Fears of an AI bubble are rising
CFOTO/Sipa USA/Alamy
The hundreds of billions of dollars being spent on AI seem to have inflated a global financial bubble that’s now fit to burst, leaving companies and investors at risk of holding vast debt that cannot be serviced by the meagre revenue brought in by current AI services. But what does that mean for the future of the technology underpinning this financial feeding frenzy?
In recent weeks, warnings of a potential AI bubble have come from the International Monetary Fund, the Bank of England, the head of the largest US bank, and even OpenAI boss Sam Altman. “This has not just been a a stock market bubble, it’s been an investment bubble, it’s been a public policy bubble,” says David Edgerton at King’s College London.
The circular nature of some of the deals between major AI players is also raising eyebrows. For example, Nvidia, which builds the GPU chips that are powering the AI boom, recently invested up to $100 billion into OpenAI so that the company could build a new data centre full of Nvidia’s own chips. OpenAI, in turn, has agreed a deal that could see it ultimately take a 10 per cent stake in Nvidia’s rival chipmaker, AMD.
Concern about the AI bubble bursting is also thrown into sharp relief when you realise the scale: at least $400 billion is being spent annually on data centres, according to Morgan Stanley Wealth Management. And while US GDP rose by 3.8 per cent in the second quarter of the year, Jason Furman at Harvard University estimates that if you removed data centres from the equation it would have barely grown 0.1 per cent over the whole first half of the year.
Carl-Benedikt Frey at the University of Oxford says this kind of exuberant deal-making isn’t unusual in the history of technology – in fact, it would be unusual if the global economy managed to invest in infrastructure for a new technology at precisely the right pace to meet demand. “It’s quite usual that you overbuild: the same thing happened with the railroad boom, the same thing happened with the dot-com bubble,” he says.
The question is whether the fallout from an AI bubble would just harm the companies involved, or could have wider impacts. Frey points out that many of these hugely expensive data centres are actually being built “off balance sheet”. This involves the creation of new companies backed by external investors or banks that build and own the assets, taking on both the risks and potential rewards.
As a result, we don’t know enough about who is exposed to this risk. A data centre could be financed by a dozen technology billionaires, or it could be high-street banks – and if their losses are large enough, then a banking crisis could send shockwaves throughout the wider economy. “That’s not to say that there’s an imminent financial crisis, but that it’s a bit opaque. And when things are opaque, there’s usually some risk,” says Frey.
Benjamin Arold at the University of Cambridge says the giveaway is the ratio of profits to company valuations, which indicates how disconnected public opinion is from the actual money businesses bring in. He says these figures for technology firms today are a red flag.
“The last time it was this low was 25 years ago, and if you remember, 25 years ago we had the dot-com bubble,” says Arold. “It’s possible that it goes well, but I would not bet my money on it.”
James Poskett at the University of Warwick, UK, believes we are heading for a correction in the AI industry that may spell the end for many companies, but he says this certainly isn’t the end for the technology itself. “It’s important not to confuse that with the idea that the technology is flawed or going to go away,” says Poskett. “There might be an AI bust, but that doesn’t mean we’re not gonna have AI.”
Just as the consolidation of numerous railway companies after a bust left us with a rail network, and the collapse of technology firms in the dot-com bust left us a legacy of extensive fibre optic networks, we’ll be left with useful technology, says Poskett.
For consumers, the AI bubble popping is likely to mean a bit less choice, maybe paying a bit more for access, maybe seeing a slower pace of updates. It could force us to face the reality that using a vastly expensive tool like GPT-5 to write an email is like using a sledgehammer to crack a nut, and that the true cost of using it had previously been hidden by the frenzied AI arms race. “At the minute, there’s a lot of free lunch, but at some point these companies have got to make a profit,” says Poskett.
Topics:
