Jevons Paradox is simple and counterintuitive. When you make a resource more efficient to use, total consumption of that resource often rises instead of falling. William Stanley Jevons spotted it in 1865 with coal and steam engines. Better engines burned less coal per unit of work, so factories and railways used far more of them. Coal demand soared.
The same pattern is showing up with AI.
Make inference cheaper and faster, and people do not stop at the old level of use. They invent new uses. Token prices have dropped hard. Efficiency gains from better models, specialized hardware, and techniques like those in DeepSeek-style systems have lowered the cost per query. Yet total compute, electricity, and token volume keep climbing. Data centers expand. Chip demand stays strong. Microsoft’s Satya Nadella put it bluntly after efficiency breakthroughs: as AI gets more efficient and accessible, use skyrockets and it becomes a commodity we cannot get enough of on an X post.
Recent discussions on Hacker News and in pieces examining the economics of AI (including analysis of the DeepSeek moment and rebound effects) keep circling this point. Large purpose-built data centers raise efficiency, which makes multi-agent workflows and always-on use practical. What was once a careful, occasional query becomes background processes, continuous monitoring, and parallel agents running for every user. The future is not one chat session. It is many agents doing many tasks at once.
Where it shows up#
- Compute and energy: Per-query energy cost has fallen dramatically, yet AI-related electricity demand is one of the fastest-growing loads on grids. Total consumption rises because the number of queries, models, and users grows faster than the efficiency gains.
- Tokens and software: Cheaper tokens turn previously uneconomic ideas into defaults. Embed AI in every customer interaction, every code review, every document draft. Usage multiplies.
- Labor and knowledge work: Efficiency can expand the total amount of certain kinds of work rather than shrink it. Lower cost of analysis or drafting can raise demand for the output, creating more overall activity even as individual tasks get automated.
Help or harm for human experience#
It can improve daily life. Cheaper intelligence makes powerful tools available to more people. Students, small businesses, researchers, and individuals in places with limited resources gain capabilities that used to require expensive teams or experts. Routine cognitive load drops. People spend less time on repetitive drafting, searching, or summarizing and more on judgment, creativity, or rest. New products and services appear that were not viable before.
It can also degrade the experience. Always-on agents and endless cheap generation risk flooding attention with noise. Energy and water demands of expanding data centers create real environmental pressure in some regions. Work that expands rather than contracts can leave people busier, not freer, if organizations simply raise output targets. Constant availability of generative tools can erode focus and the satisfaction of doing hard thinking yourself. When light became extremely efficient and cheap, cities filled with it until light pollution became a problem. Something similar can happen with cognitive abundance.
Jevons is not destiny. Demand is not infinitely elastic forever, and policy, pricing, and deliberate limits on wasteful use can shape the outcome. But right now the pattern is clear: making AI more efficient is not automatically making us use less of it. We are using more. The open question is whether that “more” ends up serving human purposes or just filling the available space.