Jevons Paradox
The Jevons Paradox describes a surprising observation: when the use of a resource becomes cheaper, total consumption often doesn't fall but rises instead. In the AI debate, it serves as an argument for why more efficient models tend to increase rather than decrease demand for electricity and chips.
Normally one would expect: if a machine achieves the same output with less energy, overall energy consumption falls. The British economist William Stanley Jevons found the opposite in 1865. Steam engines needed less and less coal per hour of work. Yet England's coal consumption rose sharply. The reason: because operation became cheaper, steam engines suddenly became worthwhile in far more factories, mines, and ships. This rule is now known as the Jevons Paradox: efficiency gains make a technology so attractive that demand outpaces the savings.
Why efficiency alone doesn't save resources
The paradox contradicts an assumption that sounds self-evident in many debates. Politicians and companies often promise to solve a problem through better technology. More economical engines are supposed to reduce oil consumption, more efficient data centers to lower electricity demand. Jevons shows that this only holds if demand stays roughly constant. But for technologies that are nowhere near universally adopted yet, demand does not stay constant.
It is important here to distinguish this from a related concept. The so-called rebound effect describes, in general, that part of the savings is eaten up again by increased use. Someone who buys a fuel-efficient car and therefore drives more often produces a rebound. One only speaks of the Jevons Paradox once the effect exceeds 100 percent. In that case, consumption ends up higher than before, not merely reduced by a smaller margin.
For investors and companies, this has concrete implications. When a technology becomes more efficient, its suppliers' revenue does not automatically decline. It can even grow, because the overall market expands. This is precisely why the term suddenly appeared in quarterly reports and analyst notes in 2025.
The mechanism of price and demand
Behind the paradox lies no magic trick, but a chain of three steps. First, efficiency lowers the cost per unit of performance. Second, the lower price makes applications possible that would not have been worthwhile before. Third, this causes the total volume of usage to rise. Whether consumption ultimately falls or rises depends on how strongly demand reacts to price.
A numerical example makes this tangible. Suppose a query to an AI system costs ten cents, and there are one million queries per day. Now the technology becomes ten times more efficient, and a query costs one cent. If the volume stays the same, costs and electricity consumption fall to a tenth. But if the low price leads to fifty million queries being made daily, total consumption is five times higher than before.
The effect does not occur everywhere. For technologies that everyone already owns and uses intensively, there is little room for additional demand. More efficient refrigerators have indeed reduced electricity consumption, because hardly anyone buys a second refrigerator as a result. The paradox takes hold primarily where a technology is still in its early stages and many potential uses remain untapped.
The argument in the AI and data center debate
In the news today, this term almost always comes up in connection with AI. When the Chinese company DeepSeek unveiled a very efficiently trained language model in early 2025, chipmaker stocks initially plunged. The worry was: if AI becomes this much cheaper, the world will need fewer computing chips. Executives at major technology companies, including Microsoft and Nvidia, explicitly pointed to Jevons in response. Cheaper AI would mean more AI, not less.
This can also be observed outside the stock market. Earlier language models were too expensive to integrate into search engines, word processors, or customer service hotlines. As the cost per response fell, they migrated into exactly these products. The consumption per individual query is significantly lower today than three years ago. At the same time, the number of queries has multiplied many times over.
Still, the paradox should not be treated as a law of nature. It is an observation about markets, not a formula with a guaranteed outcome. Eventually every sensible use case is occupied, and demand grows more slowly than efficiency. Where that point lies for AI, nobody currently knows. Anyone who encounters this argument in a news report should therefore check whether it explains something or merely serves to soothe an uncomfortable stock decline.