Stock Market Bubbles: Dot-Com to AI, Measured
A stock market bubble is a stretch when prices rise far faster than the profits behind them, on expectations that later prove too high. It is only confirmed by the fall: the Nasdaq Composite rose 529.3% in the five years to 10 Mar 2000, then lost 77.9%.
“Bubble” is the most argued-over word in markets, because it describes a price that cannot be proven wrong until after it has fallen. This page does two things with it: explains how a bubble builds, and measures the best-known one, the dot-com run of the late 1990s, against the market of September 2026.
The measuring uses daily closing prices for the Nasdaq Composite and the S&P 500. It does not use a forecast, and it does not end with a verdict on whether today is a bubble. No measure available to anyone can deliver that verdict in advance.
How it forms
It starts with something real. Each well-known bubble grew around a genuine change: railways, radio, the internet. Artificial intelligence is the change being argued over now. The new thing does change the economy, and early investors do earn large returns, which is what makes the story believable.
Rising prices then become their own argument. Buyers who missed the first leg see others profit and join in. Their buying lifts prices further, which draws in the next group. Economists call this a feedback loop; in plain terms, the price itself becomes the main reason to buy.
Expectations drift away from what the businesses earn. Share prices begin to assume that growth will be faster, last longer or be shared by more companies than it turns out to be. The gap shows up in valuation measures such as the P/E ratio, where price rises faster than earnings.
Supply follows demand. When investors pay high prices for anything carrying the theme, companies sell new shares to meet them, and the number of young companies coming to market through an IPO rises.
Borrowed money speeds both directions. Leverage lets buyers hold more than they could pay for, which lifts prices on the way up and forces sales on the way down.
The end rarely has one cause. Rates rise, a large company disappoints, or buyers simply run out. The fall that follows is what the market crash page covers.
How bubbles are measured
By price. The simplest test is how far and how fast prices rose, for example the gain over the last five years. It is objective and easy to compute, which is why this page uses it. Its weakness is that it ignores whether profits rose just as fast.
By valuation. Price relative to earnings, sales or book value. Robert Shiller’s cyclically adjusted P/E averages ten years of earnings to smooth out recessions. Warren Buffett has pointed to total market value against the size of the economy. Both describe how expensive a market is, not when that changes.
By concentration. How much of the index sits in a handful of companies. When a few names such as the Magnificent Seven drive most of the gain, a setback in one theme moves the whole index, the risk the concentration risk page describes.
By behavior. Margin borrowing, first-time investors, new share issuance and headline enthusiasm are all signs people point to. They are harder to measure consistently, and they rise in healthy bull markets too.
Dot-com to AI: the comparison
The comparison everyone makes is between the internet run of the late 1990s and the AI run since 2023. On price, the two are not close.
In the five years to its peak close of 5,048.62 on 10 Mar 2000, the Nasdaq Composite rose 529.3%. Over the final three of those years it rose 281.7%. On 25 Sep 2026 it closed at 27,068.72, up 79.9% over five years and 104.0% over three.
The S&P 500 tells the same story at a smaller scale. It rose 204.9% in the five years to its 24 Mar 2000 peak of 1,527.46, and 73.8% in the five years to its 25 Sep 2026 close of 7,743.41.
Valuations of today’s largest companies, measured on 25 Sep 2026: Nvidia traded at 28.5 times its last four quarters of earnings, Microsoft at 28.8, Apple at 39.1 and Amazon at 20.1. This site does not have an audited earnings series for 2000, so it does not quote the dot-com multiples often repeated elsewhere.
None of this says today is safe. A smaller run-up can still end in a large fall, and a market can stay expensive for years. It says that on the simplest objective measure, the 2026 run is ordinary by the standard of the last fifty years.
A worked example
Take a hypothetical $10,000 put into the Nasdaq Composite at the close on 10 Mar 2000, the top of the dot-com run, on price alone with no dividends.
By the close on 9 Oct 2002 it was worth $2,206.76. The index had fallen from 5,048.62 to 1,114.11, a loss of 77.9%.
Getting back to $10,000 then required a gain of 353.2%, because a loss is measured from the higher starting price and the recovery from the lower one: 5,048.62 / 1,114.11 = 4.532, a gain of 353.2%.
That recovery took until 23 Apr 2015, the first close above the 2000 peak, 15.1 years after it.
The same $10,000 in the S&P 500 at its 24 Mar 2000 peak fell to $5,085.31 at the 9 Oct 2002 low, a loss of 49.1%, and needed a 96.6% gain to recover. It first closed above the peak on 30 May 2007, 7.2 years later, about four months before its October 2007 peak and the fall that followed.
The original data
Every run-up here is measured the same way: the closing price divided by the last close on or before the same date five years earlier. The fall is to the lowest close before the index first closed back above its peak. Prices only, no dividends.
| Nasdaq Composite peak | 5-year gain before | Fall that followed | Years to close above the peak again |
|---|---|---|---|
| 10 Mar 2000 | 529.3% | -77.9% (9 Oct 2002) | 15.1 |
| 31 Oct 2007 | 115.0% | -55.6% (9 Mar 2009) | 3.5 |
| 19 Feb 2020 | 99.3% | -30.1% (23 Mar 2020) | 0.3 |
| 19 Nov 2021 | 201.7% | -36.4% (28 Dec 2022) | 2.3 |
| 25 Sep 2026 (latest close) | 79.9% | - | - |
On the S&P 500 the same peaks show five-year gains of 204.9% (March 2000), 101.5% (October 2007), 61.4% (February 2020) and 112.4% (January 2022), against 73.8% on 25 Sep 2026.
Where today sits in the whole record. The Nasdaq has 12,765 trading days since 5 Feb 1976 with five full years of history behind them. On 6,345 of them, 49.7%, the five-year gain was at least as large as today’s 79.9%. The highest was 533.8%, on 9 Mar 2000.
The S&P 500’s record run was not in 2000. Of its 18,054 trading days since 1955 with five years behind them, 29.8% had a five-year gain at least as large as today’s. Its highest was 224.9%, on 11 Aug 1987. Two weeks later the index peaked, and on 19 Oct 1987 it fell 20.5% in one day.
The peaks, gains, falls and recovery dates are in the run-up table for this page.
Of the 24,971 unique videos in this site’s search study, 9 have “bubble” in the title, from 9 channels, at a median of 37,340 views. Three of them are about an AI bubble. That is a small sample, and it says the question is being asked, not how it will be answered.
When it fails
The first failure is calling it too early. On the evening of 5 Dec 1996, Federal Reserve chairman Alan Greenspan asked in a speech how anyone would know when “irrational exuberance” had pushed asset values too high. The S&P 500 closed at 744.38 that day. It then rose another 105.2% to its March 2000 peak. A reader who sold on the warning missed that run, and even the bottom of the fall that followed, 776.76 in October 2002, was above the close on the day of the speech.
The second is treating a big run-up as proof. The Nasdaq rose 99.3% in the five years to February 2020 and 201.7% into November 2021. The falls that followed, 30.1% and 36.4%, were painful but recovered in months and in a little over two years, not fifteen.
The third is hindsight. Every chart of a bubble is drawn after the fall, which makes the top look obvious. At the time, the same prices had serious defenders who could point to real growth. The survivorship bias page explains why the companies that lived through a fall are the ones people remember, and the ones that did not are forgotten.
The fourth is betting on the pop. Selling short into a bubble means paying while it keeps inflating. Being right about value and wrong about timing produces the same loss as being wrong.
And the fifth is the measure itself. A five-year price gain says nothing about profits. A market that rose 80% while earnings doubled is cheaper than when it started, and one that rose 80% on flat earnings is not. Price is where this page starts, not where the question ends.
Related
The market crash page follows what happens after a peak, and drawdown explains why a 77.9% fall needs a 353.2% gain to repair. The P/E ratio page covers the valuation measure behind the table above.
For the companies at the center of today’s debate, the Magnificent Seven page lists who they are, and the Nasdaq 100 page covers the index most exposed to them.
When a market feels like a bubble, I stop asking whether it is one and start asking what I would do if it fell by half. The first question has no answer anyone can use. The second one sets position size, and that is what decides how much a bubble can actually cost.
— Michael Whitman
This page is educational, not financial advice. Test every idea on your own charts before risking money.