Jevons Paradox

Jevons noticed that more efficient steam engines didn’t save Britain’s coal. They made steam power cheap enough to use in more factories, mines and railways, so total coal burned went up. The same pattern shows up in everyday life: when a city adds lanes to a highway, driving gets easier and more people drive.

For AI, the resource is usually compute for inference, or the energy that runs it. Suppose the price of running a model drops tenfold. A company that used AI for one support workflow might now use it for ten, plus coding, search and document review. Its AI bill can go up even though each request got cheaper. That is the paradox in business terms, and it is one reason more efficient chips and smaller models don’t automatically shrink data-center demand.

It is a tendency, not a law. Total use rises only when demand responds strongly to price, and nobody can say in advance how much AI demand will grow as costs fall. Editorially, it is a useful check on claims that efficiency alone will solve AI’s energy or cost problems.

Guests invoke it directly. In episode 56, at 28:34, AMD’s Anush Elangovan responds to a question about layoffs and rising AI throughput expectations by saying he is “a big believer” in Jevons paradox: he expects much more demand for GPU compute and for people doing things with it, not less. That is his view as a hardware executive. The episode was sponsored by Galileo; see the sponsor read. In episode 78, at 1:31, Svix CEO Tom Hacohen uses the same idea for software integrations: as connecting systems got easier with REST and JSON, the number of integrations exploded. Episode 78 is sponsored by Svix, the Season 4 presenting sponsor; Conor thanks Svix at 1:56.

For the environmental side of the argument, see What AI Actually Costs the Environment.