Are We in an AI Bubble? What History Says About the Morning After

Every generation gets the mania it deserves, and ours is arriving under the banner of artificial intelligence. Ask whether we are in an AI bubble and you will get two loud answers, both delivered with total certainty. One camp says the technology is a mirage and the crash is a matter of when, not if. The other insists this time is different, that intelligence itself is being industrialised and no price is too high. I think both camps are missing the more useful question. A financial bubble and a genuine technological revolution are not opposites. They routinely arrive on the same train — and history has a lot to say about who gets off richer and who is left standing on the platform.

So let me try to hold two ideas at once, because the honest reading of this moment requires it. Yes, there are classic signs of froth. And yes, the underlying technology is real and probably durable. The interesting part is not which of those is true. It is what happens the morning after, when the froth clears and the durable part is all that remains.

The signals that make experienced people nervous

Bubbles do not announce themselves, but they do rhyme. If you have read the histories of the railway mania of the 1840s or the dot-com boom of the late 1990s, the pattern-recognition part of your brain starts firing when you look at the present. A few signals stand out.

The first is valuations that have detached from anything you can presently measure. When a company’s worth is justified almost entirely by a story about the future rather than the cash it produces today, you are being asked to buy a narrative. Narratives can be correct. They are also, by construction, unfalsifiable in the short run, which is exactly what makes them dangerous. Many analysts have pointed out that a handful of firms now carry an outsized share of the entire market’s gains, and that concentration itself is a kind of fragility.

The second signal is capital expenditure racing far ahead of revenue. Enormous sums are being poured into data centres, chips and the electricity to run them, on the expectation that demand will show up to fill all that capacity. Building ahead of demand is how you win a land grab, but it is also how you end up with a great deal of idle capacity if the demand curve bends even slightly. The gap between what is being spent to build AI and what AI is currently earning is wide, and by most accounts it is widening.

The third signal is the borrowing. When expansion is funded increasingly by debt and by circular financing arrangements — where the same money loops between chip makers, cloud providers and model companies, each booking the other’s spending as its own revenue — the system becomes reflexive. Everyone looks healthy as long as everyone keeps spending. That works beautifully on the way up and unwinds violently on the way down, because the same links that transmit confidence also transmit panic.

A financial bubble and a genuine technological revolution are not opposites. They routinely arrive on the same train.

The fourth signal is the most human one: the vocabulary of inevitability. When you are told that skepticism is simply a failure of imagination, that the only real risk is being left behind, that everything is different now — that is the emotional register of every mania that has ever happened. The certainty is the tell. Real, boring, durable technologies do not need you to suspend your judgment. They just quietly compound.

And yet the technology is genuinely useful

Here is where I part ways with the pure doom camp. It would be a mistake to look at those warning signs and conclude that the technology is fake. It is not. Large language models and the tools built on them are already doing real work — drafting, summarising, translating, writing and debugging code, sifting through documents at a scale no human team could match. Millions of people use them every day and would notice if they vanished. That is not what a hoax looks like.

This is precisely why the railway and dot-com comparisons are so instructive, and why they cut against the doomers as much as the hype. The railways of the 1840s really did remake Britain; the trains ran, the freight moved, the country was rewired. The internet of the late 1990s really was going to change commerce, media and communication, and it did — more profoundly, in the end, than even the boosters claimed. In both cases the technologists were right about the destination and badly wrong about the timing and the prices. The mania was not that the technology was worthless. The mania was that its value had been pulled forward, priced in, and then some.

So the durable-technology case and the bubble case can both be true at once. The trains were real and most of the railway companies still went bust. The internet was real and most of the dot-coms still vanished. You can believe completely in the destination and still lose everything on the way there if you buy at the top and hold the wrong ticket.

What history says about the morning after

If you want to know what a post-mania landscape looks like, the historical record is remarkably consistent. Three things tend to happen.

First, the froth clears. The valuations that depended on the story rather than the cash come down, often brutally and often faster than anyone expected. Fortunes made on paper evaporate. The people who told you that the old rules of arithmetic no longer applied go quiet.

Second, the weak firms fail. A mania funds an enormous number of businesses that only make sense in a world of infinite cheap capital and infinite patience. When the money tightens, the marginal players — the ones with no real product, no path to profit, nothing but a deck and a story — are washed out. This is painful and it is also the point. The shakeout is how capital gets reallocated from the fantasists to the operators.

Third, and this is the part the doomers always miss, the durable infrastructure survives. When the dot-com bubble burst, the fibre-optic cable that had been laid in a frenzy did not disappear. It sat there, cheap and abundant, and became the backbone that the next decade of the internet — the genuinely world-changing decade — was built on. The railway lines outlasted the companies that bankrupted themselves building them. Electricity, a speculative fever in its own day, ended up powering everything precisely after the mania broke. The technology graduates from the story that oversold it.

Applied to now, the likeliest shape of the morning after is not that AI turns out to be useless. It is that the financial structure built on top of it corrects hard, a great many companies with no defensible business die, and the models, the chips and the data centres — the actual infrastructure — carry on and get absorbed into ordinary life. The froth was never the technology. The froth was the finance wrapped around it.

Who books the gains, who pays the bill

This is the part I care about most, because it is the part that never seems to change no matter which machine we are talking about. Booms and busts are not weather. They are not things that simply happen to everyone equally. They have a distribution, and the distribution is engineered.

On the way up, the gains are booked by a specific and predictable set of people. Insiders and early holders who can sell into the enthusiasm. The intermediaries who take a fee on every transaction regardless of whether the underlying thing ever works — the underwriters, the promoters, the platforms. The founders who quietly realise secondary shares while telling everyone else to hold for the long term. By the time the story reaches the ordinary saver, the retail investor, the pension fund chasing the index, the smart money is already looking for the exit. The people who arrive last, on the strength of the very inevitability-talk I described earlier, are the ones holding the position when it turns.

The froth was never the technology. The froth was the finance wrapped around it — and someone always arranges to be gone before the bill arrives.

This is the same move, a new machine, every time. It is exactly the pattern I keep returning to in my writing about how technology gets captured: a genuinely useful tool arrives, a small group arranges to own the chokepoints, and the costs get quietly socialised while the gains get privatised. Ask who owns AI and you are really asking who is positioned to book the upside and offload the downside. The answer is not the users, and it is not the workers whose labour trained the models or whose jobs are the stated efficiency. It is whoever controls the compute, the data and the capital — which is why the fight over the AI monopoly matters more than the fight over whether the bubble pops.

Consider who actually pays when the correction comes. Not, mostly, the people who engineered the run-up; they took their money off the table in real time. The bill lands on the late retail investor, on the pension whose managers chased the momentum, on the workers at the shaken-out companies, on the communities that bet their power grids and water tables on data centres that may or may not get used, and — if the financing turns out to be systemic enough — on the taxpayer who backstops the fallout. The gains are concentrated and taken early. The losses are diffuse and arrive late. That asymmetry is not a bug in these episodes. It is who takes and who pays, laid out about as plainly as it ever gets.

How to read the moment — without hype or doom

So what do you actually do with all this? Not much good comes from either of the loud answers. Pure hype has you buying the story at its most expensive. Pure doom has you dismissing a real technology and missing what it genuinely changes. The useful stance is the uncomfortable middle, and it comes down to a few disciplines.

Separate the technology from the trade. The question “is AI real and useful?” and the question “is this a good price?” are completely different, and conflating them is how people lose money in both directions. You can be a total believer in the technology and still think the current financial structure is a fever. In fact, if you have read the histories, that is the intellectually consistent position.

Watch the cash, not the narrative. The signal that the froth is clearing will not be a change in the story — the story is always bullish until the moment it is not. It will be in the boring numbers: does the revenue start to catch up to the capex, or does the gap keep widening while the borrowing does the heavy lifting? Durable technologies eventually show up in the accounts. Manias show up in the leverage.

Ask who is selling to whom. In any moment of extreme enthusiasm, follow the direction of the smart money. If insiders are distributing shares to the public while narrating inevitability, that tells you where in the cycle you are more reliably than any forecast.

And hold both truths at once. It is entirely possible — I would say likely — that we are living through both a financial bubble and a genuine technological revolution simultaneously, and that the morning after will bring a hard correction in the finance and a durable survival of the infrastructure. The trains will keep running after the train companies fail. The models will keep working after the valuations reset.

The thing to resist is not the technology, and it is not even the boom. It is the pretence that the outcome is a law of nature rather than a distribution someone arranged. The gains and the losses in every one of these episodes fall on different people by design, and the last time I checked, the machine changes but the design does not. Read the moment clearly enough to see who is being set up to hold the bill — and then decide, with your eyes open, whether you intend to be that person.