From railways to dial-up modems, speculators have often confused new technology with guaranteed profits, only to find the future arrives more slowly than the hype funding it.
As the AI gold rush accelerates, investors are becoming jittery. The IMF’s chief economist has drawn parallels to the dotcom bubble of the early 2000s, in which $5 trillion was wiped from the market. The Bank of England warns of a “sudden correction”, and 54 per cent of fund managers surveyed by the Bank of America now think AI stocks are in a bubble. Does the data suggest it’s true?
The S&P 500 index has become a handful of tech giants with 495 other companies attached: virtually all of this year’s growth has come from Nvidia, Microsoft, Alphabet, Amazon and Meta. Spending on data centres alone is estimated to be responsible for nearly half of US GDP growth.
Bubbles are easy to detect after they burst: stocks in those sectors go down a lot, and stay low for a while. But they can be hard to spot when you’re in them. Alan Greenspan, then Federal Reserve chairman, used the term “irrational exuberance” to describe the stock frenzy of the 1990s. But how do you know whether your exuberance actually is irrational?
When discussing how overvalued or undervalued the stock market is, investors often wheel out the “Shiller price-to-earnings ratio”. Named after an American economist, it is the stock price divided by the earnings per share over the past ten years. This magic number was 31 in the summer of 1929. It was 44 in December 1999, and we know what followed. Right now, the S&P 500 is approaching 40.
Sounds scary. Except price-to-earnings isn’t everything. When companies are growing quickly, the figure will always seem big because investors hope for future profits that have yet to materialise. Nvidia, the most valuable company in the world, had a price-to-earnings ratio of 147 two years ago; now that revenues have soared, it has settled to a more modest 52.
The question of whether we really are in an AI bubble boils down to this very principle. In the past few years Google, Amazon, Meta, Microsoft and Oracle — the big companies selling computing power — have increased capital spending from $66 billion in 2020 to $393 billion this year (and should reach $684 billion by 2029). That is more than was spent on the Apollo programme in today’s money, every single year. Most of that goes on data centres: the chips that go in them, mostly made by Nvidia, plus a huge amount of energy and cooling. In fact Nvidia has become so rich that it has started investing in some of its own buyers, prompting fears of circular funding.
How much return will these companies get for their investment — and when? According to the British tech entrepreneur Azeem Azhar, revenues from generative AI are currently at about $60 billion in 2025. In other words this year’s capital spending is more than six times the revenues generated. In theory that makes the market more bubble-like than the US railway expansion of the 1870s and telecoms in the 1990s. For the bubble not to burst, AI needs to start making money fast. But there are good reasons to think it might.
One is the sheer speed of progress. METR, an AI research organisation, has studied the difficulties of tasks that AI models can do well when prompted. The tasks were measured in terms of how long it would take a human to do them.
When Chat GPT-3.5 arrived in 2022 — when we all started writing rude poems to one another in the style of Boris Johnson — the model could reliably achieve something that would take a human 36 seconds. Three years later Chat GPT-5 can carry out tasks that would take a human two hours. This is doubling every seven months: at this rate, by 2029, AI will be able to replicate a month’s worth of work — whether in software development or research — in a few seconds (bad luck, consultants).
What sets AI apart from previous advances is its adoption speed. Last year the Reuters Institute surveyed people in the US, the UK, Argentina, France, Japan and Denmark: 18 per cent had used generative AI in the past week. This year the figure had jumped to 34 per cent. It took three years for internet use to grow by a similar amount in the 1990s. Morgan Stanley thinks generative AI will produce $1 trillion in revenue by 2028.
A market correction of some sort looks likely: there are plenty of overhyped businesses loosely connected to AI whose stock price has soared. In March 2000, Cisco, the poster child of the dotcom frenzy, traded at a price-to-earnings ratio of about 200 before plummeting in value. Nvidia’s is a quarter of that, and Alphabet’s is an eighth. This time the market leaders appear reasonably priced for the money they make.