top of page

Why Cheaper AI Could Mean More Energy Not Less

Writer: Dokyun Kim
Dokyun Kim
6d
3 min read

In 1865, the English economist William Stanley Jevons noticed something that unsettled the optimists of his day. James Watt's steam engine had made coal far more efficient to use, and many assumed Britain's coal consumption would fall as a result. The opposite happened: coal use soared. By making each unit of work cheaper, efficiency opened up uses for steam power that had never been worth pursuing, and total demand overwhelmed the per-unit savings. Economists now call this the Jevons paradox, and its milder cousin the rebound effect. More than 150 years later, it may be the most important idea missing from the debate over artificial intelligence and energy.


The rebound effect comes in degrees. A direct rebound occurs when a more fuel-efficient car makes driving cheaper, so people drive more and claw back part of the savings. An indirect rebound occurs when the money saved on fuel is spent on something else that also consumes energy, such as a flight abroad. At the economy-wide level, efficiency lowers costs across whole industries, spurring growth and new products. When the rebound exceeds 100 percent, economists call it "backfire": the efficiency gain increases total consumption. Lighting is the textbook case. Over two centuries, the cost of artificial light fell by orders of magnitude, from candles to gas lamps to LEDs, and humanity responded not by using less energy on light but by lighting everything, everywhere, all night.


AI is now running the same experiment at high speed. In early 2025, the Chinese lab DeepSeek released a model that appeared to match leading systems at a fraction of the training cost, and energy stocks briefly fell on the hope that AI would need less power. Microsoft's CEO Satya Nadella responded by citing Jevons directly: as AI gets cheaper, he argued, its use would skyrocket. He was likely right. Each generation of chips performs more computation per watt, and the price of running a query has collapsed, yet data center electricity demand keeps climbing. The International Energy Agency has projected that global data center electricity use could roughly double by 2030. Cheaper intelligence, like cheaper light, is being consumed in quantities nobody planned for.


This matters because so much climate policy rests on efficiency. Appliance standards, building codes, and fuel economy rules are popular precisely because they promise savings without sacrifice. They often deliver real benefits, and most studies find rebounds for household energy are partial rather than total. But efficiency alone cannot guarantee lower emissions, because it changes the price of energy services without changing the price of the damage. If cheaper computation is powered by gas plants, a more efficient chip can still mean a hotter planet. The lesson of Jevons is not that efficiency is futile, but that it is incomplete.


The policy answer is to pair efficiency with something that holds the total in check. A carbon price captures the rebound by making the extra consumption pay for its emissions. Clean-power requirements for new data centers, or grid rules that make large new loads bring their own clean supply, can ensure that growth runs on low-carbon electricity rather than whatever is cheapest. Efficiency makes the economy more productive; only limits and prices make it cleaner. As AI becomes cheaper every year, the question for policymakers is not how efficient each chip can become, but what all those chips, together, will be allowed to burn.


Comments


bottom of page