Jevons Paradox

In 1865 the economist William Stanley Jevons published The Coal Question, arguing Britain would burn through its coal faster, not slower, as steam engines got more efficient. Engineers at the time treated efficiency as conservation: a better engine burns less coal for the same work. Jevons saw the opposite at the system level. Watt’s engine was far more efficient than Newcomen’s, and British coal consumption soared after its introduction. The paradox is that efficiency, by making a resource cheaper to use, expands the range of things worth using it for.

The Mechanism

The chain is short. Efficiency lowers the cost per unit of useful work. Lower cost raises demand for that work. If demand rises faster than efficiency improves, total resource use goes up. Efficiency also feeds growth: cheaper energy makes new industries viable, which consume still more. The per-unit and the aggregate move in opposite directions, and people reading the per-unit number mistake it for the aggregate.

The critical condition is elastic, unsaturated demand. The paradox needs people to genuinely want more of the service when it gets cheaper, with room to grow.

Rebound, Backfire, and the Modern Formalization

The modern version was formalized by Khazzoom (1980) and Brookes (1990), now called the Khazzoom-Brookes postulate. The general phenomenon is the rebound effect: the gap between projected energy savings and actual savings after an efficiency gain. It comes in three layers:

Direct rebound. The cheaper service gets used more. Efficient lighting leads to more lighting.
Indirect rebound. Money saved gets spent on other energy-consuming goods.
Economy-wide rebound. Lower effective energy prices ripple through general equilibrium, shifting prices and enabling growth.

Rebound is measured as a percentage of the expected savings clawed back. Below 100% means efficiency still nets a saving. Above 100% is backfire: total consumption rises. Jevons Paradox is the name for backfire specifically. The terms are often used loosely, but the threshold is the whole point.

When It Actually Holds

This is where the careful version matters. For most household energy services in wealthy economies, measured direct rebound is well under 100%, often 10 to 30%. You can only keep a house so warm. Saturation caps the rebound, and richer populations sit closer to saturation. Critics argue that because energy is a small fraction of total production cost in mature economies, the aggregate rebound is modest and backfire is rare.

So the paradox is conditional, not a law. It bites hardest where demand is far from saturated and the resource cost is a large share of what limits use. That describes frontier compute almost exactly: there is no ceiling on how much intelligence people will consume if it gets cheap enough, and compute cost is the binding constraint on AI applications.

The AI Case

The data is already pointing this way. Google’s per-query energy for a Gemini prompt fell 33x in a single year, yet the company’s total carbon footprint rose 48% since 2019, driven by AI buildout. Global data center electricity is projected to roughly double from 415 TWh in 2024 to 945 TWh by 2030 even as every layer of the stack gets more efficient. Cheaper reasoning tokens don’t reduce inference load, they unlock agents that emit millions of tokens per task.

The practical lesson sits inside Systems Thinking: efficiency is a lever on cost per operation, not on total consumption. If you want to bound the total, you need something that caps demand or prices the externality, because the efficiency gain will be spent, not banked. This is why per-query improvements do not resolve Data Center Externalities, and why the [private link] structure persists even as each unit gets greener.

Try it

Trace the rebound in your own tool use (an afternoon, no code). Pick a task you now hand to an LLM that you previously did by hand or didn’t do at all (drafting, summarizing, code scaffolding). Estimate how many such tasks you ran last month versus a year ago, and multiply by a per-query energy figure (~0.2 to 2 Wh). The point is to feel the elasticity: cheaper per query almost certainly raised your total count by more than the per-query cost fell. That ratio is your personal rebound.

Sources

W. Stanley Jevons, The Coal Question (1865) — the original argument
Rebound effect (conservation), Wikipedia — the Khazzoom-Brookes formalization and rebound taxonomy
The Jevons Paradox and data center carbon (SIGARCH) — the AI application