The AI bubble debate, set against 10 market bubbles since the 1630s

OpenAI launched ChatGPT on November 30, 2022, and interest in artificial intelligence took off. Investors then bid up shares of companies seen as likely to profit from AI. Nvidia $NVDA, the main supplier of chips to AI developers, became a focal point for that enthusiasm. The AI boom has also drawn heavy spending on data centers and the hardware inside them.

A technology can change the world and still be overpriced. That tension sits at the center of the AI bubble debate, and it is not new. Railways reshaped Britain in the 19th century, yet many of the companies formed to build them failed. The internet became part of daily life even as many early dot-com companies vanished.

Whether today's prices reflect future earnings or excess optimism cannot be settled in advance. Bubbles are usually identified with confidence only after they burst. What history can offer is comparison. Past manias show which patterns repeat, which ones differ and where a real innovation and an inflated market have existed side by side.

This list covers 10 market bubbles in chronological order. It starts with Dutch tulip contracts in the 1630s and moves through John Law's Mississippi scheme in France and the South Sea Company in Britain. It then covers British railway mania, the 1929 U.S. stock bubble, the Nifty Fifty, Japan's 1980s asset bubble and the dot-com era. It ends with the U.S. housing bubble and the crypto boom and bust, which included the bankruptcy of the exchange FTX in 2022.

Each slide explains how one bubble formed and draws one specific comparison with the current AI rally. Four questions run through the piece. What story justified the prices? How was the buying financed? How concentrated was the money in a few companies? And how long did it take prices to recover once the mania ended?

Some comparisons point to similarities with AI stocks, and others point to sharp differences. None of them amounts to a forecast. Together they give readers a way to judge the AI boom against market history from the 1630s to 2022.

Illustration: Quartz

Tulip mania is the earliest bubble on this list. Its lesson for the AI boom is that the size of a mania can stay in dispute long after prices fall. Prices for tulip bulb contracts in the Netherlands collapsed in early February 1637, and historians still argue over how much that crash mattered.

The collapse hit contracts for bulbs, not flowers already in bloom. Buyers agreed to prices for bulbs they would take possession of later. When confidence broke that February, those agreed prices no longer held, and the market for the contracts fell apart.

The Semper Augustus is the variety most often named as the costliest of the period. Its name appears in many retellings of the episode. The figures attached to it differ between accounts. Because the sources disagree, no single price for a single bulb can be stated with confidence.

That uncertainty sits at the center of the modern debate. The historian Anne Goldgar has argued that the economic fallout from the collapse was limited. Popular retellings often present tulip mania as a frenzy that gripped a whole country. The two pictures are hard to reconcile, and historians have not settled the question.

The gap between legend and record is what links tulip mania to the AI debate. A bubble story is simple to tell and simple to repeat. A dramatic version can travel further than the documented one, and it can harden into accepted fact.

Debate over an AI bubble runs into the same problem. A vivid historical comparison can make a case sound convincing without proving anything about current prices. Anyone weighing the AI boom against tulip mania is comparing it with an episode whose own size remains contested.

The record from 1637 raises a set of questions that can be put to any boom. Who was actually buying? What were they paying for? How far did losses spread when prices fell? Those questions apply to AI stocks as much as to bulb contracts. The tulip case shows how hard they can be to answer, even centuries later.

Illustration: Quartz

The Mississippi Bubble was a French share mania built on promised profits from colonial Louisiana, and it collapsed during 1720. Its lesson for the AI boom is that a persuasive story about a new frontier can carry a share price well beyond what a company earns today.

The scheme was the work of John Law, a Scottish financier who ran two linked institutions in France. One was the Banque Générale. The other was the Mississippi Company, formally known as the Compagnie d'Occident.

Shares in the Mississippi Company drew their appeal from two sources. The first was France's colonial trade in Louisiana, a territory on the other side of the Atlantic. The second was the company's role in handling the French state's debt. That role tied the company's fortunes to the government's own finances.

Louisiana stood for new and untapped wealth. Buying the company's shares was a way to own a piece of that trade. Buyers expected large returns from it in the years ahead. What the company earned at the time was small next to the expectations priced into its shares.

The scheme's collapse in 1720 left shareholders who had bought on the Louisiana story holding shares worth far less than they had paid.

The parallel with AI lies in the structure of the bet rather than in the technology itself. Investors value many AI companies on what the technology may earn in the years ahead, not only on current profits. In both cases, the price reflects a forecast about a frontier that has not yet been fully developed.

Any frontier story asks investors to weigh a specific risk. A new territory or a new technology can be real and still fail to produce profits fast enough to justify the price paid for it. The question for AI stocks is the one Law's shareholders faced: whether earnings eventually catch up with the story.

The comparison also has limits. Louisiana was an ocean away from the investors in France who bought into it, and few of them could judge its prospects firsthand. AI tools, by contrast, are products that businesses and consumers can use and evaluate directly. The AI boom also involves several competing companies rather than one venture run by a single financier.

Illustration: Quartz

The South Sea Bubble was the 1720 collapse in shares of Britain's South Sea Company, a firm bound up with the government's debt. Its stock surged in the first half of 1720 and then fell hard in the autumn. The episode shows how official backing and rising prices can feed each other.

The company was founded in 1711. Its central role was financial. It took on part of the British government's debt, which tied its fortunes to public finance. That link gave the shares an air of safety. A company doing business with the state looked less risky than an ordinary trading venture.

By 1720, that sense of safety had given way to a buying rush. Many investors bought for a simple reason: the price kept climbing. Each rise drew in new buyers, and their money pushed the price higher still.

The reversal came in the autumn, and it was steep. Isaac Newton was among the investors who lost money. His standing as a scientist did not protect him once prices turned.

Parliament passed the Bubble Act the same year. The law restricted the formation of joint-stock companies. It limited how freely new share-issuing ventures could be set up.

The AI comparison has two parts. The first concerns the link between a market favorite and the state. The South Sea Company's appeal rested partly on its close ties to government finance. Investors weighing AI stocks can ask a parallel question. How much of a company's value rests on policy support or official favor, rather than on what it sells to customers?

The second part is momentum. Buying because prices are rising was the engine of the South Sea rally. A price driven mainly by its own rise has little beneath it once the rise stops.

That test can be applied to the AI rally. Whether AI shares fit the 1720 pattern depends on what buyers are paying for: the revenue behind the companies, or the price trend itself.

Illustration: Quartz

British railway mania, which peaked around 1845 to 1847, showed that a technology can transform an economy while many of the companies selling it fail their investors. Railways were real and lasting, but many of the companies promoted to build them never laid track or never paid.

The mania was a speculation in railway shares. Buyers were acquiring stakes in specific companies, not in the general idea of rail travel. That distinction sits at the center of the episode. A railway line could be a sound idea, and the company promoting it could still be a poor investment.

Some of the companies never got past the planning stage. Their lines were never built, so shareholders held a claim on a project that produced nothing. Other companies did build but never paid a return to the people who financed them. In both cases, investors lost money even as rail transport proved its worth.

The episode separates two questions that tend to blur during a boom. The first is whether a technology will matter. The second is whether the price paid for a particular company reflects what that company will earn. A clear yes to the first question does not answer the second.

The current AI boom raises the same split. Generative AI tools are in commercial use, and companies along the supply chain sell real products, from chips to software. The railway case shows that the usefulness of a technology does not, on its own, settle whether any given AI stock is priced sensibly.

The 1840s also show what can happen when capital flows freely toward a popular technology. Promoters can float more companies than the market can support. Investors may struggle to tell which ventures will deliver. Whether a similar gap opens among AI companies is a question that only their results over time can answer.

For investors, the railway era offers a narrow lesson. Backing a transformative technology and backing a profitable company are different bets. The 1840s are an early record of how far apart those two bets can land.

Illustration: Quartz

The Dow Jones Industrial Average peaked at 381.17 on September 3, 1929. It did not close above that level again until November 1954. That wait of about 25 years is the clearest measure of what the 1929 bust cost investors who bought at the top.

The decline itself was fast. The Dow hit its low of 41.22 on July 8, 1932, less than three years after the peak. That was a fall of about 89% from the 1929 high.

A drop that deep changes the arithmetic of recovery. An investor holding $100 of the index at the peak would have held about $11 at the 1932 bottom. To get back to even from there, the index had to rise by roughly 825%. A loss of 89% and a gain of 825% describe the same distance, measured from opposite ends.

The climb back from the low took about 22 years. The slide had lasted under three. The recovery took more than seven times as long as the collapse that made it necessary.

The 25-year span refers only to the index's price level. The Dow is a price index, so its headline number leaves out the dividends its member companies paid along the way. An investor who collected and reinvested those payments would have regained the 1929 value sooner than the bare index figure suggests.

For AI stocks, the 1929 record speaks to duration rather than direction. It does not say whether today's AI share prices are too high or too low. It shows something narrower. When a market climbs to a peak and then breaks, the time needed to regain that peak can stretch across decades.

The timing of a purchase shaped the result. Someone who bought the Dow in September 1929 needed about a quarter century to break even on the index price. Someone who bought near the July 1932 low held the same stocks but started from a far lower price.

A market decline has two dimensions: how far prices fall and how long they stay down. Coverage of a crash usually captures the first. The Dow's path from 1929 to 1954 shows that the second can matter just as much to the people holding the shares.

Illustration: Quartz

The Nifty Fifty were about 50 large, well-known U.S. growth stocks that investors in the late 1960s and early 1970s treated as safe to hold at any price. They fell sharply in the 1973–74 bear market, showing that crowding into a short list of trusted companies offered no protection.

The group never had an official membership list. Polaroid and Xerox are among the names most often counted in it. Both were household brands built on products people used every day. Polaroid sold instant cameras that produced a print within minutes. Xerox became so closely tied to photocopying that its name turned into a verb.

The trade rested on one assumption. If a company's earnings would keep rising for decades, the price paid today seemed almost irrelevant. That idea earned the group another nickname: "one-decision" stocks. The only decision an investor supposedly needed to make was to buy. Selling was never meant to be necessary. Investors could hold and stop worrying about valuation.

That reasoning failed once the wider market turned, and a reputation for safety did not stop the losses.

The companies were not frauds or empty promises. Their products had real customers. The weakness lay in the price investors paid. A strong business can still be a poor investment when buyers pay so much that years of future growth are already built into the share price.

That is the clearest link to the AI boom. Much of the market's enthusiasm for artificial intelligence has gathered around a small number of very large companies. Nvidia $NVDA, the leading supplier of AI chips, sits near the center of that group. When a few companies carry most of the hopes for a new technology, a broad index can come to depend heavily on how they perform.

The parallel has limits. The Nifty Fifty spanned different industries, from instant cameras to office copiers, rather than a single technology theme. The AI trade is built around one technology and the hardware and software that support it. A setback for that one technology could touch all of those companies at once.

The question the episode raises for AI stocks is not whether the companies are real. It is whether their prices assume a future with no room for disappointment.

Illustration: Quartz

Japan's Nikkei 225 stock index peaked at 38,915.87 on December 29, 1989. It did not close above that level again until February 2024, about 34 years later. That wait shows how long a market priced for perfection can take to regain its high.

The bubble reached well beyond shares. Land and stock prices both rose sharply in Japan during the 1980s. Japanese banks commonly lent against real estate. Rising land values supported more borrowing, and that borrowing pushed asset prices higher still.

The same loop ran in reverse after prices turned. Falling land values eroded the collateral behind bank loans. Lenders were left holding debts that borrowers struggled to repay. Japan's slow growth in the 1990s became known as the lost decade.

The length of that recovery is what makes Japan useful for the AI comparison. Investors at the end of 1989 paid prices that left little room for disappointment. Japan's economy did not deliver the growth those prices implied. The index remained below its record through the 1990s, the 2000s and the 2010s.

An overpriced index is not the same as a set of failing companies. Many businesses listed on the Nikkei kept operating through the long slump. Their shares had simply been bought in advance for results that took far longer to arrive than buyers expected.

AI stocks face a version of the same test. Share prices for companies such as Nvidia $NVDA rest heavily on expectations of future demand for AI chips and services. Those expectations may prove right. The Japanese case shows the cost if they run too far ahead of what companies later deliver. An index can then stay below its old high for decades.

The comparison also carries a contrast. Japan's bubble ran through a national economy, across both property and equities. It also ran through a banking system that had lent heavily against land. The AI rally centers on technology companies and their shares rather than on household and corporate real estate. That difference would shape how any correction travels. The Nikkei's history cannot show how far one would reach.

Illustration: Quartz

The dot-com bubble is the historical episode most often set beside the AI boom. In the late 1990s, investors poured money into internet companies and into the networks built to carry their traffic. When the bubble burst, the Nasdaq $NDAQ Composite lost about 78% of its value from peak to trough.

The Nasdaq Composite tracks the stocks listed on the Nasdaq exchange, where many technology companies trade. It peaked at 5,048.62 on March 10, 2000. The decline that followed lasted more than two years. The index bottomed at 1,114.11 on October 9, 2002. It did not regain its 2000 high until 2015, about 15 years after the peak.

Pets.com became the best-known symbol of the era. The online pet supply retailer went public in February 2000, weeks before the Nasdaq topped out. It was liquidated in November 2000. Its run from stock market debut to shutdown fit inside a single calendar year.

The closer parallel to AI sits below the consumer websites. During the dot-com years, companies spent heavily on telecom equipment and fiber-optic networks. The bet was that internet traffic would grow fast enough to fill all that capacity. The build-out ran ahead of the demand that existed for it. Much of that fiber went unused for years after the crash.

The AI build-out follows a similar pattern. Technology companies are spending heavily on data centers and the specialized chips that train and run AI models. Nvidia $NVDA sells those chips, which places it where the network equipment makers stood in the late 1990s. Its sales depend on how much its customers decide to build. The equipment makers faced the same exposure when telecom spending slowed after 2000.

The question the fiber build-out faced now applies to AI infrastructure. If demand for AI services grows into the capacity being built, the spending will have laid the groundwork for a new industry. If demand falls short, the AI build-out will resemble the telecom overbuilding of 2000. The test is how much revenue AI products generate compared with what the infrastructure cost to build.

Illustration: Quartz

The U.S. housing bubble of the 2000s was built on borrowed money, and that is why its collapse spread far beyond homeowners. Lenders made mortgages to borrowers with weak credit, and those loans were bundled into mortgage-backed securities held across the financial system.

Debt sets housing apart from many earlier manias. Most home buyers pay with a mortgage rather than cash. A buyer who purchases shares with savings can lose the money invested. A homeowner who borrowed owes the balance to a lender. If that homeowner stops paying, the loss moves to the lender as well.

Securitization stretched that chain further. A lender could make a loan, sell it and pass the risk of default to whoever bought the resulting security. Banks and other investors held those securities. Mortgage risk had been divided among many securities and many holders. That made it hard to tell in advance which institutions were exposed. When subprime borrowers, those with weak credit histories, stopped paying, the securities lost value. The losses landed on institutions that had never dealt with the homeowners directly.

The spread became unmistakable when Lehman Brothers filed for bankruptcy on September 15, 2008. Lehman was a major Wall Street investment bank. Its failure showed that a housing downturn had become a problem for the banking system itself, not only for households.

The comparison with the AI boom concerns financing more than technology. When a bubble is paid for mostly with shareholder money, falling prices hurt the people who own the shares. When it is paid for with borrowed money, falling prices can also damage lenders. Those lenders in turn supply credit to businesses and households that had no part in the boom.

That distinction applies to AI stocks and to the companies building data centers and chips. The relevant question is not only whether prices are too high. It is who would absorb the losses if prices fell. In 2008, the answer turned out to be banks, and through them the wider economy.