Is AI a Bubble? What 400 Years of Manias Reveal

First, What Is a Bubble, Really?

The word gets thrown at any price that feels too high, but that is not what it means. A high valuation is not a bubble. A bubble is when the price of an asset separates from any reasonable estimate of the cash that asset can produce, and keeps rising anyway, propelled by the belief that someone else will pay more tomorrow. Tulips were not worth a craftsman's annual wage because of the flowers they would grow. They were worth it because the next buyer was sure to pay more. Until, suddenly, there was no next buyer.

Economists have spent a century describing the shape this takes, and three frameworks come up again and again. Understanding them is the difference between reacting to a headline and reading a market.

The economist Hyman Minsky argued that stability itself breeds instability. During long stretches of calm and rising prices, the way deals get financed quietly degrades through three stages. First comes hedge finance, where borrowers earn enough cash to cover both interest and principal. Then speculative finance, where they can cover only the interest and must keep rolling the debt over. Finally Ponzi finance, where the cash flow covers neither, and the whole structure depends on the asset price continuing to rise. The instant prices stop rising, the Ponzi borrowers are forced to sell, which pushes prices down, which forces more selling. That tipping point now carries his name: the Minsky moment.

Charles Kindleberger, who studied centuries of financial chaos in his classic Manias, Panics, and Crashes, broke the cycle into five acts: displacement (a genuinely new technology or a flood of cheap money shifts what people believe is possible), boom (credit expands and prices climb), euphoria (the mania phase, when the public piles in and "this time is different" becomes gospel), distress (the smart money quietly heads for the exits), and revulsion (the crash, as everyone runs for the same door at once).

Geographer Jean-Paul Rodrigue distilled the emotional journey into the chart above: a stealth phase where smart money quietly accumulates, an awareness phase where institutions follow, a mania phase where the public floods in and a "new paradigm" narrative takes hold, and a blow-off phase where the bubble bursts and runs through denial, fear, capitulation, and despair. Notice that the story almost never identifies its own ending in advance. At the very top, the dominant feeling is not fear. It is the serene conviction that the old rules no longer apply.

Four Centuries of Manias: A Field Guide

The most useful thing about bubbles is that we have a long, well-documented record of them. The technology changes, the asset changes, the decade changes, but the arithmetic of the aftermath is stubbornly similar. Here is the tour, with the receipts.

Tulip Mania (1637). The original. At the peak of the Dutch frenzy, a single rare bulb could trade for several times a skilled worker's yearly income. When confidence broke in February 1637, prices collapsed to as little as 1 to 5 percent of their highs, a decline of more than 95 percent. The bulbs were always just bulbs. The price was a story about the next buyer.

The South Sea Bubble (1720). Shares of Britain's South Sea Company rocketed from around £130 to nearly £1,000 in a matter of months on the promise of trade riches that never materialized, then crashed about 80 to 85 percent by December. Among the ruined was Sir Isaac Newton, who lost a fortune and reportedly sighed, "I can calculate the movement of the stars, but not the madness of men."

Railway Mania (1840s). Britain went mad for railway stocks, the disruptive technology of its day. The index of railway shares roughly doubled, then fell about 66 to 70 percent from peak to trough as it became clear that far more lines had been promised than could ever pay. The railways themselves were real and transformative. The share prices were not.

The Crash of 1929. The Dow Jones Industrial Average peaked at 381 in September 1929 and bottomed at 41 in July 1932, a staggering 89 percent decline that helped usher in the Great Depression. It did not reclaim its 1929 high until 1954, a full quarter century later. This is the benchmark every other crash is measured against.

The Nifty Fifty (early 1970s). A group of fifty blue-chip growth stocks, household names like Coca-Cola, Disney, and Xerox, were considered so dependable you could buy them at any price. At the 1972 peak they traded at more than double the market's earnings multiple. In the 1973 to 1974 bear market the S&P 500 fell roughly 45 percent, and the favorites fell far harder: Xerox lost 71 percent, Avon 86 percent, Polaroid 91 percent. "Quality at any price" turned out to have a price after all.

Japan's Nikkei (1989). The most sobering case for anyone who assumes markets always bounce back quickly. The Nikkei 225 peaked at 38,915 on the last trading day of 1989, then fell nearly 80 percent over the following years. It did not return to that level until February 2024, a recovery that took 34 years.

The Dot-Com Crash (2000). The closest historical rhyme to today, and we will return to it. The Nasdaq Composite hit 5,048 in March 2000, then fell about 78 percent to its 2002 low as a generation of profitless internet companies evaporated. The index did not set a new high until April 2015, 15 years later.

The Global Financial Crisis (2008). A housing and credit bubble, not a tech one. The S&P 500 fell about 57 percent from its October 2007 peak to its March 2009 low, and US home prices dropped roughly 30 percent nationally. The market recovered to new highs by 2013, relatively quickly, in part because central banks intervened on a scale never seen before.

China 2015 and Crypto. More recent reminders that the pattern never retires. China's Shanghai Composite shed about half its value after its 2015 peak. Bitcoin fell 84 percent after its 2017 top and about 78 percent after its 2021 top, though crypto's recoveries have come in years rather than decades.

Who Is Actually Driving the AI Boom?

Before we judge whether this is a bubble, it helps to know who is actually in it. The AI boom is not an abstraction. It is a specific cast of companies and a very specific flow of money, and the names have gotten enormous, fast.

The labs building the models

At the center sit the AI labs. OpenAI, maker of ChatGPT and the GPT-5 family, has confidentially filed to go public in late 2026 at a valuation between $852 billion and $1 trillion. Microsoft owns roughly 27 percent of it, a stake worth around $228 billion, plus a share of its revenue. OpenAI now takes in about $2 billion a month and still expects to lose money for years.

Its rival Anthropic, maker of the Claude assistant, has quietly leapfrogged it. A fresh raise pushed Anthropic to a $965 billion valuation, the most valuable AI startup in the world, and it now reports more annual revenue than OpenAI does. Amazon and Google are its largest backers, and it too is preparing for an eventual IPO. Then there is xAI, Elon Musk's lab behind the Grok assistant, which merged into SpaceX in early 2026 to form a roughly $1.25 trillion private giant, now the most valuable private company on earth. Musk's stated ambition runs past chatbots to "orbital data centers," AI infrastructure launched into space to power future missions. Google with Gemini and Meta with its new Superintelligence Labs round out the American field.

The race is global, and that matters. China's DeepSeek stunned markets in early 2025 by matching top US models at a fraction of the cost, and Alibaba's Qwen now leads the world in open-source downloads. Europe's Mistral is in the mix too. If a cheaper model can do the same job, much of the spending below may never be needed, and that is exactly the kind of shock that pricks a bubble.

The picks and shovels

Every model needs chips, and this is where the real money has flowed. Nvidia, whose GPUs train nearly every major model, is worth more than $5 trillion. TSMC, which actually manufactures the chips, sits above $2 trillion, as does Broadcom, which designs custom AI silicon. AMD trades around $760 billion.

The wildest moves, though, are in memory and storage, the corner of the boom most people miss. AI systems are starved for high-bandwidth memory, and suppliers cannot make it fast enough. Micron (ticker MU) has climbed roughly 150 percent in a year toward a $900 billion market value, its advanced memory sold out for quarters. SanDisk (ticker SNDK), spun out of Western Digital in early 2025, has rocketed more than 1,000 percent as its flash storage became the backbone of AI data centers. When a once-sleepy memory chipmaker rises elevenfold in a year, you are looking at either a genuine supercycle or a textbook mania, and reasonable people disagree about which.

Tying it all together is a web of cloud and capital. Microsoft, Amazon, and Oracle rent out the computing power, and the AI-cloud specialist CoreWeave, backed by Nvidia itself, sits on a $100 billion order backlog. The result is staggering concentration. According to JPMorgan, just four chipmakers and four cloud giants have swelled from a combined $3 trillion seven years ago to about $18 trillion today, roughly a fifth of the entire developed world's stock market. When that much value rests on so few names all betting on the same story, a stumble by any one of them stops being a company problem and becomes a market problem.

The Case That AI Is a Bubble

Now to the present. Take the skeptics seriously, because the strongest version of their argument is built from numbers that are not in dispute.

Valuations are at historic extremes. The Shiller CAPE ratio, which compares prices to a decade of inflation-adjusted earnings to smooth out the noise, hit 42.6 in late May 2026. It has only been higher once in 140 years: 44.2, in December 1999, weeks before the dot-com peak. The Buffett Indicator, the total value of US stocks divided by GDP, sits near 236 percent, against roughly 146 percent at the 2000 top. By that gauge, the market is more stretched today than it was on the eve of the most famous tech crash in history.

The market is dangerously concentrated. A handful of mega-cap technology companies now drive the bulk of the S&P 500's returns and earnings. In the first quarter of 2026, three companies alone, Alphabet, Amazon, and Meta, accounted for roughly 71 percent of the entire index's earnings growth. When an index this large leans on this few names, a stumble by any one of them is no longer a company problem. It is a market problem.

Spending is sprinting ahead of revenue. This is the heart of the bear case. The four largest hyperscalers plan around $725 billion in capital spending in 2026, up an extraordinary 77 percent from the prior year, and Goldman Sachs models a baseline of $7.6 trillion of AI infrastructure spending through 2031. Against that torrent, the revenue is thin. The venture firm Sequoia framed it as the "$600 billion question": the annual revenue the industry would need to justify its spending, a gap that has only widened. OpenAI now generates roughly $2 billion a month, and even Anthropic, which recently passed it as the revenue leader, brings in only about $47 billion a year against more than $700 billion of annual spending. The arithmetic does not yet close, and free cash flow at the largest spenders is projected to turn negative for the first time in 35 years.

The money may be moving in a circle. Skeptics point to a web of deals in which the same dollars appear to fund demand for the companies that paid them. Nvidia agreed to invest up to $100 billion in OpenAI, which in turn commits to buy enormous quantities of Nvidia chips. Similar arrangements link OpenAI to Oracle (around $300 billion), to Microsoft (a $250 billion cloud commitment), and to AMD. When a supplier funds its own customers, reported demand can look stronger than underlying demand actually is.

The accounting may flatter profits. Investor Michael Burry, who famously shorted the 2008 housing bubble, argues that hyperscalers stretch the depreciation of their chips over five or six years when the real useful life, given how fast Nvidia iterates, may be closer to two or three. He estimates this could overstate industry profits by roughly $176 billion between 2026 and 2028. Nvidia disputes this, noting that older chips remain in heavy use. The point for our purposes is simply that the profits underpinning the bull case rest partly on an accounting assumption that reasonable people contest.

Physics may cap the dream. Even if demand is real, the electricity may not be there. Global data-center power demand is set to exceed 1,000 terawatt-hours in 2026, roughly double its 2023 level, and analysts at Gartner warn that power shortages could constrain as much as 40 percent of AI data centers by 2027. You cannot run a trillion-dollar buildout on a grid that physically cannot deliver the electrons.

The Case That It Is Not a Bubble

Now the other side, with equal seriousness, because it is not a strawman. Some of the most experienced investors alive look at the same market and see something structurally different from 1999.

This time, the leaders actually make money. The defining feature of the dot-com era was companies with no profits and, in many cases, no revenue, trading at fantasy valuations on stories alone. Today's AI leaders are among the most profitable enterprises in human history. Microsoft, Alphabet, Amazon, Meta, and Nvidia generate enormous real earnings and throw off mountains of cash. A high valuation on a hugely profitable company is a very different animal from a high valuation on a company that has never earned a cent.

The buildout is funded by cash, not debt. The telecom companies of 1999 built on borrowed money and collapsed when the debt came due. Today's hyperscalers are largely funding their spending out of operating cash flow, from balance sheets carrying far less leverage than the doomed "high spenders" of 2000. The asset manager Allianz, which supplies some of the most-cited bear-case statistics, nonetheless concludes that the buildout is "war-proof for now" and that solvency risks are "overdone," precisely because the spenders are so financially strong.

The technology is already generating real value. Unlike many dot-com products that had no users, AI tools have been adopted by hundreds of millions of people and tens of thousands of businesses within a few years. The demand for computing power is not theoretical: the world's most advanced chips and memory are sold out. High-bandwidth memory is spoken for through 2026, and DRAM prices jumped 90 percent in a single quarter. That is what genuine scarcity, not speculative hoarding, looks like.

Maybe a trillion dollars is just the new baseline. Here is the most uncomfortable possibility for the bears. What if the old valuation yardsticks are miscalibrated for a handful of globally dominant, capital-light, wildly profitable platforms? If a company can grow earnings at double digits almost indefinitely on the back of a genuine technological shift, a valuation that looks insane on a 1990s spreadsheet might simply be the market repricing what a dominant business is worth in the 21st century. The four most dangerous words in investing are "this time is different." But they are dangerous precisely because, occasionally, they are true.

So, Are We in a Bubble?

Here is where an honest article refuses to perform a confidence it has not earned. The verdict is genuinely unresolved, and the most telling evidence is this: the single report most often quoted by the bears, the Allianz analysis with its alarming statistics on spending and cash flow, is the same report that concludes the boom is sturdy and solvency fears are overblown. The best-informed analysts are looking at identical data and reaching opposite conclusions. That is not a failure of research. It is the actual state of the world.

What we can say with confidence is which bubble symptoms are present and which are absent. The valuations are unmistakably at bubble-era extremes. The concentration is extreme. The spending is outrunning revenue in a way that has historically ended badly, and the circular financing is a genuine yellow flag. Those are real.

But the central feature of the most catastrophic bubbles, an edifice of profitless companies built on borrowed money, is largely absent. The leaders are real businesses earning real money, funding themselves with real cash. That does not make them immune. Cisco was a real, profitable business in 2000, and its stock still fell about 88 percent and has never fully recovered, because the price had simply run too far ahead of even a great company's prospects. A bubble does not require fraud or fiction. It only requires a wonderful thing priced as if nothing could ever go wrong.

So the truthful answer to "is AI a bubble" is: the market is pricing AI for a future of nearly flawless execution, and history's verdict on that kind of pricing is unkind even when the underlying revolution is real. Whether that makes today a bubble or a new baseline depends entirely on a number nobody yet knows, which is how much cash all this computing will actually produce, and how soon. The technology is almost certainly transformative. Whether the stocks are correctly priced for it is a separate question, and it is the only one that determines whether you make or lose money.

How to Navigate a Bubble, Whichever It Turns Out to Be

Here is the liberating part. You do not actually need to win the "is it a bubble" debate to handle it well. Some of the most legendary investors in history got the call right and still lost, because being early is indistinguishable from being wrong when the bills come due. Two cautionary tales from the year 2000 say it all.

Julian Robertson, who ran the storied Tiger Management, correctly judged that internet stocks were absurd and positioned against them. He was right. But the market stayed irrational longer than his investors could stay patient, and he was forced to close his fund in March 2000, at the exact moment the bubble finally peaked. He was vindicated within weeks, and his investors were not around to see it.

Stanley Druckenmiller, running money for George Soros, also knew it was a bubble. So he did the opposite of Robertson: he kept buying technology stocks into the top, chasing the gains, and got caught in the collapse. "It would have been nice to go out on top, like Michael Jordan," he later admitted. "But I overplayed my hand." Two of the sharpest minds of their generation, both correct about the bubble, both wrecked by their timing in opposite directions. The lesson is not "be smarter." It is "build a process that does not require perfect timing."

That is the entire point of risk management, and it long predates AI. Here is how disciplined traders have historically survived markets exactly like this one.

For traders who use options, this is also where defined-risk structures earn their keep. A protective put can cap the downside on a position you want to keep holding, and strategies with strictly limited risk let you express a view on a volatile name without exposing yourself to its full, brutal range. The tools matter less than the principle behind them: decide where you are wrong before you put the trade on, and let that decision, not your emotions at the top or the bottom, run the exit.

Frequently Asked Questions

Is AI a bubble in 2026?

It is genuinely contested among serious analysts. Valuations are at bubble-era extremes and AI spending is sprinting ahead of AI revenue, which are real warning signs. But unlike the dot-com era, today's AI leaders are highly profitable and fund their spending with cash rather than debt. The honest answer is that the market is priced for nearly flawless execution, and whether that proves to be a bubble depends on how much revenue AI actually generates in the next few years.

How is the AI boom different from the dot-com bubble?

The biggest difference is profitability. Many dot-com companies had little revenue and no profits, and the telecom buildout was financed with debt. Today's AI leaders are among the most profitable companies in history and largely fund their data-center spending from operating cash flow. The valuations rhyme with 2000, but the financial foundations are far stronger. The main echo is that spending is racing ahead of demand, just as it did in the 1999 to 2000 fiber buildout.

What would signal that the AI bubble is bursting?

Watch the gap between spending and revenue. The warning signs would be AI revenue growth stalling while capital spending stays high, hyperscalers pulling back their spending guidance, cracks appearing in the circular financing deals, or a broad loss of faith that triggers forced selling, the classic Minsky moment. A sharp break in the leading chip and infrastructure names on heavy volume would likely be the first market signal.

How far do bubbles usually fall, and how long do they take to recover?

Historically, major bubbles have fallen anywhere from about 50 percent to nearly 90 percent from peak to trough. Recovery time varies enormously and is not tied to how real the technology was. The 2008 crash recovered in about five years, the dot-com Nasdaq took 15 years, and Japan's Nikkei took 34 years to reclaim its 1989 high. The depth of the fall and the length of the recovery are two separate risks.

How can I protect my portfolio if AI is a bubble?

This is educational, not personalized advice, but the time-tested principles are: avoid betting more on any single idea than you can afford to lose, take partial profits as positions rise rather than holding for the exact top, use trailing stops so an exit is decided by rules rather than emotion, keep some safe assets as ballast, and resist averaging down into a crash. The aim is to survive whichever way the market breaks, not to predict the top.

Can the market keep going up even if it is a bubble?

Absolutely, and that is the trap. As the economist John Maynard Keynes is credited with warning, markets can stay irrational longer than you can stay solvent. Bubbles routinely run far higher and far longer than skeptics expect, which is why shorting them early has destroyed even brilliant investors. Participating with discipline and a clear exit plan has historically been far safer than trying to call the top.