Too big to compute? AI’s bubble trouble

There have been whispers for some time now that, like the dot.com bubble, the current AI boom is looking a little fragile. To quote Ray Dalio, the Bridgewater Associates founder, “There’s a lot of bubble stuff going on… This looks quite a lot like 1998 or ’99”.

The American economy is, for the moment, a tale of server farms and big money bets. Data centres are sprouting across US suburbs and farmland like a new kind of industrial corn, and markets are buying the harvest in bulk. Chip manufacturer Nvidia crossed the $5 trillion mark this autumn, a valuation that says less about its present cash flows and more about the scale of belief in an AI future that will pay for all the megawatts. Meanwhile, some regulators have begun to squint at the financial plumbing. The Bank of England recently published a blog warning about the surge in loans to AI-linked data centres and the web of “circular” financing that is knitting together chip vendors, cloud providers, and potential tenants. 

One of the reasons for the concern is that the capex numbers are staggering. This year, Microsoft notched up $24.2 billion of capex in a quarter and guided to $30 billion next quarter - both records. Amazon told investors it expects $125 billion of capital spending this year and disclosed $34.2 billion of cash capex in Q3 alone, much of it for AI‑ready capacity. 

Harvard’s Jason Furman estimates that investment in information‑processing equipment and software contributed 92% of US GDP growth in H1 2025; strip that out and growth was roughly 0.1% annualised. In other words, without data centres, the US economy would have been pretty much flat. Yet the deployment pay‑off inside companies is modest. The Stanford AI Index 2025 finds that most businesses reported only single‑digit cost savings and sub 5% revenue rises from AI to date, while survey data shows adoption racing ahead of productivity. 

So far, so 1999, but it’s the financial plumbing where things start to look a bit more 2008. Grace Blakeley recently wrote in a Substack article: “Nvidia invests in OpenAI; OpenAI signs vast compute leases with Oracle; Oracle buys, and books more Nvidia hardware; everyone in the loop reports rising orders and rising valuations, even as cash sloshes round the same drain.” Put simply, the optics of demand can be funded by its own suppliers. 

Normally, customers pay suppliers. Here, suppliers are funding customers so they can buy more from the same suppliers, while financiers bankroll the buildings to house it all. If all of this has the whiff of 2008, it’s because the structures feel similar, even if the assets have shifted from suburban homes to server farms.

What’s different from 2008 is the assets. Instead of subprime mortgages, its long‑dated leases to hyper-scalers with investment‑grade profiles and campuses whose value depends on power and interconnects. 

While they are no doubt stronger than NINJA mortgages, they are not invulnerable. Refinancing risk and rate risk persist and crucially the assets are bespoke and alternative uses are limited if the demand or economic model disappoints. Even optimists concede the onsite jobs are low once construction ends; the macro uplift relies on productivity gains that, so far, haven’t scaled like the marketing has. 

As Carl-Benedikt Frey, Associate Professor of AI & Work at the University of Oxford, put it: “The vast bet on AI infrastructure assumes surging usage, yet multiple US surveys show adoption has actually declined since the summer. Unless new, durable use cases emerge quickly, something will give, and the bubble could burst.”

The further worry is that the macro backdrop is not particularly positive. By October this year, 22 US states were in or near recession, collectively representing about a third of US GDP. When asset prices are built on expectations while real‑economy indicators lag significantly, you don’t need a PHD in economics to sense fragility. 

The market’s defence is plausible; every general‑purpose technology overbuilds its infrastructure first and extracts value later. Fiber transformed communication; railways reshaped trade; cloud found its workloads. Perhaps GPUs and gigawatts will look prudent in hindsight. The issue is that bubbles are not bubbles because the underlying tech doesn’t work, they are bubbles because the timelines don’t work. 

Perhaps the lesson is not to let the marketing hype get away from reality. Today’s valuations, and the debt structures beneath them, presume the payoff arrives before the bill is due. If it doesn’t, let’s hope we don’t have another 2000 - or worse, another 2008 - on our hands.