The AI Bubble: What 100 Years of Warnings Actually Show
Paul Samuelson, the Nobel laureate economist, wrote a line in Newsweek in 1966 that still gets quoted: "Wall Street indexes predicted nine out of the last five recessions! And its mistakes were beauties." The same pattern applies to speculative bubbles, only more so.
The more useful question is therefore not who is warning today.
How often warnings came, how often bubbles burst
Right now the Bank for International Settlements, the Bank of England and a growing share of fund managers are warning about an AI bubble. In the BofA Global Fund Manager Survey from July 2026, 43 percent of managers called AI stocks a bubble and 48 percent did not. Google searches for "AI bubble" rose 950 percent year over year in autumn 2025.
How often have people warned about a bubble over the last hundred years, and how often did one actually burst? There is solid data on this from finance research, and it comes out considerably clearer than the tone of the current debate suggests. The short answer: warnings are near-continuous, and very little has burst.
The Goetzmann dataset
The most solid work here comes from William Goetzmann, professor of finance at the Yale School of Management. He analysed price data from 41 markets between 1900 and 2014. That is 3,514 market-years in total.
His definition of a bubble is narrow and testable: a market doubles within one calendar year, then loses more than half its value. The result:
- Four cases where the collapse came in the following year.
- Ten cases when the post-boom window is extended to five years.
Ten cases out of 3,514 market-years. Goetzmann himself cautions against reading an exact frequency into this, since annual data systematically undercounts short cycles. But the order of magnitude stands: the pattern everyone warns about occurs very rarely.
What happens after a sharp run-up is more revealing still. Goetzmann examined 58 cases where a market had doubled within a year:
| App | Time horizon after the boom | Market doubles again | Market loses over 50% |
|---|---|---|---|
| 1 year | Following year | 6.9% | 6.9% |
| 5 years | Following five years | 22.4% | 17.2% |
Over a one-year window, both outcomes are exactly equally likely. Over five years, the market rose again more often than it collapsed. Goetzmann's own phrasing: following a boom, a market is as likely to double again as it is to crash.
Investors overestimate crash risk by a factor of ten
Robert Shiller has asked investors the same question since 1989. What is the probability of a catastrophic stock market crash in the next six months, on the scale of 28 October 1929 or 19 October 1987?
The answers across 1989 to 2015, analysed by Goetzmann, Kim and Shiller:
- Mean: 19 percent.
- Median: 10 percent.
- Historical base rate for the same event over six months: 1.7 percent.
Investors put the crash probability at more than ten times its historical level. This is not a one-off error but a stable pattern across 26 years of surveys.
A common misreading
Shiller's "Crash Confidence Index" is frequently misread. It reports the share of respondents who think a crash is less than 10 percent likely. A high reading therefore signals confidence, not crash expectation.
The four mega-bubbles of the last 100 years
Set the bar at collapses with macroeconomic consequences and four remain. Not forty, not four hundred.
| App | Period | Peak | Maximum drawdown |
|---|---|---|---|
| Wall Street | 1921–1929 | 3 Sep 1929, Dow 381 | −89.2% by July 1932 |
| Japan | 1986–1991 | 29 Dec 1989, Nikkei 38,915 | −80% by 2003 |
| Dotcom | 1995–2000 | 10 Mar 2000, Nasdaq 5,049 | −78% by Oct 2002 |
| US housing | 2002–2007 | 9 Oct 2007, S&P 500 at 1,565 | −57% by March 2009 |
Two details that usually get lost in comparisons. First, the Japanese bubble is conventionally dated 1986 to 1991, but the Nikkei peaked back in late December 1989. Second, in the subprime crisis US house prices peaked in mid-2006 according to the Case-Shiller index, roughly fifteen months before equities. Anyone warning about an equity bubble in 2006 had the trigger right and the timing wrong.
Alongside these were sector-level speculative waves that never turned systemic: the Hunt brothers' silver squeeze in 1980, the uranium bubble of 2007, the crypto cycles of 2017 and 2021, the SPAC and meme-stock boom of 2020/21. They destroyed wealth without causing a recession.
Why every major technology triggers a bubble debate
Peter Oppenheimer, chief global equity strategist at Goldman Sachs Research, analysed 51 significant technological innovations between 1825 and 2000. Railways, electricity, the telephone, radio, the automobile, semiconductors, the internet.
In 73 percent of those cases, a speculative bubble formed in equity prices.
That is the number that actually matters for the AI debate. A bubble forming around a significant new technology is not the exception, it is the base case. It is not evidence that the technology is overrated. It is a side effect of capital moving faster than implementation does.
What happened afterwards is more instructive still. The dotcom crash wiped out hundreds of internet companies. The Nasdaq took fifteen years to reclaim its old high. But the internet did not go away. Amazon fell more than 90 percent and later became one of the most valuable companies in the world. The fibre that telecom firms laid on borrowed money during the boom stayed in the ground and was used cheaply by whoever came next. Railways went the same way: the investors of the railway mania lost their money, the tracks stayed put.
What this means for the AI debate
At Polyfactor in Hamburg we build AI systems for mid-sized companies every day, and the bubble question comes up in almost every first conversation. My answer consists of three statements that are true at the same time and only appear to contradict each other.
There are overvaluations
The ten largest holdings in the S&P 500 made up around 40.7 percent of the index at the end of 2025. At the dotcom peak the figure was roughly 23 percent. Concentration today exceeds that level.
The Bank of England noted in its July 2026 Financial Stability Report that valuations look stretched even when the thirty largest AI stocks are excluded. The Bank for International Settlements pointed out in June 2026 that the five largest hyperscalers have committed capital exceeding their earnings and free cash flow. Direct-lending funds, it added, have quadrupled their exposure to the AI and IT sector to around 15 percent of their portfolios.
Enormous capital is flowing in
For 2026, Amazon has guided to roughly $200 billion in capital expenditure, Microsoft around $190 billion, Alphabet $175 to $185 billion, and Meta $130 to $145 billion. Goldman Sachs models aggregate capex of around $765 billion in one scenario, and the bank explicitly stresses that this is a scenario rather than a forecast. Capital at that scale will not earn its expected return in every case. A correction is likely.
The value created holds
This is the part the bubble debate most often drops. Nvidia traded at a forward price-to-earnings ratio of around 20 in early August 2026, which is not what a classic bubble stock looks like. Revenue in the first quarter of fiscal 2027 came in at $81.6 billion, up 85 percent year over year. Unlike 1999, there are real revenues and real demand behind the valuations.
"Capital markets price expectations, a company prices outcomes. Conflating the two is the most expensive mistake in this debate." — Jamin Mahmood-Wiebe, Founder of Polyfactor
A correction in capital markets and a durable productivity gain are not mutually exclusive. Historically they mostly arrived together. That is precisely the lesson from 73 percent of technological breakthroughs since 1825.
Frequently asked questions
How often has a bubble actually burst in the last 100 years?
Four times at a global level with macroeconomic consequences: Wall Street 1929, Japan 1990/91, dotcom 2000, US housing 2007/08. In Goetzmann's dataset of 41 markets and 3,514 market-years, ten cases met the narrow criterion of "doubling in one year, then halving within five".
How high do investors rate crash risk?
At 19 percent per six-month period on average, with a median of 10 percent (Shiller surveys, 1989 to 2015). The historical base rate for an event of that scale is roughly 1.7 percent. The overestimate exceeds a factor of ten.
Does every new technology produce a bubble?
Not every one, but most. Goldman Sachs Research examined 51 significant innovations between 1825 and 2000, among them railways, electricity, radio and the internet, and found a speculative bubble in equity prices in 73 percent of cases. Bubble formation around major technological breakthroughs is the base case, not the exception.
Is the AI bubble comparable to the dotcom bubble?
Only partly. Index concentration is higher today than in 2000, with the ten largest S&P 500 holdings above 40 percent against roughly 23 percent then. Unlike 1999, however, the leading names are backed by real revenues and strongly growing demand. Nvidia's forward price-to-earnings ratio stood at around 20 in early August 2026.
What happens to the technology after a bubble bursts?
Historically it survived in almost every case. After the railway mania the tracks stayed in the ground, after the fibre boom the cables stayed laid, after the dotcom crash the internet remained. Investors lost, users benefited, often at far lower cost than before.
Scope of this article
The piece you have just read interprets historical market data and describes what follows from it for business decisions. It is not investment advice and not a recommendation to buy or sell securities. Companies and figures named here serve to place the technology cycle in context, not to value individual stocks.
Sources and further reading
For how to demonstrate the actual business value of AI internally, our piece on measuring AI agent ROI is the logical next read. How far implementation has progressed in the German mid-market is covered in the article on the AI agent adoption gap. And why off-the-shelf software alone no longer creates an edge is the subject of SaaS is dead.
Primary sources: William N. Goetzmann, "Bubble Investing: Learning from History", NBER Working Paper 21693. Goetzmann, Kim and Shiller, "Crash Beliefs from Investor Surveys", NBER Working Paper 22143. Peter Oppenheimer, "Why AI stocks aren't in a bubble", Goldman Sachs Research. BIS Annual Economic Report 2026. Federal Reserve History on 1929 and the Great Recession.
