Jevons Paradox

In 1865 the economist William Stanley Jevons noticed something counterintuitive about coal. Engineers had made the steam engine far more efficient — the same work now took far less coal. The expectation was that Britain would burn less coal. The opposite happened: cheaper steam power made it economical to put engines everywhere — factories, ships, mines, railways — and total coal consumption soared.

The efficiency gain didn’t shrink demand. By lowering the price of useful work, it unlocked demand that hadn’t existed at the old price. That is Jevons Paradox.

Strip away the coal and the paradox is just price elasticity of demand. When the price of something falls by X%, demand rises by some percentage. If demand rises by more than X%, total spending — and total quantity consumed — goes up. The good is “elastic.”

AI is, fundamentally, a machine for making cognitive tasks cheaper. Writing a first draft, summarising a document, generating code, drafting an email, transcribing a call — each gets dramatically cheaper per unit. The Jevons question for every one of them is the same: how elastic is demand for the outcome that task served?

This is why it’s a primitive and not a one-off article. The same lens reappears at every layer:

• Compute. Cheaper inference per token doesn’t reduce GPU demand — it makes new product categories (agents, always-on assistants, reasoning that runs for minutes) economical, so total compute demand climbs. See the cost-of-a-query model.
• Power. The literal, original version: more efficient datacenters lower the cost of compute, which pulls in more compute, which raises total grid draw. Efficiency is not a substitute for building power.
• Token pricing. Every order-of-magnitude price cut a lab ships expands the set of applications that pencil out — and usage grows faster than price fell.
• Jobs. The contested one, below.

The fear is mechanical: AI does the task, so the job disappears. That’s right only for a job that is a single isolated task — and only where AI is reliable at it. (When the task is self-contained and verifiable, jagged intelligence is good enough; when it’s precision-critical and hard to check, it isn’t — see where building with AI goes wrong.)

But most jobs aren’t a task — they’re a task in service of a broader outcome. A lawyer doesn’t produce contracts; they reduce a client’s risk. A developer doesn’t produce code; they ship working products. When the per-task cost collapses, the ROI of the outcome rises, and demand for the outcome expands. The work reshapes into one of three modes:

• Less time, same outcome — the task is now cheap; you finish sooner.
• Same time, more outcome — you do more of the valuable thing per hour.
• More time, far more outcome — the outcome is now so cheap to pursue that you do dramatically more of it (the Jevons case).

Which mode you get depends on the elasticity of the outcome. Inelastic outcomes (fixed demand) shed hours; elastic ones (latent, price-sensitive demand) expand the work.

When someone claims an efficiency gain will reduce total consumption — of compute, energy, or human labour — ask the Jevons questions in order:

1. What outcome does this task serve? Don’t reason about the task in isolation.
2. How elastic is demand for that outcome? Is there latent demand waiting at a lower price?
3. What was priced out before? Jevons effects come from use cases that weren’t economical at the old cost.

Two companion ideas sharpen the jobs case. The lump-of-labour fallacy explains why the total quantity of work isn’t fixed in the first place. Baumol’s cost disease explains where human labour concentrates — and gets more valuable — once AI absorbs the tasks it’s good at.