Back in 1865, Jevons noticed a fascinating trend where highly efficient steam engines ended up consuming greater amounts of coal instead of conserving it. We are witnessing this exact same pattern unfold today with artificial intelligence, as every leap in efficiency results in higher overall usage rather than a reduction. For example, when token prices dropped, inference budgets did not actually shrink. Lower costs simply encouraged development teams to take on more ambitious projects. They began utilizing significantly longer contexts and implementing a higher number of retries. We now routinely see automated agents querying a model thirty times to accomplish something a person would have asked about just once. Ultimately, if you are mapping out capacity requirements with the expectation that better efficiency will lead to decreased demand, you have the fundamental direction of this relationship entirely reversed.