Why AI Breaks the Economy at the Cost Curve, Not the Job Count

Almost every public argument about AI and work makes the same mistake. It treats an unprecedented event as a familiar one, then reaches for the nearest comforting analogy. AI will take some jobs but create others. We survived automation before. Regulation will keep humans central. Society adapted to the Industrial Revolution and it will adapt again.

These are stories, not analysis. They all assume AI is another tool that slots into the economy the way the spreadsheet or the loom did. It isn’t. AI is cognition supplied at scale, and cognition is the input that every white-collar industry runs on. When you automate the input that the whole system depends on, the comforting analogies stop applying. The Discontinuity Thesis is the attempt to follow that fact through to its conclusion instead of flinching from it.

Unit Cost Dominance is the whole engine

The argument rests on one plain observation. When AI can produce cognitive output cheaper, faster and at acceptable quality compared with a human, firms move toward AI in that domain. Not because anyone is enthusiastic about it, but because a firm that doesn’t is undercut by one that does.

This is the point I call Unit Cost Dominance. A rational firm picks the production method with the lowest marginal cost. AI, especially when paired with a single human checking the output rather than a team producing it, pushes the marginal cost of cognition toward the floor. The reasons are unglamorous and they compound:

• No salaries, healthcare, pensions or holiday.
• No skill plateau, and a step up in capability with each model release.
• Near-zero cost to replicate, and effectively instant scaling.

Once a production method has those properties, substitution stops being a choice. If one firm in a sector automates its cognitive work and the cost gap is real, the rest follow or they lose. This is the multiplayer prisoner’s dilemma in plain clothes. Everyone might prefer a world where no one defects, and everyone defects anyway, because the firm that holds out is simply outcompeted by the firm that doesn’t. Goodwill loses to the cost curve every time.

The economy runs on participation, not just consumption

Most economists arguing about AI fixate on jobs: how many are lost, how many appear, how fast people reskill. That framing misses the load-bearing question, which is about productive participation, not headcount.

The distinction matters because two systems can look identical from the outside and be completely different underneath:

• In one, most adults do economically valuable work, earn wages, and spend them. Production and demand are wired together through the same people.
• In the other, most adults produce little but keep consuming through redistribution. The wage-demand circuit has been cut and replaced with a transfer.

Post-war capitalism is defined by the first arrangement: mass participation in production, not merely mass consumption. The wage is what links the two. Once AI outcompetes human cognition across enough sectors, that participation thins out, and the link breaks.

You can still keep everyone fed. Dividends, a national fund, UBI, whatever the mechanism. But that does not preserve the system, it preserves consumption while the engine underneath is swapped out. This is where the optimistic stories quietly cheat: they downgrade capitalism to a distribution scheme and ignore that the productive side of the circuit is the thing actually at risk.

Why regulation and “human-only zones” don’t hold

The hopeful counter-arguments all reduce to a few moves. We’ll regulate AI. We’ll ring-fence human-only tasks. We’ll tax usage. We’ll slow deployment to protect jobs. Each runs into a structural problem.

Coordination is the thing we are worst at

If one country deploys AI fully across its economy, it gets higher productivity, lower costs, cheaper exports, faster innovation and a military edge. Every other country then has to follow or accept becoming less competitive. There is no body, not the UN, the G7 or the OECD, with the authority to stop that. We have not solved tax competition, and we have barely moved on climate, both of which are slower and lower-stakes than this. Expecting AI to be the coordination problem we finally crack is not a plan, it is hope wearing a suit.

You cannot draw a stable line through cognitive work

This is the part the regulatory proposals never survive. Cognitive tasks do not come in discrete units you can fence off. They sit on a gradient, and AI walks up the gradient one notch at a time. What starts as support becomes substitution, and you rarely notice the line being crossed:

• Spell-check, then writing suggestions, then drafting, then full composition, then the decision the writing was for.
• Code completion, then code generation, then whole-system design.
• Image touch-ups, then full image generation, then the creative direction itself.

This is the boundary problem, the Sorites paradox applied to labour. At no single step can you say “here is where the human task ends and the machine task begins,” because each step is only marginally more capable than the last. Arms control could work because warheads are countable, discrete objects you can inspect and cap. Cognition has no such edges. You cannot regulate a gradient, and because cognition is continuous rather than discrete, there is no stable place to put the fence. That is what quietly kills most serious “human-only” proposals, not bad faith but a category error about what is being protected.

The strongest objection, and where it breaks

The best argument against all of this is the historical analogy. AI is like every previous technological wave. It destroys some work, creates new work, and people move up the value chain to do the things machines can’t. It is a good argument because it has been right before. The problem is what makes this wave different.

Every past revolution automated one layer and left cognition to humans. Electricity and the engine automated muscle. Computers automated arithmetic. The internet automated distribution. In each case the human moved “up” to the cognitive work of supervising, deciding and directing the new machines. The ladder had a higher rung to climb to, and that rung was always cognition.

AI automates the rung itself. When cognition becomes cheap, fast and scalable, there is no higher cognitive layer for displaced workers to retreat into, because that layer is exactly what is being automated. “New jobs will emerge” was a reasonable inference when the next rung was reliably human. It becomes an article of faith when the mechanism that produced those rungs, human cognitive advantage, is the thing under competition. For the analogy to hold, there has to be an absorption channel: somewhere productive, wage-sustaining and scalable for displaced workers to go. Past revolutions had one. The case I have not seen answered is which channel survives when the competitor is general cognition rather than a single-purpose tool.

What comes after the wage economy

None of this is a prediction that society ends. The claim is narrower and, I think, harder to dodge: the specific arrangement we have lived under for roughly eighty years ends, and what replaces it is not capitalism but something else wearing some of its clothes.

The political form is wide open. It could be AI-funded UBI, national AI dividends, sovereign ownership of the models, compute cooperatives, a techno-socialist settlement or an AI-powered neo-feudalism where a few owners capture nearly everything. The economic substance is the same across all of them. Humans become consumers rather than producers, and value creation runs on automated cognition.

This is not automatically dystopian. A world where a national compute fund pays every citizen a comfortable income could be genuinely prosperous. But it is a discontinuity, a break rather than a continuation, and it is worth being honest about that. If every citizen receives a generous payment from a state-owned AI engine, capitalism did not survive. It ended and was replaced by a post-labour system that keeps people comfortable while the productive engine changes species. Redistribution can make the transition humane. It cannot make it a continuation.

Where this leaves us

Strip out the reassurance and the chain is short. Once AI becomes the cheapest producer of cognition in a domain, competitive defection forces adoption across that domain. As this spreads, productive participation falls. The wage-demand circuit that defines this version of capitalism comes apart, and a successor system has to take its place. The argument is structural and probabilistic, not a guarantee with a date on it, but each link follows from the one before for ordinary economic reasons, not exotic ones.

It also fits things the comfortable model struggles with: why white-collar wages stall while output rises, why productivity gains stop translating into hiring, why “future-proof” roles keep getting shorter half-lives, why any AI advantage a firm finds is common within weeks, and why capital owners are gaining ground this fast. None of that requires the thesis to be true, but it is what the thesis predicts, and the optimistic story has to keep explaining it away.

I am not asking anyone to accept a verdict. I am asking for the missing piece. If the historical analogy is going to hold this time, name the absorption channel that is wage-sustaining, scalable and resistant to cheaper machine cognition. Until someone does, the discontinuity is the more honest reading of where this is heading.