The two arguments on this site, and why they are one argument
The front page of Crash Lab watches a financial bubble: the debt paying for AI data centres, the behind it, the leverage that has moved out of banks into places nobody has to . Six times a day it asks how close that is to breaking.
This section is about what happens next, and it starts from an uncomfortable place. We built the original CrashLab in 2025 and 2026 to argue that AI would displace a large share of the workforce and that the economy has no answer ready. Then we built a machine to watch the AI boom, and the machine keeps reporting that the boom is financed on terms that cannot hold.
A reader is entitled to ask whether those two things contradict each other. If the AI trade is a bubble, does that not mean the job losses were overstated too?
No. The crash does not refute the displacement argument. It accelerates it. That is the claim this section exists to make, and it rests on four things.
1. The bubble is in the financing, not the capability
Nothing about a credit event makes a trained model worse at its job. The thing that is overextended is the capital structure around AI — vendor guarantees, backstops on buildings full of chips, twenty-four-year amortisation schedules against five-year assets, private credit funding data centres at that assume nothing goes wrong. That is what we write about six times a day, and that is what breaks.
When it breaks, the models keep running. The capability was never the fragile part.
2. A crash makes AI cheaper to deploy, not dearer
This is the part people get backwards. An overbuild does not vanish when its financiers do — it gets sold.
The precedent is exact and it is the one the entry keeps returning to: the telecoms of 1999 to 2001 laid enormous quantities of fibre on borrowed money, went bankrupt, and the fibre stayed in the ground. It was bought for cents by companies that had not paid to install it, and it became the cheap bandwidth that made the next decade of the internet possible. The capital was destroyed. The capacity was not.
Do that to compute and the result is the same. Data centres and accelerators come out of bankruptcy at a fraction of build cost, in the hands of operators with no debt to service. The cost of running a model falls exactly when every company in the economy is desperate to cut costs.
3. After a crash, cutting costs stops being a and becomes survival
In an ordinary year, a company automating a department is making a choice. After a crash it is not.
Revenue falls. Credit gets expensive or stops being available at all — that is what a credit event means. Boards and investors stop asking about growth and start asking about the cost line, and they ask every quarter. Every firm in the economy is under the same instruction at the same time: get the cost base down, now.
Labour is the largest line on most cost bases. And for the first time there is a tool sitting there that can take a real bite out of it, freshly cheapened by the same crash that created the pressure. A management team that hesitates gets replaced by one that will not.
This is the leg of the argument we think is most underrated, and it is the one that turns a slow trend into an event. Companies that would have automated a department over five years do it in two quarters, because the alternative is not slower growth — it is not making it.
4. Under competition, replacement is not optional anyway
The original argument stands on its own and the crash only sharpens it. Once one firm in an industry demonstrates that the work can be done for a tenth of the payroll, every competitor has to match it or lose on price. The first movers are rewarded immediately — margin, share, a stock price that goes up on the announcement. The holdouts are punished just as fast.
Nobody has to believe AI is good for society for this to happen. They only have to believe their competitor will do it.
So the sequence is worse than either half alone
Put those together and you do not get a crash or mass displacement. You get them in an order that makes both worse:
- The AI financing bubble bursts. Capital is destroyed, credit tightens, the economy takes the ordinary damage a financial crisis does.
- The capability survives and its cost collapses, because the infrastructure was built with somebody else's money and is now for sale.
- Every firm is simultaneously under maximum pressure to cut costs, holding a newly cheap tool that cuts the biggest cost there is.
- The layoffs land on an economy that has already spent its cushion on the crisis — with a weakened tax base, elevated debt, and a central bank that has already used its tools.
The standard response to mass unemployment is for the government to replace lost income until demand recovers. That is what 2008 and 2020 did. It works when the shock is cyclical, the number of people is a few percent of the workforce, and the government goes into it with fiscal room.
None of those three will be true. The displacement is not cyclical — the jobs do not come back, because the thing that replaced them is cheaper and does not get tired. The number is not a few percent. And the fiscal room will already be gone, spent on the crash.
What we are honest about
We do not know the timing. The front page measures the crash six times a day, from evidence, and says plainly when it is guessing. This section is an argument about what follows, and arguments about the future are not measurements. The displacement could begin before the crash, or years after it, or in a slow grind that never has a date attached.
What we are confident about is the ordering — that a crash in AI financing makes AI-driven displacement faster and more brutal, not slower and gentler. And that is the opposite of what most people assume when they hear the word bubble.
Where this goes
There are three pages under this one. Read them in order if you have the time.
Why the jobs go — what makes this technology different from every previous wave, why replacement happens in a step rather than a slope, and why the new jobs are not coming.
The spiral, and why UBI cannot stop it — how layoffs cause layoffs, and the four separate reasons the one policy answer on offer does not add up. This is the arithmetic, and it is worse than it is usually presented.
Three futures — the same starting conditions, three policy responses, three endings. Two of them are bad.
The conclusion those pages arrive at is not a technocratic one, and we are not going to dress it up as one. If the machines can do the work, the question of who eats is decided entirely by who owns the machines. Redistribution does not settle it, because redistribution leaves ownership where it is and has to be re-fought every year, from a tax base that is shrinking. The thing that has to change is ownership, and changing who owns the productive capacity of a society is not a policy adjustment. It is a revolution, and we should call it one.