In 1843, the British economy plunged into railway mania. The public had seen the success of the earlier railway projects, and a newly-minted class of middle-class investors wanted their piece of the rapidly-expanding pie. Within two years, stock prices doubled, and the number of railroad securities listed on the London Stock Exchange trebled. In the press, rapturous commentary heralded “the arrival of a time when the whole world will have become one great family, speaking one language, governed in unity by like laws, and adoring one God”.
Such forecasts were hasty. So was much of the investment. The railway bubble burst, and speculators suffered bruising losses. Three decades later, a similar bubble arose in the US. This bubble, like the British one, was soon to burst. Around 18,000 businesses failed within two years.
As a share of American GDP, investment in artificial intelligence has surpassed the telecoms boom of the dotcom bubble. In its degree of concentration, it stands second only to the railroad boom of the late 19th century. AI, then, might be approaching a moment of truth. The technology was meant to speed up the economy; now, it appears to be slowing it. Unless the trend is soon reversed, the biggest tech companies might be in for a sharp revaluation.
In fact, if you remove those companies from the stock index, so-called American exceptionalism disappears. Outside the top 10 companies, the indices have been stagnating since the start of 2025. Similarly, minus that top 10, corporate earnings have been flat for the past three years — coincidentally, the period that followed the release of ChatGPT, the event that sent the AI boom into overdrive. With the sector’s capital expenditure having risen tenfold in just the past three years, the AI bubble is now sucking up more than half of America’s investment. So insatiable is this bubble that there may not even be enough cash in America to satisfy OpenAI. Sam Altman attempted to rustle up additional cash by going abroad, seeking to raise $7 trillion from the UAE and other investors.
That rush to the top doesn’t leave much for everyone else. With all that money pouring into AI, investment in the rest of the economy has begun to decline, slowing growth. That helps to explain the rapid deceleration in the US economy now underway, a marked contrast to the recent good fortune of the “Magnificent Seven” tech companies. With job growth flat, consumption now adds less to the economy than a handful of massive AI investments.
And the gains made by AI are, for now at least, someone else’s loss. Not only is the industry sucking in capital, but it is ravenous for electricity. Data centres consume 10 times the power of traditional servers, and are helping raise American electricity prices — up some 6.5% over the past year. AI centres also consume huge volumes of water for cooling, leading to supply strains and increased prices in many areas. That keeps upward pressure on the inflation rate, raising costs to both consumers and businesses alike and slowing the interest-rate cuts that might facilitate investment in small and medium firms.
In short, the US is currently placing its money on one huge bet: that AI will unleash a productivity revolution so great it will ultimately re-start the economy it’s presently impeding. Americans are doing their bit to make it happen, with ordinary retail investors driving the stock market rally of the last few months and OpenAI, the makers of ChatGPT, reporting that Americans prompt their chatbot some 330 million times a day.
Yet for all that, we’ve yet to see much evidence of a transformation in output. Over the same three years that the AI revolution has been in full swing, labour productivity has grown at barely 1% a year, maintaining a decades-long trend of declining growth across Western countries. And while some might retort that productivity revolutions unfold over many years — several decades passed between the invention of the steam engine and electricity and significant changes in labour productivity — the maths behind the AI boom make little sense. Unlike the railroads built in the late 19th-century US bubble, which are still around today, AI investment is going into data centres that are set to depreciate rapidly. Nvidia, for instance, warranties its GPUs for only five years.
In short, this bet needs to pay off soon. And while the record profits at the tech giants leading the American AI boom might suggest that’s already happening — Nvidia’s earnings have quintupled and its share price increased tenfold in the last three years — those results don’t actually tell us much. Nvidia isn’t making money off AI so much as the AI mania, capitalising on the sale of its chips — the “picks and shovels” of the new goldrush — to other companies that then use them to build AI. In a recent earnings statement, Meta said it did not expect the gains of its AI investment to materialise for another two years. Instead, it is still making money off advertisers on its Facebook and Instagram sites, just as Apple still makes its money from iPhone sales.
“Nvidia isn’t making money off AI so much as the AI mania.”
The record profits of the Magnificent Seven, therefore, do not support a narrative of economic transformation. Instead, we’re awaiting the breakthrough application that will lift economic productivity to a new plateau. Enthusiasts maintain that, once the technology spreads, AI will add one per cent or more to America’s annual rate of economic growth. But there’s no shortage of sceptics who question the calculations behind that optimism.
Among them is Goldman Sachs’ Jim Covello, who says that for AI to justify the current wave of investment, it must solve a trillion-dollar problem. It’s hard to see that happening. Daron Acemoglu has argued that AI’s productivity impact will be negligible, because once it gets through the easy-to-learn tasks which are currently exciting everyone, like search tools, the returns will rapidly diminish. This contention is borne out by recent macroeconomic research from Japan. The neuroscientist Eric Hoel calls this the “supply paradox of AI”: “the easier it is to train an AI to do something, the less economically valuable that thing is. After all, the huge supply of the thing is how the AI got so good in the first place.” Take the example of copy-writers. Because so many of them produce so much material, they become easy to replace. Their knowledge is cheap. This phenomenon won’t carry over to fields where knowledge is expensive.
Instead, a closer look at Meta’s bottom line may reveal what’s really going on, and why the optimistic view of AI’s productivity impacts could rest on shaky foundations. The company credits its new AI for helping steer more customers to its products, thereby boosting sales and revenue. We see a similar trend at Google, whose searches previously led people to websites, but which now keep them on the Google site where they can read quick summaries of the linked content: more eyeballs on Google, more ad revenue to it, but less for all those sites.
In other words, the tech giants have been using AI to capture revenue that previously went elsewhere. On the face of it, that’s the market doing its thing, and it’s that very possibility which excites corporate executives. They believe AI will reduce their need for labour, allowing them to similarly suck up money that would have gone elsewhere — in this case, to their workers.
This, arguably, has been the tale of the information age: new technology doesn’t necessarily raise economy-wide output that much. Instead, it replaces labour and directs the resulting cost-saving to owners, helping to worsen income and wealth inequality. Just as the websites that make those AI summaries possible are now being plundered to the benefit of the AI giants, the LLM chatbots they’re developing are trained using content whose creators receive no compensation. Without that motherlode of free inputs, involuntarily subsidised by their creators, AI might make no sense. As one investor put it in 2023, “imposing the cost of actual or potential copyright liability on the creators of AI models will either kill or significantly hamper their development.”
That route to higher profits raises two problems. The first is that the resulting inequality could stoke a political populism that makes the economy more fragile. We might never have got the volatile policy-making of this Trump administration were it not for the sense of malaise caused by globalisation’s decimation of industrial towns and industrial jobs. If AI really displaces workers, one can’t rule out a radically redistributive politics of the Right or Left. If anything, current trends favour it.
But in addition to the possibility that the AI revolution could end up eating itself, this type of job destruction can produce an economic paradox. While professionals might lose jobs to AI, not everyone will. Microsoft recently commissioned a study estimating which jobs will be most affected by AI and least; not surprisingly, those most insulated from its effects are manual labourers and personal-care workers (including health care), with knowledge-workers taking the biggest hit.
In other words, the jobs AI will disrupt least lie in those sectors of the economy where labour productivity tends not only to be lowest, but most resistant to transformation. This gives rise to what’s known as the Baumol-Bowen effect. As workers displaced by AI are forced to take new jobs in lower-skilled occupations, their productivity will decline. In consequence, the needle on economy-wide productivity may budge little.
It may yet be that someone will invent an application for AI that transforms productivity across the economy and lifts everyone’s output and income to a new plane. But time may be running short, because right now, the boom is becoming a bust for many.
So: are we facing another railroad episode? It’s certainly true that AI is starting to resemble a speculative bubble, not unlike the dot-com boom. Although the Magnificent Seven are making profits that justify their valuations, their future growth depends on the AI revolution continuing for years at its current rate or greater.
But that growth depends on AI being deployed in revolutionary new ways. If that doesn’t happen, then those growth prospects will change, and the markets could tumble. Given the weighting of the big tech companies in the US stock market — Nvidia alone is worth nearly a tenth of the market’s total value — a fall in share prices could not only hurt all those retail investors who rode the wave, but ricochet through the financial system.
In the long term, it might be good for America if the market goes back to investing in all the other sectors of the economy. In the short term, if this bubble bursts, it will be very painful. Both investors and the Trump administration are going all-in on their AI bet. If it succeeds, they’ll parade like champions. But if it fails, they’ll be running for the exits. And given last week’s lacklustre response to ChatGPT’s newest update, investor patience may be wearing thin.
