Jevons Paradox in AI Inference

When the cost of doing something falls, people do more of it — sometimes so much more that total resource use rises rather than falls. That is the core of the Jevons Paradox, and it is increasingly invoked to explain why cheaper AI is unlikely to mean less AI energy. ## The classical paradox In 1865, William Stanley Jevons argued in The Coal Question that improvements in steam engine efficiency had not reduced Britain's coal consumption — they had multiplied it. More efficient engines lowered the effective price of mechanical work, which expanded the range of profitable applications: railways, ironworks, textile mills, shipping. Each task burned less coal per unit of output, but there were now vastly more units of output. The aggregate went up. Modern energy economists formalized this intuition as the Rebound Effect (Energy Economics), with the strongest version — the Khazzoom-Brookes postulate — claiming that economy-wide efficiency gains generally increase total energy demand at long horizons. ## Mapping onto AI inference AI inference is unusually well suited to this dynamic. Per-token costs have fallen rapidly as providers compete and as architectural advances — model distillation, Mixture of Experts routing, quantization, speculative decoding, better KV-cache reuse — compound on top of hardware gains. Public price data shows order-of-magnitude declines for frontier-class models over two to three years, and even sharper drops for smaller task-specific models. Cheaper tokens do not merely substitute for the prior workload. They unlock workloads that previously failed a cost-benefit test: chat interfaces replacing keyword search; AI agents that loop over a problem dozens of times, calling tools and re-prompting themselves; classifiers running on every email or log line; always-on assistants embedded in IDEs and browsers. Each new use case is individually cheap. Collectively, they push aggregate inference compute — and the electricity and GPU capacity behind it — upward, even as per-query footprint shrinks. ## Counter-arguments and limits The Jevons framing is not a law of nature. Three caveats matter. First, the rebound effect is bounded by demand saturation: at some point users do not want another agent loop, just as households eventually stop adding light bulbs. Second, much AI energy growth is driven by training and frontier-model scaling rather than inference rebound, so attribution is messy. Third, efficiency gains can shift load onto cleaner grids or off-peak windows, decoupling compute growth from emissions growth even if kilowatt-hours rise. What the paradox does usefully predict is direction: absent a binding constraint — regulation, grid capacity, chip supply — efficiency improvements in inference should be expected to expand the total AI footprint rather than contract it. Per-query metrics flatter the picture; system-level metrics are the honest ones.