Is the AI Bubble Real? A Builder’s Answer

Between 1844 and 1846, British investors committed the equivalent of the nation’s entire annual economic output to railway shares. Parliament authorised more than 9,000 miles of new line. A great deal of it was never built. When the mania broke, the railway index lost roughly two-thirds of its value, and middle-class families who had put their savings into companies with no track and no trains lost everything. Then the dust settled, and Britain found itself with the densest rail network in the world, most of which is still carrying passengers today.

I keep that episode in mind whenever someone asks me whether AI is a bubble, because the question is built on a false choice. A bubble and a genuinely important technology are not alternatives. More often than not they are the same event seen from two angles. Jeremy Grantham, who has correctly called more bubbles than almost anyone alive, puts it more bluntly than I would dare to.

So the useful questions are not whether AI matters. It plainly does. The useful questions are narrower: how stretched are the economics, who gets hurt when they snap back, and what is left standing afterwards. Let me take them with the actual figures, because the figures have become hard to look away from and most of the commentary is still arguing about whether robots will take our jobs.

The gap between what is spent and what is earned

In 2026 the big cloud providers will spend somewhere around 660 to 690 billion dollars in capital expenditure, much of it on AI, against roughly 51 billion dollars in direct AI revenue. More than ten dollars going out for every dollar coming back. When cloud computing was at a comparable stage of adoption in 2011, that ratio was about two and a half to one.

Ten to one is not, by itself, proof of anything mad. I run my own products at a loss while they find their feet; every worthwhile technology spends years underwater before it surfaces. But a ratio that size is a wager, and the wager is that enormous revenue is arriving soon, at a scale and speed with no real precedent.

OpenAI is where that bet is most concentrated. The company reported around 13 billion dollars of revenue in 2025 and has reportedly told investors to expect 100 billion by 2028. No company has ever climbed from roughly ten billion to a hundred billion that fast. When a podcast host gently pointed out that a business with 13 billion in revenue had signed something like 1.4 trillion dollars in compute commitments, Sam Altman’s answer amounted to: revenue is higher than you think and growing steeply, and I am betting it keeps going. Strip away the confidence and what is left is the arithmetic. The 100 billion projection is not a description of OpenAI. It is the load-bearing assumption holding up 1.4 trillion dollars of promises.

Why the hardware is the problem

Here is the part the excitement skips, and the part I pay most attention to as someone who actually pays compute bills. The infrastructure of this boom does not behave like the infrastructure of the booms it gets compared to.

Railways, canals, the fibre-optic cable strung across the ocean floor in the late 1990s: those were things built once that then lasted. The railways ran for a century. The dotcom fibre sat dark for years before becoming the literal backbone of the internet you are reading this on. That is the saving grace of an infrastructure mania. The money burns, but the steel and the glass remain, and someone eventually puts them to work.

The hardware at the centre of the AI boom rots. The expensive part is the chips, and an economist recently described them as “digital lettuce,” which is the most honest two words written about this whole subject. A motorway laid this year will still be a motorway in 2050. A top-end GPU bought this year is hot, hammered around the clock, and quietly humiliated by a faster one inside three years. You are not building the railway. You are buying the lettuce, and the clock starts ticking the moment it reaches the loading dock.

This is not a figure of speech reaching for effect. It sits underneath a serious accounting fight. Michael Burry, the investor who made his name shorting the 2008 housing bubble, argues that the hyperscalers are writing their chips down over five or six years when the real working life is closer to two or three. If he is right, they are understating depreciation by something like 176 billion dollars between 2026 and 2028, and flattering their profits by roughly the same amount. The lettuce is being booked as though it were tinned (won’t rot, won’t need replacing) and the difference is being counted as earnings.

The money that runs in a circle

The third thing to understand is the circularity, because it is what can turn an ordinary slowdown into something faster and worse. And here I do not need an analogy, because the real version is more striking than anything I could invent.

Nvidia invests in OpenAI. OpenAI uses the money to buy Nvidia chips. Nvidia books the revenue, its share price climbs, and the higher valuation helps fund the next round of investment into the next customer who will spend it on more chips. Microsoft and Nvidia together put around 15 billion dollars into Anthropic; Anthropic is expected to spend heavily on their cloud and chips in turn. The same dollars travel a loop, and at every stop on the loop somebody records them as fresh demand.

None of this is illegal or even unusual at the dawn of a major technology. But it means a meaningful slice of what looks like roaring demand is the same capital lapping the track. A loop is a wonderful thing while everyone keeps running. The trouble is that it unwinds exactly as efficiently as it built up. If one runner stumbles, the circle does not slow down gently. It reverses.

So is it different this time?

How the AI boom compares with the boom-bust cycles of past major innovations

Investment as a multiple of the pre-boom trough

Lay the AI investment curve over the great manias of the past, each measured as a multiple of where it started, and the shape speaks for itself. By year three, AI has already climbed steeper than railway mania, steeper than the dotcom boom, past the canal mania and the Roaring Twenties without slowing. The body that produced the underlying work is not a doom-mongering newsletter. It is the Bank for International Settlements, the central bank used by central banks, and close to the most temperamentally cautious institution in global finance. When that building starts drawing charts like this, it is worth a moment of your attention.

Steeper is not stronger. It means more violent on the way up and more violent on the way down. The comparison earns its keep precisely because it cuts both ways: three of the four historical booms on that chart burst, and all three left behind infrastructure that remade the world. The railways ruined their investors and gave Britain a century of trains. The dotcom crash vaporised paper fortunes and left the fibre carrying these words to your screen. The pattern is never “bubble, therefore worthless.” It is “bubble, and then a long, quiet harvest of real value by people who were not the ones who paid for the planting.”

What is different this time is the perishability. The Victorians did not have to re-lay their track every three years. If the AI revenue arrives late, you cannot simply wait a decade for the world to grow into the capacity you built, because by the time the world is ready the hardware has spoiled and the invoices for its replacement are already in the post.

What actually happens when it corrects

I do not think the outcome is “AI is fake and it all goes to zero.” That is the lazy version of the bear case, and it is wrong. I know the technology is real and useful because I use it every day to build things that work and that people pay for.

The likelier shape is more particular, and more interesting. The largest spenders, Microsoft, Alphabet, Amazon and Meta, are among the most profitable companies in the history of money. They can swallow writedowns that would erase a smaller firm without missing a quarter. They are not the fragile part. The fragile part is the layer underneath: the specialist cloud providers built entirely on borrowed money and leased chips, and the labs burning cash at a rate that would frighten anyone who had to do it with their own. OpenAI expects to keep burning more than half its revenue through 2027, with positive cash flow not arriving until the end of the decade.

Set that against Anthropic, whose models I happen to build on, which forecasts cutting its burn to about a third of revenue in 2026 and to single digits in 2027, aiming to break even around 2028. Same boom, two completely different relationships with gravity. That is the detail that ought to end the lazy version of this argument. “Is AI a bubble” treats the whole industry as one lump. It is not one lump. Some of these companies are building something they can pay for, and some are running a bet that only clears if the most optimistic revenue forecast in corporate history lands dead on schedule. When the correction comes, it will tell them apart with no mercy at all.

How I build, knowing all of this

None of this stops me building. It changes how I build, and the changes happen to make for better products regardless of what the market does next.

I make ordinary, deterministic code carry as much of the weight as it can. The model does the part that genuinely needs judgement; plain, predictable, cheap software does everything else. That is partly a hedge against the day the compute subsidies thin out and the true price of a model call lands on my desk. But it is mostly just how you build something that does not fall over, because every part of a product that depends on a model behaving today exactly as it did yesterday is a part you do not fully control.

I treat the cost of a model call as a real number rather than a rounding error. A lot of what people are enjoying right now is priced as though inference were nearly free. It is not free. It is subsidised, and subsidies have a habit of ending the moment the people paying for them want their money back. A product that only makes sense while a venture capitalist three steps upstream is quietly covering its running costs is not a product. It is a seat in someone else’s bet. So I build as though I will eventually pay the real price, on the assumption that one day I will.

I keep a person in the loop wherever being wrong is expensive, and I treat accuracy as something to design around rather than a gate the model passes or fails. That is a whole essay in itself, so the short version will have to do: a tool that is right most of the time and honest about the rest is worth more, and is being more honest about its own economics, than one performing a confidence the underlying technology has not earned.

And I do not stake my own work on the hype surviving. The things I build are meant to still be useful if the valuations halve, if the cheap models get expensive, if half the names at the top of today’s table are gone in five years. That is not gloom. It is the same habit I bring to a programme that is running hot: assume the rosy forecast is wrong, build so that you are fine when it is, and let yourself be pleasantly surprised if it holds.

So, is the AI bubble real? The boom is real and historically extreme, and the BIS chart puts that beyond much argument. A painful correction in the financing looks more likely than not, and it will land hardest on the people who built as though the best case was the only case. The technology will come through it, the way the railways came through their mania and the fibre came through its own, to be put to proper use by people who kept their heads the first time round.

The whole trick, if you are building now, is to be one of those people in advance. Use the technology, because it is genuinely good. Refuse the story being sold around it. Build as though the lettuce will spoil. It will.