# Jevons Paradox The Jevons paradox says that making a resource more efficient to use can *increase* how much of it we consume in total. Efficiency lowers the effective price, a lower price opens up uses that weren't worth it before, and if demand grows fast enough, total consumption goes up. It's counterintuitive because we tend to reason one use at a time ("this engine burns half the coal, so we'll burn half the coal"), while the economy reasons across all the uses that just became affordable. ## Where it comes from William Stanley Jevons (1835-1882) was an English economist and logician. In 1865, in *The Coal Question*, he looked at Britain's coal and noticed something odd: James Watt's much more efficient steam engine hadn't reduced coal consumption. It had exploded it, because cheaper effective power made steam worth using in far more places. His conclusion is the famous line: > "It is wholly a confusion of ideas to suppose that the economical use of fuel is equivalent to a diminished consumption. The very contrary is the truth." A fun detail: Jevons also built the "logic piano" in 1869, a mechanical machine that solved simple logic problems. So a model that makes cheap, typed logical decisions being named after him is more fitting than it first looks. ## The mechanism (and when it doesn't happen) The paradox isn't a law. It depends on how demand reacts to price: - **Price elasticity of demand.** If a 10% price drop increases usage by less than 10%, efficiency still saves resources overall (a *partial rebound*). If usage grows by more than 10% (elasticity above 1), total consumption rises. That's the full paradox, which economists call *backfire* - **The rebound effect** is the general name for the part of efficiency gains that gets eaten by extra usage. In the 1980s, economists Daniel Khazzoom and Leonard Brookes formalized the energy version (the Khazzoom-Brookes postulate): more energy efficiency tends to raise overall energy demand - **Everyday example:** more fuel-efficient cars make each kilometer cheaper, so people drive more So the useful question is always "how much hidden demand is waiting at a lower price?" Where demand is saturated, efficiency saves. Where it's huge and latent, efficiency backfires. ## Jevons and AI The paradox became an AI talking point in late January 2025, right after DeepSeek-R1 showed near-frontier reasoning at a fraction of the cost. Satya Nadella posted "Jevons paradox strikes again!", arguing that cheaper AI would push usage up, not down. Erik Brynjolfsson has made a similar argument about jobs: occupations like radiologists, translators or coders might grow if AI makes their output cheap enough. [[Andrej Karpathy]] invoked it when [[Claude Fable 5]] came out: as working software "comes out on a tap", his own demand for software grows (explainers, dashboards, single-use apps). [[TypeSafe AI]] made it the name of their product. [[Diogo Almeida]] on Latent Space: "Jev is optimized for intelligence per dollar, hence the name Jev. Jevons Paradox." The pricing is designed for it: $0.042 per million input tokens, and output tokens are free. And the docs actively push you toward more usage: - "Asking a question you might not need is close to free." That's the whole idea behind [[Speculative Fan-Out]]: put every question you *might* need in one request (13 questions batched came out 12.2x cheaper than 13 separate calls in their cookbook) - A prose linter asking ten yes/no questions per paragraph, or reranking every BM25 hit with a model, only makes sense when one judgment costs a fraction of a cent - TypeSafe says it processes more than 1 trillion tokens a day (self-reported, a week after launch) That's Jevons in action: the unit price of a judgment collapses, so we ask for many more judgments, and total spend on judgment can go up. ## Where the analogy wobbles - **Tokens aren't coal.** The Register questioned whether AI tokens follow the same demand elasticity as energy markets. The paradox needs a lot of latent demand; for many business tasks, you only need so many classifications - **Frontier prices are going the other way.** Per [[AI API prices are rising]], frontier models got *more* expensive through 2026 while everything below the frontier got cheaper. Jevons applies to the cheap tier, where demand is still exploding - **It's easy to use as a slogan.** "Jevons paradox" gets invoked to justify any price cut. The claim only holds if you can point at the new uses the lower price enables ## How I use it When a tool makes something much cheaper (AI inference, storage, a note-taking workflow), I don't ask "how much will I save?" I ask "what becomes worth doing now that wasn't before?" That's usually where the value is, and it's also where the budget surprise comes from. If you move your classification calls to a model that's 400x cheaper, expect to make far more of them, and plan your costs (and rate limits) for the new volume, not the old one. ## References - [Jevons paradox (Wikipedia)](https://en.wikipedia.org/wiki/Jevons_paradox) - [The Coal Question (Wikipedia)](https://en.wikipedia.org/wiki/The_Coal_Question) - [Jev: System One models for Prod, not God, with Diogo Almeida (Latent Space)](https://www.latent.space/p/jev) - [Introducing System One Models & Jev (TypeSafe AI)](https://typesafe.ai/blog/introducing-system-one-models-and-jev) - [TypeSafe AI debuts model for machines that plays Doom (The Register)](https://www.theregister.com/ai-and-ml/2026/09/16/typesafe-ai-debuts-model-for-machines-that-plays-doom/5296711) - [Parallel questions cookbook (TypeSafe docs)](https://docs.typesafe.ai/cookbooks/parallel_questions) ## Related - [[Jev]] - [[TypeSafe AI]] - [[Machine Native Intelligence]] - [[Speculative Fan-Out]] - [[AI API prices are rising]] - [[AI Cost Management]] - [[AI Inference]]