The Jevons Paradox
When making something more efficient makes us use more of it — an 1865 idea that became AI's favorite argument.
The Coal Question
In 1865, Britain ran on coal, and a 29-year-old economist named William Stanley Jevons published a book asking how long that could last. The comfortable answer of his day was that it would last longer and longer: James Watt's steam engine extracted several times more work from a ton of coal than the Newcomen engine it replaced, and every further improvement would surely stretch the reserves further.
Jevons looked at the record and found the opposite. Each leap in engine efficiency had been followed by Britain burning more coal, not less. His explanation was blunt:
"It is wholly a confusion of ideas to suppose that the economical use of fuel is equivalent to a diminished consumption. The very contrary is the truth."
— William Stanley Jevons, The Coal Question (1865)
Efficiency had made steam power cheap enough for uses that were previously unthinkable — pumping deeper mines, driving mills, hauling trains, crossing oceans. Every improvement widened the set of places where burning coal made economic sense. Efficiency didn't shrink coal's role in the world. It multiplied it.
The Mechanism
There is nothing mystical here — the paradox is a price effect. Nobody wants coal, or tokens, for their own sake. What you actually buy is what the resource does: motion, heat, light, answers. When efficiency improves, the effective price of that useful work falls. And when a price falls, two things happen: existing users consume more, and marginal uses that never made sense before suddenly do.
Whether total consumption ends up lower or higher turns on one hidden variable: how hungry demand is at the new price — what economists call elasticity. If demand is nearly satisfied, efficiency delivers what intuition promises: the same work from less resource. If demand responds but modestly, part of the saving is eaten — the rebound effect. And if demand is highly elastic, new consumption outruns the efficiency gain entirely. That full reversal — called backfire — is the Jevons Paradox proper.
"Jevons Paradox Strikes Again!"
In January 2025, the Chinese lab DeepSeek released R1, a reasoning model near the level of OpenAI's o1 at roughly one-thirtieth of the price per token. The market read it the intuitive way: if intelligence just got that much more efficient, the world would need far fewer chips. Nvidia lost about $600 billion in market value in a single day — the largest one-day loss any company had ever recorded.
Microsoft's CEO Satya Nadella replied with a 160-year-old idea: "Jevons paradox strikes again! As AI gets more efficient and accessible, we will see its use skyrocket, turning it into a commodity we just can't get enough of."
The time since has mostly sided with him — so far. The price of a given level of capability has kept collapsing, yet total token consumption, data-center investment, and energy demand have kept climbing. Cheap intelligence did not shrink the market; it found new places to be spent: agents that run for hours, coding tools that burn millions of tokens per task, context windows that hold whole codebases. The pattern keeps refreshing itself — GLM-5.2, an open-weight frontier model at roughly a sixth of its Western rivals' price, is the current data point, and its release was met not with less demand for compute but with more.
Worth keeping in view: "so far" is doing real work in that paragraph. The paradox names a recurring pattern, not a guarantee. If today's build-out overshoots actual demand, the story gets retold as a bubble. Both readings currently have serious defenders.
The Knobs
Like any intuition pump, the Jevons Paradox has hidden knobs — parameters that quietly decide whether it applies. Turn them and the conclusion changes.
The elasticity knob
The paradox needs hungry demand. Lighting backfired for two centuries — cheaper light meant more lit hours, more lit rooms, more lit cities. Then demand saturated: nobody lights their living room more because LEDs are cheap, and total lighting energy in rich countries has actually fallen. Efficiency wins in the end — once people have enough.
The bottleneck knob
Consumption can only grow until it hits the next constraint. For AI, candidate walls include power grids, chip fabrication, capital — and human attention, which no efficiency gain has ever multiplied. A paradox can be real and still get stopped by whatever is scarce next.
The time-horizon knob
Jevons himself extrapolated wrong. He feared Britain's greatness would exhaust itself with its coal; British coal use peaked in 1913 and is near zero today. Backfire dominated for decades — then substitution and saturation won. Ask not just whether the paradox applies, but for how long.
Economists genuinely dispute how cleanly any of this maps onto AI. Demand for energy is bounded by physics and comfort; demand for intelligence might be bounded by very little — or by more than we think. That open question is exactly why the lens is worth carrying.
How to Use It
The next time someone says "this efficiency gain means we'll need less" — of compute, of energy, of programmers, of anything — run three questions:
1. Is demand elastic? Are there uses waiting at a lower price, or is the need mostly met? The paradox only fires when cheaper means hungrier. Nobody heats their house more because the furnace got efficient — but people ask a model things they would never have paid a consultant to answer.
2. What unlocks? Which previously-absurd uses become reasonable at a tenth of the cost? The interesting answers are never "current uses, done cheaper." They are the uses nobody bothered to imagine at the old price — agents that run for hours, code review on every commit, a tutor for every student.
3. Where does the bottleneck move? Consumption grows until it hits the next constraint. For AI that might be power grids, chip supply, capital — or human attention, which no efficiency gain has ever multiplied. Name the next wall, and you know where the story goes after the paradox plays out.
And run it in reverse, too. When someone declares that usage will inevitably skyrocket because "Jevons," ask the same questions. The paradox is often invoked as a spell — one word that settles the argument. It settles nothing. It only tells you where to look.
Key Takeaways
- In 1865, Jevons observed that more efficient steam engines led Britain to burn more coal, not less
- Efficiency is a price cut on what a resource does; if demand is elastic, new uses outgrow the savings
- Rebound means part of the saving is eaten; backfire — the full paradox — means consumption rises outright
- DeepSeek R1's cheap reasoning crashed Nvidia for a day; Nadella's "Jevons paradox strikes again" has looked mostly right since
- The paradox has knobs — elasticity, bottlenecks, time horizon — and it eventually broke for both coal and lighting
- Use it as a lens, not a law: ask what unlocks at the new price, and what becomes scarce next
