Why the bursting of the AI bubble only helps those who built beforehand
Half of tech is waiting for the AI bubble to burst as if the bursting itself were the liberation. It isn’t. When it bursts, what comes is not the reward for those who saw it coming, but the bill for everyone who hung half their business on the API of a single American lab. Who ends up paying is decided not at the moment it pops, but in the months before.
This is not a forecast about share prices, it is an owner’s decision. Anyone who today rests their processes, their knowledge, their competitive edge on foreign models buys a dependency whose price only comes due when the market turns. Not deciding is also a decision, only a worse one, because it feels careful. The rest of this piece is about where the value goes when the bubble turns, and why the falling price only lands in the hands of those who built beforehand.
The direction is turning, visibly
That it is turning is no longer a claim, it is there in the numbers. The four big American hyperscalers plan around 725 billion dollars in investment for 2026, up 77 percent on the previous record. Amazon alone is heading toward roughly 200 billion. It is one of the largest private capital bets ever made, and it runs on the assumption that demand keeps growing in a straight line.
At the same time the ground under that bet is cracking in three places. First the model itself. An open model like DeepSeek V4 comes close on real tasks, level on some tests, a few points behind on the hardest, at roughly a thirtieth of the cost per token. For most applications that gap is irrelevant. The gap between 28 and 2500 dollars a month is not. Second the price of compute. An H100 on the spot market now costs around one to three dollars an hour, down from over seven in early 2024. What was scarce and expensive yesterday is turning into a commodity with a falling price.
Third, and this is the most honest crack, the companies do not believe their own bet. They depreciate their Nvidia chips over five to six years, while the real product cycle runs two to three. Meta alone pulled 2.9 billion in depreciation out of a single year just by stretching the useful life. Michael Burry puts the sector’s understated depreciation at 176 billion between 2026 and 2028, with earnings at individual firms overstated by more than 20 percent. Anyone who books the life of their most expensive asset as longer than the vendor ships new generations is polishing the balance sheet. It is exactly this bigger-is-better logic that Nicolas Colin reads, rightly, as the turning signal.
Local models are not the answer
Colin draws the conclusion you now hear everywhere: local models are the way out. Away from the frontier labs, toward the good-enough, cheap model on your own hardware, closer to your own edge. This is where the analyst and the owner part ways.
Because the dependency does not disappear when the model gets cheap. It moves. I have argued elsewhere that enforcement today needs no territory, only the dependency someone else hangs on. The same logic bites one layer down. A model from Shenzhen at a thirtieth of the price does not dissolve the dependency, it relocates it. Away from the American lab’s API, toward the silicon the local model runs on, the repository its weights load from, the channel it updates through. Each of those layers has its own switch, and most of them still sit somewhere else.
That is the mistake underneath the whole sovereignty hope. Decentralization gets mistaken for independence. An American bill wants every exported AI chip to report its location continuously, and it explicitly examines remote shutdown. Run your local model on exactly that silicon and you have changed the location of control, not the fact that it lies elsewhere. The cheap model on someone else’s hardware is not a leap from renting to owning. It is a move into another apartment owned by the same landlord. The real question is not which model you pick, but how fast you can leave the one you picked when the switch moves.
What the turning point acutally shifts
To think this bigger, Carlota Perez has the map. Across five technological revolutions she described the same pattern. An installation phase, in which financial capital pours into the new infrastructure, speculates and overshoots, until a bubble bursts. Then the turning point, and only after it the real deployment, when the technology works its way through the whole economy. Railways, electricity, the car, the microchip, the same arc every time.
The part most people skim past: at the turning point it is not the provider that changes, it is the kind of capital that captures the value. In the build-out phase the money is made by whoever builds and finances the infrastructure. In deployment the money is made by whoever applies it. Value shifts from financial capital to productive capital, from those who raise data centers to those who turn them into an advantage in their own business. Not the lab that builds the model, but the machine builder who pours thirty years of fault and service data into a system a competitor would need years to assemble.
Applied to AI, this is uncomfortable for both camps. The frontier labs are the financial capital of this round, brilliant and overfunded, and their moment of peak valuation sits in installation, not after it. But the liberation Colin sees is no liberation either. The turning point does not hand the value to everyone who now switches to a cheap model. It hands it to those who build an advantage out of the application that no model swap can collect back. Perez calls it deployment. An owner calls it the homework you do before the crash.
Anthropic seems to have understood this. It is turning Claude into the workplace for programming, while the model underneath becomes a commodity. Whoever holds the chokepoint no longer sells the model, but the place where the work happens.
Three questions that can no longer be delegated
Three questions turn sharp, the ones you could comfortably push to a provider in the cheap phase. First: who can switch off the models your operation runs on? Second: whose balance sheet takes the hit when the switch is pulled, the provider’s or your own? Third: what of all this can still be outsourced at all?
The honest answer to the third is the whole piece. The model is delegable, it is even meant to be replaceable. Compute is delegable, it becomes a commodity. Not delegable are the two layers your edge hangs on: your own data, the accumulated knowledge from documents, decisions, histories, and the ability to swap the model above it with a single line of configuration. Whoever owns or can organize those two can withstand any switch above them, because they can replace the provider underneath. Whoever does not, only gets to choose which landlord they belong to.
The value arrives in two waves
So the forecast, plainly. The value ends up in one place: your own data and the ability to swap the model above it in days. It arrives there in two waves. The first is known, it slid off the model onto the data and the orchestration around it, and whoever was paying attention has priced it in. The second is just beginning. When confidence drains from the market, the capital that chased ever bigger models has to go somewhere, and it moves onto exactly this layer. That repricing is in no valuation yet.
What follows is not the same for owners and investors. The investor allocates, buying onto this layer before the capital crowds in after the crash, taking the repricing with him. The owner builds, securing their own data and the ability to switch while the exercise is still cheap and unhurried, not once everyone needs it. One shapes with money, the other with their own operation. Both windows share the same expiry date, the moment the herd starts to run.
Whoever sowed before the direction turns harvests the falling price and a rising value in their own business. Whoever only grabbed the cheap model changed landlords and calls it independence.
