Are We in an AI Bubble? What the Numbers Say, and What They Don't

Let's do an uncomfortable exercise. OpenAI, the company that makes ChatGPT, was valued in February 2026 at 730 billion dollars in its funding round: the largest private deal in history, with 50 billion put in by Amazon and 30 billion each by Nvidia and SoftBank. That figure places it above the value of almost every publicly traded company in the world. Now the number that doesn't add up: OpenAI does not make money. It loses billions every year.

Holding both sentences at once — worth more than almost anyone, hasn't earned a dollar — is to understand why half the financial world is asking the same thing: are we facing the greatest technological revolution of our era, or the greatest bubble? The honest answer, the one this article tries to give, is that both can be true at the same time, and that separating them demands looking at the numbers coldly and stating plainly what is known and what is not.

The money on the table

The figures are dizzying and deserve to be taken seriously before being judged. The big labs and the companies around them are investing hundreds of billions in building data centers, buying chips, and training ever-larger models. In October 2025, analyst Julien Garran went so far as to call the phenomenon "the biggest and most dangerous bubble the world has ever seen," estimating that the share of the U.S. economy devoted to AI investment is, proportionally, seventeen times larger than the one that fed the dot-com bubble of the late nineties. Investor Ray Dalio, among others, has pointed to parallels with that episode, which ended in a historic crash.

The trick that really worries: money going in circles

There is a detail in this story that worries more than the big numbers, because it smells of the worst moments of other bubbles: the so-called "circular deals." The most cited case: in September 2025, Nvidia announced a 100-billion-dollar investment in OpenAI. But that money came with an expectation — that OpenAI would use much of it to buy more chips from Nvidia itself. In other words, the manufacturer finances its customer so the customer will buy from it.

When money circulates like this, among a handful of giants paying one another, two things turn murky. The first is that valuations can inflate on their own, because each deal raises the price of the next without a single new dollar entering from a final customer. The second is that, if one day the real demand for AI falls short of expectations, the losses aren't spread out: they concentrate and multiply inside that same closed circle. This is not proof of fraud — that must be said clearly — but it is exactly the kind of structure a prudent observer watches.

Why this isn't (yet) the dot-com bust

And here is where the easy comparison breaks, because there is one difference that changes almost everything. The dot-com bubble of 2000 swelled with companies that, in many cases, had neither revenue nor users: they were worth a fortune on a promise and a web page. The artificial intelligence of 2026 is not at that point. ChatGPT and its rivals have hundreds of millions of real users who use them daily, and they generate growing revenue, even if it doesn't yet cover the astronomical costs of training and running the models. There is a product. There is demand. There is technology that works. In 2000, that was missing.

The difference matters because a bubble can burst in two very different ways. It can blow up like the dot-coms, when the market suddenly discovers there was nothing behind it. Or it can deflate more slowly, when there is something real behind it but far more has been paid than that something yet returns. AI, if it falls, would fall the second way, not the first.

The real crack: spending runs ahead of return

The most honest point in this whole debate isn't in the valuations, but in productivity. A study by the U.S. National Bureau of Economic Research, published in February 2026, found that around 90% of the firms surveyed did not yet report a measurable impact of AI on their productivity or their way of working. Executives project future gains — around 1.4% more productivity — but these are promises, not earnings collected.

There lies the tension that defines the moment: money is being spent today, at a historic scale, betting on a return that tomorrow will have to appear. If that return arrives — if AI truly transforms the economy the way the railway or electricity did — today's valuations will, in time, look almost modest. If it takes too long, or arrives smaller than promised, the adjustment will be painful. No one, not the most brilliant analyst, knows today for certain which of the two will happen. Anyone who tells you otherwise, in either direction, is selling you a certainty that does not exist.

So, is it a bubble or not?

The question, put that way, has no useful answer, because a bubble is only confirmed when it bursts and only disproven when time turns it into a bargain. The useful question is another one: how long can the real return on AI take to arrive before the money — which is impatient — loses faith? That is the countdown running beneath the entire sector, quietly, while ever-larger rounds are announced.

The likeliest outcome is neither total collapse nor instant validation. It's a sorting: some of today's bets will prove brilliant and others will turn out to be expensive follies, and not always the ones that look like each today. The technology, almost certainly, will stay and mature; part of the money, almost certainly, will evaporate along the way. Telling the two apart in advance is, precisely, the hardest and most lucrative job of this moment. And it is a job that, for now, no one has solved.