The Jevons Paradox for AI: Why "It's Too Late" Is Exactly Backwards

The short version of the Jevons paradox is a story. Here's the data that turns it into a fact you can act on. From late 2022 to late 2024, the price to run a GPT-3.5-class model fell from about $20 per million tokens to roughly $0.07 — a ~280x drop in two years. Epoch AI, tracking the cheapest model that matches a given capability across six benchmarks, puts the median decline at ~50x per year, accelerating to ~200x per year after January 2024. If demand were fixed, total AI spend should have collapsed. It did the opposite.

• Price per unit of intelligence: down ~300x in roughly two years.
• Token volume over the same window: up thousands of percent (numbers below).
• Net result: total AI revenue and spend went up, not down. That gap — price down, total bill up — IS the Jevons paradox, measured.

The price drop is only half of it. The other half is how fast usage climbed to fill the new headroom. As the per-token cost cratered, the number of tokens people actually ran went vertical. These aren't projections — they're reported run-rates from the companies serving the traffic.

Jevons isn't magic and it isn't guaranteed. It fires when demand is elastic — when there's a large backlog of things people would do if only the price dropped. Cheap coal in 1865 met a backlog of factories, railways and ships that weren't worth building at the old price. Cheap tokens in 2026 meet a backlog of dashboards, internal tools, support bots, data cleanups and one-off scripts that were never worth a developer's week. When the price crosses the line where those become worth doing, they all get built at once. That's the surge.

1. Efficiency cuts the cost per use (a better engine; a cheaper model).
2. The lower cost crosses a threshold — uses that were 'not worth it' flip to 'worth it'.
3. Because the backlog of those uses is huge, the new volume swamps the per-unit saving.
4. Total consumption — and total spend — rises. The market widens instead of clearing.

You don't have to take the theory on faith; the market ran the experiment live. When DeepSeek shipped a frontier-ish model at a fraction of the training cost, US AI stocks dropped on the fear that cheaper AI meant a smaller pie. Within a day, Microsoft CEO Satya Nadella posted "Jevons paradox strikes again!" and argued the opposite: "As AI gets more efficient and accessible, we will see its use skyrocket, turning it into a commodity we just can't get enough of." He was reading the same curve you just saw. Cheaper input, bigger total market.

• The reflex (markets, Jan 2025): cheaper AI = less revenue, sell.
• The Jevons read (Nadella, same week): cheaper AI = use skyrockets = bigger pie.
• Eighteen months of token-volume data since then sided with the Jevons read.

On June 9, 2026, Andrej Karpathy — now at Anthropic — named the same effect for software directly. Reacting to a major model release, he wrote that "working software increasingly comes out on a tap," so "the Jevon's paradox kicks in" and his own "demand for software [is] growing substantially." Sit with who's saying it. The man can write any tool himself, and his appetite for software is going up, not down. Cheap creation didn't satisfy his demand. It exposed how much he'd never bothered building because it used to cost too much. That backlog — yours, his, every business's — is the elastic demand the paradox needs.

• When code is cheap, the bottleneck moves from 'can we build it' to 'what's worth building'.
• Every bespoke single-use app that wasn't worth a dev-week is suddenly worth an afternoon.
• Real-world echo: per SaaStr, Claude Code went from its May 2025 public launch to a run-rate above ~$2.5B by early 2026 — that's the unbuilt backlog turning into spend.

Read the curve honestly: total demand grows, but it does not get handed out evenly. When coal got cheap, the winners weren't people who admired steam engines — they were people who built mills, railways and ships on cheap power. Same shape now. Cheap tokens reward whoever turns them into something a specific buyer will pay for, not whoever merely has access. Everyone has access. Access stopped being the edge the day the price fell 300x.

• Saturated at the input layer (tokens are a commodity, racing to ~free).
• Wide open at the application layer (the backlog of 'now-worth-doing' jobs is enormous).
• The scarce thing is no longer the model. It's a clean offer pointed at one buyer's problem.

The paradox rewards motion, so here's a concrete way to use it this week instead of nodding and closing the tab. Pick one narrow job that just crossed the worth-doing line.

1. Name a 'now-worth-doing' job. Find a task a specific business does by hand because a developer was always too expensive for it. Example: a dental clinic that re-types every new patient's intake form into their practice software.
2. Price the old vs new cost. Old: a custom integration nobody would quote under a few thousand dollars. New: an AI workflow that reads the form and writes the record for cents per run. That price collapse is your opening — it's Jevons at the level of one invoice.
3. Build the thin layer between the model and the buyer. Not the model — the workflow, the packaging, the 'it just works' setup and support. That layer is what the buyer pays for; the tokens underneath are nearly free.
4. Ship the small version now. One clinic, one workflow, one monthly fee. 'I'll wait until it settles' is how you opt out of a market that is still widening — the volume tables above are the cost of waiting.

When something genuinely useful gets cheaper, people don't buy less of it — they find a hundred new reasons to buy more. AI is the cheapest it has ever been and the best it has ever been at the same time, and the receipts show total spend rising right through the price collapse. 'Saturated' and 'too late' describe a shrinking market. This one is measurably doing the opposite.

• Cheaper + better → more total demand and more total spend (because demand is elastic).
• The builder/reseller market is widening, not closing — the token-volume curves prove it.
• The risk was never being late. It's mistaking 'everyone has access' for 'everyone has an offer.'