Here is a commonsense intuition: if every artificial-intelligence computation consumes less energy, the total should fall. It is false, and indeed the opposite happens. As AI grows more efficient and cheaper, its usage soars, and total consumption climbs instead of receding. An English economist named this trap one hundred and sixty years ago, about coal. It is called the Jevons paradox, and it has just struck the most watched technology of our age.
1 "Jevons strikes again"
It all begins on a day of market panic.
The scene
The day cheap AI shook Wall Street
On 27 January 2025, the stock of chipmaker Nvidia lost nearly $589 billion of value in a single session, the largest single-day loss in American market history. The cause: a Chinese start-up, DeepSeek, had just shown that cutting-edge AI could be had for far less. The same day, Microsoft chief Satya Nadella posted a line that became famous: "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."
Nvidia, 27 January 2025
−$589 bn
in one session, an all-time Wall Street record.
Nadella's word
Jevons
an 1865 paradox propelled to the heart of the debate.
The reversal
The market first read efficiency as a threat: if AI is cheaper to run, perhaps fewer chips will be needed. Nadella read the opposite, and history is proving him right for now: efficiency does not cut the bill, it unleashes it. The cheaper the computation, the more we want of it. What remains is to understand why this trap is so reliable.
2 An idea from 1865
Let us go back to the source.
The origin
Coal, the steam engine and the error of intuition
In 1865, the economist William Stanley Jevons observed a troubling fact: steam engines that burned ever less coal had not lowered England's coal consumption, they had sent it soaring. The cheaper steam became, the more it was used, in more industries, for more purposes. His formulation has endured: "It is a confusion of ideas to suppose that the economical use of fuel is equivalent to diminished consumption. The very contrary is the truth."
Rebound and "backfire"
Economists have refined the idea. The rebound effect measures the share of energy savings reclaimed by the rise in usage that the lower price induces. As long as the rebound stays below 100%, efficiency still reduces consumption. But when it exceeds 100%, we speak of "backfire": total consumption increases because of efficiency, not despite it. This is precisely what the Jevons paradox describes. Three conditions trigger it: progress that lowers the cost of use, demand highly sensitive to price, and a market still far from saturation.
3 The DeepSeek shock
Yet AI meets these three conditions to perfection.
The trigger
Cutting-edge AI at a fraction of the price
In January 2025, DeepSeek released a reasoning model, R1, presented as rivalling the best American systems. The reported training cost of its base model: about $5.6 million, against a hundred million or so for comparable Western models. The figure is disputed, it covers only the final training run, but the signal is plain: cutting-edge AI is becoming affordable. Venture capitalist Marc Andreessen called it "AI's Sputnik moment."
The founding misunderstanding
Note the irony: what panicked the markets was the announcement of cheaper AI. The short-term logic said: less spending, hence fewer chips, hence less profit for Nvidia. The logic of Jevons answers: cheaper means more uses, hence in time more computing power demanded, not less. The whole debate over AI's energy hinges on this gap in reading.
4 The collapsing price
For the collapse in costs is dizzying.
The fall
Ten times cheaper every year
The cost of running an AI at a given level of performance is collapsing by about 40 times a year, according to Epoch AI's measurements. Sam Altman, OpenAI's chief, puts it another way: "The cost to use a given level of AI falls about 10× every 12 months, and lower prices lead to much more use." To match the performance of a 2022 model, you had to spend some $20 per million words processed; two years later, a few cents were enough.
Cost at equal performance
÷ 40 / yr
the measured fall in the price of inference (Epoch AI).
Altman's rule
÷ 10 / yr
"lower prices lead to much more use."
The line that says it all
"Lower prices lead to much more use." Without naming it, Altman describes exactly the Jevons mechanism: the fall in price does not lower total spending, because it unlocks latent demand. Every cent saved per computation opens up whole uses, unthinkable yesterday, commonplace today.
5 The exploding usage
And usage follows, at a staggering pace.
The evidence
Fifty times more computation in a year
The figures speak. In thirteen months, the volume of text processed each month by Google's models rose from about 9.7 trillion to 480 trillion tokens, a fiftyfold increase. ChatGPT doubled its weekly audience in 2025, from 400 to more than 800 million users. And Nvidia shipped 3.7 million graphics processors in 2024, more than a million above 2023, even as each chip was more efficient than the last. Efficiency climbs, and volumes climb faster.
Monthly computation (Google)
× 50
in thirteen months, through May 2025.
Chips shipped (Nvidia)
3.7 M
in 2024, +1 million despite greater efficiency.
The snake eating its own tail
Our readers will recognize a cousin of the paradox we have already explored: an AI whose success forever calls for more means to feed it. Where the analysis of demand showed the commercial gearing, the Jevons paradox shows its physical translation: every gain in efficiency is at once swallowed by a surge in usage. Efficiency per computation is real; it is simply outpaced by the number of computations.
6 The rebound of reasoning
A new development even amplifies the movement.
The turning point
When AI "thinks," it consumes ten to a hundred times more
Recent models no longer merely answer: they "reason," unspooling a long internal monologue before they conclude. This reasoning swallows ten to a hundred times more computation per question than a simple answer. The cost per word falls, but the number of words per task explodes. And energy shifts from training, a one-off, toward inference, that is, daily use, which already accounts for 80 to 90% of computation and will run without end, billions of times a day.
Why efficiency is never enough
This is the heart of the paradox applied to AI: a model is trained once and queried endlessly. As long as every drop in cost opens new uses, autonomous agents, generated video, ever-present assistants, efficiency per query is caught up and then surpassed by the multiplication of queries. To make AI leaner per computation is to make possible a number of computations with no apparent limit.
7 The grid and the water
At the end of the chain, there are wires and rivers.
The physical bill
The price of the electron and the drop
According to the International Energy Agency, data centers consumed about 415 terawatt-hours of electricity in 2024, or 1.5% of the world's electricity; this figure would double by 2030, to nearly 945, the equivalent of Japan's entire consumption, with AI as the prime driver. In the United States, residential bills climbed about 11% in 2025, and Virginia, the world capital of data centers, had to create a special tariff category so that households would not subsidize the servers. To this is added cooling water, tens of billions of liters a year.
Data centers, 2024 → 2030
415 → 945
terawatt-hours, a doubling (IEA).
US bills, 2025
≈ +11%
on the residential price of electricity (EESI).
The return of the atom
A striking symbol of this hunger: to power its servers, Microsoft has signed to restart a reactor at Three Mile Island, the site of America's worst nuclear accident. Google and Amazon, for their part, are betting on small modular reactors. AI, meant to usher us into an immaterial age, is reviving power plants and rekindling the coal we were about to put out.
8 Maybe it's not so simple
Let us stay rigorous: the thesis has its limits.
The counter-arguments
What could belie the worst
The rebound is an observed fact; "backfire" beyond 100% remains a thesis. Three serious objections deserve a hearing, and they do not say "AI consumes nothing."
Three nuances, given equal weight
①
Decoupling has already happened. Between 2010 and 2018, data center computation grew more than sixfold, yet their energy consumption rose only 6% (Science journal, 2020). Efficiency can therefore, for a time, prevail over usage.
②
The grid may hit a ceiling. Connecting a new center takes four to ten years, against two to three to build it. For want of electrons, part of the demand simply will not be able to materialize.
③
Projections diverge enormously. Consumption forecasts for 2030 range from single to fivefold depending on assumptions; some, alarmist, have already been belied. The uncertainty is real.
A conditional validity
The economist Steve Sorrell, a reference on the subject, reminds us that backfire at the scale of an entire economy is rarely demonstrated: often plausible, hard to prove. The honest position therefore holds in a single sentence: AI's rebound is solidly observed in the short term, but to attribute it mechanically to "backfire" remains an interpretation, not a law.
9 Leaner, really?
What remains is to draw the lesson.
The moral of the paradox
Necessary, but insufficient
The great lesson of Jevons is not that we should renounce efficiency, which would be absurd. It is that efficiency is not enough. As long as nothing caps consumption, neither a carbon price, nor a physical limit, nor a political will, every gain in efficiency turns into a surge in usage. The leanness of a computation does not make the leanness of a system. Counting the watts per query is pointless if one does not count the queries.
The compass
①
For the citizen. Be wary of the "our AI is greener" argument: it can be true per computation and false in total. The right question is not unit efficiency, but aggregate consumption.
②
For the investor. The rebound illuminates the rush for energy: nuclear power, grids, cooling, metals. AI's constraint may not be the chip, but the electron. Which makes the power grid the real chokepoint.
③
Clear-sightedness. Aim not for "efficient computation," but for "net-zero computation": efficiency in the service of an energy envelope that, for its part, does not grow without end.
The opposite of intuition
Jevons wrote it already: "the very contrary is the truth." One hundred and sixty years later, the line applies word for word to artificial intelligence. The leaner it grows per computation, the more it devours in total, because its very leanness is what multiplies its usage. Efficiency is not the solution: without a safeguard, it is the fuel of the problem.
Read alongside: the snake eating its own tail, the other face of the AI paradox. Reference: abbreviations & acronyms (TWh, AI, IEA, GPU).