Jevons is eating your AI budget
Every efficiency gain in inference has produced the same result since 1865: more consumption, not less. Why your AI bill grows as models get cheaper, and what to do about it.
In 1865, William Stanley Jevons noticed that more efficient steam engines did not reduce coal consumption. They increased it, because cheaper steam made coal worth burning in places it had never been worth burning before. Every CFO approving an AI budget in 2026 should have this paragraph framed on their wall.
Inference costs have fallen roughly an order of magnitude in eighteen months. Ask any finance team what happened to their total AI spend over the same period and you will hear the same answer: it went up. Not because anyone was careless, but because each price cut turned another hundred use cases from absurd to reasonable. The organization did not decide to use more AI. The price decided for it.
This is the Jevons paradox applied to tokens, and it breaks the most common budgeting assumption in the market right now: that efficiency gains will show up as savings. They will not. They will show up as usage. The only question is whether that usage creates value at the margin, and almost nobody is measuring the margin.
The measurement gap is where budgets quietly die. Most organizations can tell you their total inference spend and almost none can tell you their cost per successful outcome: per resolved ticket, per accepted draft, per correct extraction. When the denominator is missing, every efficiency gain looks like progress and every consumption increase looks like adoption.
The fix is not to use less AI. That is the Luddite reading of Jevons, and it is wrong. The fix is to route consumption deliberately: small models for easy tasks, big models for hard ones, decision models for classification, and rules for anything a lookup table handles. Routing is the only mechanism we have seen that lets organizations capture efficiency gains as savings instead of watching them evaporate into volume.
The broader market lesson is uncomfortable for vendors. If the Jevons dynamic holds, the winners are not whoever sells the cheapest inference. They are whoever helps customers decide what not to run. Judgment about consumption is becoming more valuable than the consumption itself. That sentence describes a very different industry than the one currently buying Super Bowl ads.
Keep Thinking
Independent analysis: no reselling, no vendor commissions. On the side, we help a small number of companies implement what we write about.
Work with the lab