If they return in force, rising rates and the resulting liquidity squeeze will make it much harder for AI firms to keep funding their expansion plans through heavier borrowing, and could send the stock market plummeting.

This is how the endgame unfolds. Problems surface first in the debt markets, then spill over into other financial markets. Symptoms started to appear this summer in the U.S. and Asian markets with tremors in semiconductor shares, but other AI stocks may be at imminent risk. It’s hardly a surprise, then, that a research note from Bank of America recently warned that “bubble-like instability” was spreading worldwide.

Every boom sows the seeds of its own demise. Big tech stocks had been rising since 2009, but began to falter four years ago, as the Fed aggressively hiked rates. The release of ChatGPT and the end of Fed tightening then gave the tech bull a second wind. As the Bank for International Settlements noted, “The larger the boom, the larger the eventual bust.”

Each bubble since the crash of 1929 has been larger than those of the past. One reason is the financialization of the economy. Financial markets were roughly equal to global GDP in 1980, but nearly three times as large in 2000, when the Internet boom went bust. Today, they are five times larger. As markets have grown to dwarf the economy, the potential fallout of market crashes has grown with them. A 25 percent drop in the S&P 500 index, for example, would be about half the decline of 2001 or 2008 but would generate equally large losses for American households.

When a tech bubble deflates, the fallout tends to follow a standard trajectory. Lenders pull back, firms at the heart of the boom cut costs or close, bankruptcies pick up, building projects are shelved. Laid-off workers compete for jobs elsewhere, depressing wages in other industries. Some economies slow; a few may tip into recession. The economies that gained most at the height of the AI frenzy would be the first to suffer some combination of these effects. The shock would radiate from the United States to Japan, South Korea, and Taiwan, then down the AI stack through smaller suppliers, and eventually to the miners and power producers. Every burst bubble causes economic damage. This one would be felt around the world.

And the fallout could be magnified by distinguishing features of the AI era. To an unusual degree, growth in economies up and down the AI stack is driven by the twin investment and market booms. Rising stock prices are in turn having an inspiring “wealth effect” on consumer spending, as households feel richer and spend more when the value of their investments rises. This narrow dependence on AI could amplify the shock when the bubble deflates.

Another new twist is the debt backdrop. Never has the world entered a crisis of any kind carrying a debt burden this large. Driven mainly by government borrowing, total global debt spiked over the last year at the fastest pace since the COVID-19 pandemic, to more than $350 trillion. That is more than three times the size of the world economy. Once the AI boom ends, few governments will have the money to stimulate battered economies or bail out affected businesses.

Any attempt to rescue AI companies would meet serious resistance. Most Americans say they are more concerned than excited about AI. Seven in ten expect it to leave fewer jobs for humans. Data center projects across the United States now face determined community opposition. Local revolts blocked or delayed projects worth some $130 billion in the first quarter of this year alone. Politicians may read those numbers and think twice before stepping in to rescue faltering AI firms.

Each bubble since the crash of 1929 has been larger than those of the past.

The worst bubbles form in industries that offer little in the way of innovation and are financed mainly by debt. The real estate bubble of the 2000s is a classic example. When it burst, the result was the long and painful Great Recession of 2008–9. Overextended lenders and borrowers took years to dig their way out of trouble, and the crash left behind nothing of value—no new inventions, no new industries, just deflated home prices. By contrast, the dot-com boom was a classic “good bubble” in a productive industry, financed by equity rather than debt. After it popped in 2001, the drop in stock prices was painful but fleeting. With debts low, borrowers did not struggle for long, and the resulting recession was short and mild. And this time, the crash left behind something of tremendous value: the Internet.

AI is a hybrid. It is a “good bubble” in an innovative technology financed by huge flows of both equity and debt. The fallout from the most recent big bubbles offers no clear framework for estimating the immediate economic losses of such a crash. This is uncharted territory, complicated by the fact that overextended governments cannot easily step in to stem the losses.

When good bubbles burst, they leave behind a streamlined network of stronger survivors, as was the case with the Internet and the railroad, canal, automobile, and television industries before it. Survivors of the AI crash will push the technology forward to the benefit of users worldwide and, in the rosiest scenario, provide the hoped-for boost to productivity growth.

The question is how big this boost might be—and when it will materialize. Often, new technologies raise productivity growth only decades after they appear in commercial form. Cars replaced horse carriages on the streets of New York in the 1910s but made their maximum impact on productivity only after the construction of the national highway system in the 1950s—decades after the market crash of 1929 wiped out hundreds of automakers, leaving only Chrysler, Ford, and General Motors standing.

A similar shakeout and delay could occur before artificial intelligence makes its mark. The most recent forecasts for AI’s potential boost to annual productivity growth range from 0.1 percentage points, which would barely be noticeable, to nearly 2.0 percentage points, which would be extraordinary. An increase that large would roughly double the sustainable economic growth rate worldwide, ushering in a golden age of prosperity at astonishing speed.

Earlier stages of the digital revolution offer reasons to temper those expectations. The Internet was also expected to spark a productivity boom and did so, briefly, in the late 1990s. But that boom eventually faded, overwhelmed by larger forces including the rise of the regulatory state and increasingly generous government bailouts. An AI-driven productivity boom would face similar headwinds.

That AI is advancing at a dizzying pace does not mean it is advancing outside history or the norms of human behavior. Manias create imbalances too large to last. When the crash comes, the damage will be severe but not permanent. To borrow a line from chemistry, nothing is created, nothing is destroyed, everything is transformed. The technology will keep moving forward even as the speculative frenzy around it fades, leaving behind a more balanced global marketplace, more rational spending by companies, and a reordering of the crazed AI world that exists today.