The Jevons Paradox: Why Making AI Cheaper Makes Us Use Far More

In 1865 a young English economist named William Stanley Jevons noticed something that ought to have been reassuring and was instead deeply unsettling. Britain’s steam engines were getting dramatically more efficient — squeezing more work out of every lump of coal — and yet the country was burning far more coal than ever before, not less. The Jevons paradox, as it came to be known, is the observation that when a resource becomes more efficient to use, we don’t bank the savings. We use so much more of it that total consumption climbs. I keep returning to this nineteenth-century coal story because it is the clearest lens I know for what is happening right now with artificial intelligence, and for why the promise that AI will keep getting “cheaper and greener” deserves a hard second look.

What Jevons actually saw

Jevons was writing at the height of Britain’s coal anxiety. The engines of the Industrial Revolution ran on coal, and thoughtful people worried the country would run out. The intuitive answer was efficiency: build engines that waste less, and the same coal will last longer. James Watt’s improved steam engine had done exactly that, using a fraction of the fuel of the earlier Newcomen design to do the same job.

Then Jevons pointed at the ledgers. As engines got more efficient, coal consumption didn’t fall — it exploded. His explanation was elegant and uncomfortable. Efficiency doesn’t just let you do the same work for less fuel. It makes the work itself cheaper, and cheaper things get done far more often. A more efficient engine made steam power economical for factories, mines, railways and ships that could never have justified it before. Every drop in the cost of using coal opened a new frontier of places worth using it. The savings per unit were real; they were simply overwhelmed by the explosion in the number of units.

Efficiency doesn’t just let you do the same work for less fuel. It makes the work itself cheaper, and cheaper things get done far more often.

This is the crucial move, and it is worth slowing down on. We tend to treat efficiency and conservation as the same virtue. They are not. Conservation is a decision to use less. Efficiency is a decision to get more out of each unit — and getting more out of each unit is, historically, one of the most reliable ways to make people want a great deal more of it. Jevons was not saying efficiency is bad. He was saying that efficiency, left to the logic of markets, is an accelerant, not a brake.

Why AI is a textbook case

Now hold that idea up against the last few years of AI. Almost every headline about the technology is, at bottom, an efficiency story. Models are getting cheaper to run. Chips are getting faster per watt. Techniques like quantisation, distillation and smarter inference keep cutting the cost of producing a given answer. The reassuring narrative writes itself: the environmental cost of AI will take care of itself, because each query keeps getting lighter.

Jevons would recognise this reasoning instantly, and he would tell you to watch the total, not the unit. When the cost of a single AI query falls, we do not run the same number of queries and pocket the difference. We put AI into places that could never have justified it before. A capability that was a costly novelty becomes a default. It gets embedded in every search box, every document editor, every customer-service flow, every phone. Features that would have been absurd to compute a few years ago — summarise this, rewrite that, generate an image for a throwaday message — become free-feeling reflexes precisely because the per-use cost collapsed.

Each of those uses is cheap. There are just an astonishing and rapidly growing number of them. And every one of them runs somewhere physical. This is the part the word “cloud” is designed to make us forget — that the cloud is physical, a landscape of enormous data centres full of hot silicon that has to be powered and cooled around the clock. Efficiency gains at the chip level don’t escape that reality. They feed it, by making it economical to build more of it.

The resource bill nobody put on the invoice

So what does the total actually look like? Here I have to be careful, because precise figures in this space are slippery and often quietly self-serving. But the direction is not in serious dispute. By many estimates, data-centre electricity demand — driven substantially by AI — is on a steep upward curve, and several forecasts have it roughly doubling over the second half of this decade. Grid operators in a number of regions have reported pausing or reconsidering plans to retire fossil-fuel plants specifically because of the new load. I’ve written more about that trajectory in AI’s energy consumption; the short version is that the appetite is growing faster than the efficiency savings are shrinking it.

Electricity is only half the story. The other half is water. Large data centres are commonly cooled with water — a lot of it, reportedly, evaporated away for good rather than returned — and the newer AI-heavy facilities are among the thirstiest. That matters enormously depending on where the cooling happens. A water-hungry facility in a wet, cold region is one thing. The same facility in a hot, water-stressed one is quite another, which is exactly the tension I dug into in AI’s water use in India. The resource is drawn from a shared, local, finite pool, whether or not the people drawing from that pool ever get to use the AI it powers.

The person whose reservoir is drawn down to cool a data centre is very rarely the person whose query it answered.

And that is the question Jevons’s paradox forces to the surface — not just “how much,” but “who pays?” The efficiency gains flow to the companies building and selling AI, and to the users who get a smoother, cheaper product. The resource bill — the electricity that stresses a grid, the water that evaporates from a stressed watershed, the emissions from a coal plant kept running past its retirement — lands somewhere else entirely. The person whose reservoir is drawn down to cool a data centre is very rarely the person whose query it answered. Efficiency, in other words, doesn’t just multiply consumption. It widens the distance between the people who enjoy the benefit and the people who absorb the cost.

This is where the paradox stops being an economics curiosity and becomes a question of power. When the story is “AI is getting more efficient, so don’t worry,” the effect is to close down exactly the conversation we most need to have — about total demand, about who bears the load, about whether some uses are worth their physical cost at all. This is the deeper pattern in how technology gets captured: a technology arrives wrapped in the language of progress and inevitability, and the framing itself — “efficiency will save us” — does the work of making its costs feel like someone else’s problem, or no problem at all.

A tendency, not an iron law

I want to be honest about the limits of the argument, because the Jevons paradox is often wielded as if it were a law of physics that guarantees efficiency always backfires. It isn’t, and it doesn’t.

Economists distinguish between the direct rebound — you use more of the thing that got cheaper — and the total effect once everything settles out. Sometimes the rebound is partial: consumption rises, but not enough to wipe out the savings, so you still come out ahead. The paradox in its strong form, where efficiency causes total consumption to increase, tends to show up under particular conditions: when the resource is cheap enough that price is the main thing holding back demand, when there’s a vast reservoir of unmet uses waiting to be unlocked, and when nothing external is capping the total. AI in 2026 happens to sit squarely in all three of those conditions, which is why I think the paradox applies so forcefully here. But that’s an argument about this case, not a universal decree.

The genuinely useful lesson from Jevons is not fatalism. It’s that efficiency and restraint are different things, and only one of them actually reduces total consumption. If we want AI’s physical footprint to shrink rather than balloon, efficiency alone will not deliver it — the savings will keep getting spent on new uses. What reduces a total is a decision to cap it: cleaner and genuinely additional energy tied to the new load rather than displacing it, honest accounting of water in the places it’s actually drawn, siting choices that respect local scarcity, and a willingness to ask whether a given use is worth its physical cost at all. Those are choices. They don’t emerge automatically from a falling price per query — if anything, the falling price makes them harder to insist on.

The uncomfortable takeaway

What Jevons gives us is a warning against a very seductive form of complacency. Every time someone reassures you that AI is getting more efficient and therefore greener, remember the coal. The engines really did get more efficient. Britain really did burn more coal than ever. Both things were true at once, and it was the second one that shaped the century.

Efficiency is not restraint. Cheaper is not smaller. A more efficient AI is not a lighter one on the planet — it is very likely a heavier one, because efficiency is precisely what makes it worth deploying everywhere. Seeing that clearly doesn’t require pessimism about the technology. It requires refusing the story that the physical bill will quietly shrink on its own, and insisting instead on the boring, contestable, political work of deciding how much we build, where, powered by what, and paid for by whom. That decision is ours to make. The paradox only guarantees that if we don’t make it deliberately, the market will make it for us — in the direction of more.