Jason A. Hoffman

Jason A. Hoffman, PhD

What is it about human beings that we keep arguing against Jevons’ paradox?

In 1865, William Stanley Jevons published The Coal Question and observed that more efficient steam engines had not reduced Britain’s coal consumption. They had increased it. Efficiency lowered the cost of work done with coal, the lower cost admitted new uses, and the new uses consumed more coal than the old inefficiency ever had. The energy-economics literature has spent the years since studying the pattern under the name of the rebound effect, and its record carries a distinction worth stating: rebound, use expanding when cost falls, is robust across domains; full backfire, consumption rising past its pre-efficiency level, is domain-dependent, and it is strongest where demand is elastic and the latent uses run deep. The argument that gets re-run in every technology era is the one the record least supports: that this time, efficiency will satiate us. The interesting question is not whether Jevons was right. It is why we keep flinching.

I can find three roots, and they are not equally deep.

The first is a bias in the strict sense. Our only native model of demand is the body, and bodies satiate. Hunger ends. Sleep completes. Every appetite we have direct experience of is bounded, so we project the bound onto aggregates: once there is enough light, enough bandwidth, enough compute, we will be done. But aggregate demand is not a quantity. It is a curve. Every efficiency gain admits the tranche of uses that were previously uneconomic, and the tranche was always there, waiting on the price. Individual appetites saturate; markets do not. The error is scaling personal physiology up to civilization, and it is hard to unlearn because the prior was installed by the body rather than by argument.

The second root is structural rather than psychological. The consumption Jevons predicts comes from uses that do not exist yet, so they cannot be enumerated in the argument. This is Bastiat’s distinction applied to demand: the saved fuel is seen; the marginal user the saving creates is unseen. Nobody in 1995 could say what cheap bandwidth was for, because the answer was YouTube and the answer did not exist for another decade. Reasoning stops at the visible demand set because the visible demand set is all there is to reason from.

This root reads differently for one profession. Latent demand being invisible is not a reasoning failure to lament. It is the arbitrage. The product person’s entire job is seeing the tranche below the current price point before it exists. Everyone else’s bias is their alpha, which makes product work professional belief in Jevons.

The third root is the deepest, and it is moral rather than cognitive. Efficiency is coded as thrift, and thrift is supposed to be rewarded with restraint: use less, want less, be finished sooner. Jevons says the reward for thrift is appetite. Accepting that means accepting something uncomfortable about what the market is. It is not an external degenerate we can moralize at from a distance. Aggregate demand is us, summed. Every individual purchase arrives wrapped in a story: for work, for the kids, self-care. Aggregation deletes the stories and leaves revealed preference, and wanting without narrative cover is what addiction looks like. The psychology literature has a precise version of the cut in Berridge’s distinction between wanting and liking, two systems that can and do decouple. The market ledgers wanting only. Satisfaction appears nowhere in GDP.

Underneath all three roots is one mistake about prediction. The past predicts the future exactly insofar as the process generating it is stationary, and what you are entitled to extrapolate is the generator, not the values. Hume showed there is no non-circular justification for even that much; induction is unjustifiable and unavoidable at the same time. But if you are going to induce, induce on the thing that repeats. The anti-Jevons argument induces on levels: current demand, current uses, current prices. The 160-year record does not repeat at the level of any of those. It repeats at the level of the response function: use responds when cost falls, and it overshoots into backfire exactly where demand is elastic and latent uses run deep. Wanting is stationary. Particular wants are not. The live question for a new domain is not whether Jevons applies. It is how much elasticity is waiting.

The same flinch appears wherever aggregation can be refused. Jevons denial refuses to aggregate over uses; “I am healthy, why fund healthcare” refuses to aggregate over time; “I have no kids, why fund schools” refuses to aggregate over people. Each assumes stationarity of the self while denying stationarity of the system, which is exactly backwards. Every one is a claim to be finished.

The version in front of us right now is token pricing. The analysis takes a point-in-time snapshot of token costs and prices, extrapolates the decline, and concludes that foundation model companies are doomed. Same autopsy. The analysis prices tomorrow’s costs against today’s demand set, when the price decline is the tranche-admission mechanism itself. A use that pencils out at sixty dollars per million tokens is a different market from the one that appears at six dollars, and both are different from the uses that arrive at cents. Each demand class is invisible until its price point exists. This is the 1995 bandwidth error, re-run with better spreadsheets. And token demand expands on three axes at once: more users, more tasks, and more tokens per task.

There is a second mistake stacked on the first: the analysis treats the model as if it were the company. A frozen model is a depreciating asset whose price falls toward marginal cost as fast followers match it. True, and not the question. The model is a sample; the company is the generator. The margin lives in the gap between the frontier and the trailing edge, which is a flow, not a stock. The third axis is Jevons operating inside the task: as price per token fell, tokens per job exploded, because reasoning chains and agents spend efficiency gains deeper inside the work. The want was never tokens. The want was cognition, and cognition sits on the wanting side of the ledger. And if backfire lives where latent uses run deep, cognition is the deepest domain there is: the backlog is everything not yet worth doing at the current price of thinking.

The honest cut is that the induction says the market grows. It does not say any given lab survives. Some will not. But that is a share question, and the doom analysis derived it from a size argument, which is the tell that it is flinch rather than analysis. The demand-curve axis in the Vertical Machine series is this same question asked from inside the machine.

And the counterparty has already voted. The gigawatt build-out is not a bet on today’s demand set. It is a bet on the generator, priced by the operators holding the best proprietary demand curves in the world. You can argue with the induction. You are arguing with the order book.

Arguing against Jevons is arguing that we are finishable. We are not.