The 1860s Coal Paradox That Explains Why More Efficient AI Means More Energy Usage, Not Less

You would think that making a technology dramatically more efficient would mean we use less of the resource behind it. A British economist showed the opposite in 1865, using coal. The same logic is now playing out with artificial intelligence — and it explains why every price cut and efficiency breakthrough seems to leave us using more computing power, not less.

What’s actually true: When a resource becomes cheaper or more efficient to use, total consumption of it often rises rather than falls, because the lower cost opens up uses that were never worth it before. This is Jevons paradox, and it has held up across coal, electricity, fuel economy, and now the computing power behind AI.

What’s misleading: The common assumption — that a more efficient model or a cheaper chip shrinks overall demand for energy and hardware — usually gets the direction backward. Efficiency is often the very thing that drives the demand curve upward.

In 1865, William Stanley Jevons published The Coal Question, a book worried that Britain’s industrial dominance rested on a finite pile of coal. The conventional wisdom of the day was comforting: engineers were making steam engines far more efficient, so surely Britain would burn through its reserves more slowly. James Watt’s improvements had already cut the coal needed for a given amount of power by a wide margin. Better engines, the thinking went, meant less coal.

Jevons argued they had it exactly backward. Cheaper, more efficient engines made steam power economical for industries and uses that could not previously justify it. Coal-powered machinery spread into more factories, more mines, more railways, more corners of the economy. Each engine sipped less coal per unit of work — but there were now vastly more engines doing vastly more work. National coal consumption did not fall. It climbed steeply.

That is the heart of the paradox. Efficiency lowers the effective price of a service. A lower price expands demand. If demand expands far enough, total resource use rises even as the resource is used more sparingly for any single task. Economists later gave the broader version a name, the “rebound effect,” with Jevons paradox describing its most extreme form — where the rebound more than erases the savings.

The mechanism is not magic, and it is not universal. It depends on how sensitive demand is to price. For a service people already use as much as they want, making it cheaper changes little. But for a service with enormous unmet demand, where high cost is the main thing holding people back, a price drop can crack the whole thing open. Coal-powered industry in the 1860s was exactly that kind of service. So is artificial intelligence today.

The cost of running AI models has been falling at a pace that is hard to hold in your head. The price of getting a given quality of answer out of a model has dropped sharply in just a few years, driven by better chips, more efficient architectures, cheaper inference methods, and fierce competition between providers. On a per-query basis, AI keeps getting cheaper and more efficient.

If the naive intuition held, that would mean the industry needs less computing power and less energy over time. The opposite is happening. As each query gets cheaper, entirely new uses become viable. Things nobody would have paid a dollar for become worth a fraction of a cent — so people do them constantly. Summarizing every email. Drafting every first pass. Generating code, images, and transcripts on demand. Embedding a model into every app, every search, every customer-service chat. Running agents that make thousands of calls to finish a single task a person used to do by hand, if at all.

Cheaper AI did not shrink demand for compute. It multiplied it. The International Energy Agency has documented the striking version of this: the energy used per AI task is improving at a rate it calls unprecedented in energy history, and yet total electricity demand from data centres is set to keep climbing, not falling — precisely because the cheaper each task gets, the more of them the world runs. The efficiency is real. So is the surge in total use that the efficiency created. That is Jevons, updated for silicon.

The comparison to coal is close enough to teach and different enough to watch. In both cases a general-purpose technology got cheaper, and the price drop pulled in a long tail of uses sitting just out of reach. In both cases the people making the efficiency gains were often the same ones promising those gains would ease the strain on the resource. And in both cases the resource was not only the obvious one. For coal it was ultimately the mines and the air. For AI it is electricity, water for cooling, and the manufacturing capacity for advanced chips.

Here the careful framing matters, because Jevons paradox gets dragged into arguments it does not settle. The economists’ version is precise: the rebound effect is well documented, but its size varies enormously by sector, and a true “backfire,” where efficiency actually raises total use, is real but not guaranteed everywhere. Whether AI’s efficiency gains produce a full backfire in energy terms is still being measured, and serious analysts disagree about how much of the projected data-centre demand is Jevons-style rebound versus a one-time buildout that eventually plateaus.

The industry tends to lean on the paradox as reassurance — do not worry that efficiency is pointless, because the extra use creates extra value. There is something to that. More useful AI, more widely available, is a real benefit, not just wasted computation. But “efficiency created more value” and “efficiency reduced total resource use” are different claims, and the second one keeps turning out to be false.

Then there is the popular version, the one in headlines and feeds: that a newer, more efficient model is automatically greener, or that a cheaper chip means a lighter footprint. That is the framing worth the most skepticism. A more efficient model can sit right alongside a much larger total energy bill, because the efficiency is exactly what makes the larger scale possible. To their credit, some engineers and analysts say this plainly — the per-task number and the total number can move in opposite directions at once, and marketing tends to quote whichever one flatters the story.

Efficiency is not the same as reduction. When a technology becomes a general-purpose tool that many people want more of, making it cheaper tends to increase how much of it we use in total — sometimes far more than the efficiency saved. Jevons saw it in coal in 1865. The pattern repeated through electric lighting, engines, and bandwidth. AI is the current, and possibly the largest, example.

What is not settled is where the AI curve lands. Will demand for AI-driven services eventually saturate the way older technologies did, letting efficiency finally bite into total consumption? Or is the tail of possible uses so long that we are nowhere near the ceiling? The answer decides whether today’s data-centre buildout is a spike or a new baseline, and it turns on things we cannot yet measure: how good the models get, how many tasks they credibly absorb, and whether energy and chip supply become the binding constraint before demand does.

The one thing the paradox tells you with confidence is which mistake to avoid. If someone points to a more efficient AI model and concludes that we will therefore use less computing power and less energy, treat it the way Jevons treated the coal optimists of his day. The efficiency is probably real. The reduction usually is not.

1. Jevons, W.S. The Coal Question; An Inquiry Concerning the Progress of the Nation, and the Probable Exhaustion of Our Coal-Mines. Macmillan and Co., 1865. econlib.org
2. International Energy Agency. Energy and AI. IEA, Paris, 2025. iea.org
3. International Energy Agency. Key Questions on Energy and AI. IEA, Paris, 2026. iea.org

Disclaimer: This content includes personal opinions and interpretations based on available sources and should not replace medical advice. This content includes interpretation of available research and should not replace medical advice. Although the data found in this blog and infographic has been produced and processed from sources believed to be reliable, no warranty expressed or implied can be made regarding the accuracy, completeness, legality or reliability of any such information. This disclaimer applies to any uses of the information whether isolated or aggregate uses thereof.