Jevons Paradox

In 1865, William Stanley Jevons noticed that James Watt's more efficient steam engine had not reduced Britain's coal use. It had multiplied it. The engine made coal cheaper per unit of work, so factories, mills, and railways bought more work. Total coal burned went up, not down. Jevons's claim, in The Coal Question: efficiency is not conservation. It is a price cut on the service the resource provides, and price cuts grow markets.

How it works

The mechanism is three layers stacked.

Direct rebound. A more efficient car costs less per mile. Drivers respond by driving more miles. Studies of US passenger vehicles put this rebound at roughly 10-30% of the engineered fuel saving (to verify: Greene & Sorrell rebound surveys, ~2014-2020).
Indirect rebound. Money saved on fuel gets spent on flights, appliances, larger homes — each with its own energy footprint.
Economy-wide rebound (the Jevons case proper). When an efficiency gain unlocks a new use — making the resource viable for applications previously priced out — demand can expand faster than efficiency improves. Coal in 1865. Compute in 2026.

The strong Jevons claim is layer 3: total consumption rises in absolute terms. The weaker rebound claim is that efficiency gains are partially clawed back. Almost no economist disputes the weak version. The strong version is fought over.

Where it shows up

Domain	Efficiency gain	What happened to total use
Lighting	LEDs ~90% more efficient than incandescent (lumens/watt)	Global lighting energy consumption broadly flat-to-rising; lit surface area on Earth grew ~2% per year, 2012-2016 (Kyba et al., Science Advances, 2017)
Passenger vehicles	US new-car fuel economy roughly doubled 1975-2020	Total US vehicle miles traveled roughly tripled in the same window (FHWA data, to verify)
Air travel	Jet fuel per seat-km down ~40% since 1990 (IATA, to verify)	Global passenger-kilometres roughly quadrupled 1990-2019
AI inference	Cost per token down ~10-100x in 2023-2025 across frontier models	Total inference compute up by a larger multiple; data center electricity demand projected to double by 2030 (IEA, 2024)
Cryptocurrency mining	Joules-per-hash for Bitcoin ASICs down ~1000x since 2013	Bitcoin network energy use rose to ~150 TWh/year by 2024 (Cambridge CBECI)

The pattern is not universal. Domestic refrigeration in the OECD looks like a genuine efficiency win: fridges got ~75% more efficient since 1980, and household fridge electricity dropped, because the service (one fridge per home) saturated. The lesson: rebound is shaped by whether the underlying demand is bounded (one fridge per household) or unbounded (more miles, more compute, more lumens of advertising).

What's contested

The climate-policy stakes are real. If economy-wide rebound is small (~20-30%), efficiency standards work roughly as planned and the gap between projected and actual emissions cuts is manageable. If rebound averages 50%+ across sectors, half of every efficiency mandate is fictional accounting. If it exceeds 100% in some sectors (backfire — pure Jevons), efficiency policy in those sectors increases emissions.

The empirical literature is messy. Sorrell's 2007 UKERC review put economy-wide rebound between 10% and 60% with wide uncertainty. Saunders, Brookes, and Khazzoom argue for higher numbers and historical backfire. The IEA and most environmental ministries assume low rebound in their models, which critics call motivated reasoning. The honest position in 2026: nobody has a clean empirical handle on long-run general-equilibrium rebound, and the answer probably differs by sector and by how mature the technology is.

The deeper unknown is whether rebound saturates. Lighting can only get so bright before further lumens have no buyer. Compute may have no such ceiling, because every efficiency gain unlocks new applications (real-time agents, video generation, scientific simulation) that did not previously exist as demand.

Why this has to do with other realms

Jevons is the economic shadow of concept second law thermodynamics. The second law says useful energy degrades; Jevons says cheap useful energy gets used more. Together they explain why "decoupling growth from emissions" is harder than the absolute-decoupling charts suggest — efficiency expands the size of the economic engine even as it lowers the per-unit cost of running it.

The pattern also recurs in concept induced demand from transport planning: build a wider highway, get more traffic, not less congestion. Same shape, different resource.

An open question

If AI inference costs fall another 100x by 2030, what new categories of demand appear that today don't exist as markets — and does the resulting electricity load reshape grid planning faster than renewables can be built?

Key sources

The Coal Question by William Stanley Jevons (1865) — the original argument, still readable, still uncomfortable.
The Rebound Effect: an assessment of the evidence — Steve Sorrell, UKERC (2007) — the canonical empirical review.
Kyba et al., "Artificially lit surface of Earth at night increasing in radiance and extent," Science Advances (2017) — the lighting backfire datum.
IEA, Electricity 2024 — projections for data center load growth.
to verify: Saunders, "The Khazzoom-Brookes Postulate and Neoclassical Growth" (~1992) — the strong-rebound theoretical case.

Abhishek's take

I see the same shape when a costing sheet or fit note gets faster. A vendor sample that once forced one decision now produces three alternate drops, because the cheaper decision loop makes the range larger. Efficiency does not reduce judgment on the floor; it moves the bottleneck to restraint.

Abhishek