A Victorian economist's inconvenient finding
William Stanley Jevons was born in Liverpool in 1835, the son of an iron merchant. Financial difficulties cut short his studies at University College London in 1854, forcing him to take a post as an assayer at the British colonial mint in Sydney, Australia. It was there, assessing metals far from England's industrial heartland, that he grew fascinated by political economy.
He eventually returned, earned his master's degree in 1863, and climbed the academic ladder through Owens College Manchester and Queen's College Liverpool before landing a professorship at University College in 1876. He was elected a Fellow of the Royal Society in 1872. Irving Fisher later described Jevons's 1871 book The Theory of Political Economy as the start of the mathematical method in economics. Jevons was also the first economist to construct index numbers, and his work on marginal utility, published the same year that Karl Menger published similar findings in Vienna, helped shift economic theory from classical to neoclassical foundations.
But the work that concerns us here came six years earlier. In 1865, Jevons published The Coal Question, considered the earliest work to extensively address energy resource depletion in an industrialized world. His observation, tucked inside Chapter 7, "Of the Economy of Fuel," was simple and deeply uncomfortable: more efficient steam engines had not reduced Britain's coal consumption. They had increased it.
The mechanism is straightforward. When you make a resource cheaper to use per unit of output, you lower the cost of every activity that depends on it. Lower costs expand demand. New uses appear that were previously uneconomical. The total consumption of the resource rises, often well past the original baseline.
Jevons argued, in his own words, that "contrary to common intuition, technological progress could not be relied upon to reduce fuel consumption." British coal consumption went on to triple by 1900. Jevons had no real answer to the problem he raised. Britain could burn through its coal quickly or slowly. That was roughly the range of options available.
He also noted the undesirable side effects of coal burning, particularly air pollution, while acknowledging that curtailment was unlikely without deep societal reform. He died in 1882, near Hastings, before seeing how thoroughly his prediction held.
The paradox has surfaced consistently across different technologies and time periods. Each example follows the same structure: efficiency lowers per-unit cost, consumption expands to fill and then exceed the original footprint.
The average UK resident in 2000 consumed 75 times more artificial light than their ancestor did in 1900, and roughly 6,000 times more than in 1800. LED lighting is cheaper to run, so people install more of it and run it longer. The reduced cost per lumen drove up total lumen-hours consumed.
The same pattern appears with cars. More fuel-efficient vehicles make each mile cheaper to drive, so people drive more miles, trade up to larger vehicles, or add a second car to the household. Research from Nissan found that electric vehicle drivers travel over 370 miles more per year than petrol or diesel car users.
In Germany during the 1980s, building insulation programs reduced heating costs per square foot. Within a few years, occupants responded by raising their comfort standards, heating larger spaces, and erasing most of the efficiency gains. The efficiency improvements created room for new demand, and demand filled it.
When DeepSeek released its R1 model in late January 2025, claiming performance comparable to OpenAI's frontier models at a fraction of the training cost, Nvidia lost $600 billion in market value in a single day. Markets read cheap AI as deflationary for the chip industry.
Microsoft CEO Satya Nadella read it differently. He wrote on LinkedIn that same day: "Jevons paradox strikes again! As AI gets more efficient and accessible, we will see its use skyrocket, turning it into a commodity we just can't get enough of."
UBS analysts made the comparison concrete, noting that DeepSeek's 97% cost reduction in AI training mirrors the internal combustion engine's roughly 100x efficiency leap, which was followed by a 223x growth in global oil demand after 1900.
Google announced in 2025 a 33-fold reduction in energy consumption per Gemini query. Each text prompt now uses approximately 0.24 watt-hours. That is genuine technical progress. But as AI inference gets cheaper, total query volume grows. Global data centers consumed approximately 415 terawatt-hours in 2024, about 1.5% of worldwide electricity. According to a 2024 International Energy Agency report, data centers handling AI training and computation were already at 2% of global electricity consumption and set to double by 2026, surpassing all of Canada's power usage.
The emissions numbers tell the same story. Google reports a 48% rise in emissions since 2019. Microsoft reports a 21% increase since 2020. Both companies are drawing significantly more water: Google up 20%, Microsoft up 34%, during a period when roughly half the world's population is moving toward water scarcity.
A paper published at the 2025 ACM Conference on Fairness, Accountability, and Transparency examined exactly this dynamic, finding that efficiency gains in AI are systematically reinvested to expand markets and stimulate new demand rather than reduce total consumption. A July 2025 report in Nature Cities noted that generative AI platforms concentrated in major urban tech hubs now consume approximately 600 megawatt-hours of electricity daily, equivalent to over 50 times the annual energy use of an average US household.
Stanford economist Erik Brynjolfsson has pointed out that a Jevons-style rebound is plausible in AI-affected occupations. If AI makes coders or radiologists more productive, that productivity lowers the effective cost of their output, which can expand demand for that output enough to increase total employment in those fields rather than reduce it. The efficiency gain doesn't shrink the market. It grows it.
That dynamic is still playing out. But the energy version is already measurable, and Jevons would have recognized it immediately. He spent his career arguing that efficiency, without demand restraint, does not conserve resources. It accelerates their consumption. That argument is 160 years old and still running ahead of the engineering curve.