Jevons Paradox
A reader's summary of William Stanley Jevons's 1865 observation that a more efficient engine increased coal consumption rather than cutting it, the economic formalization that followed a century later, and why the same reasoning keeps resurfacing in arguments about whether cheaper AI inference will actually reduce total compute and energy demand.
The paradox at a glance
Jevons Paradox describes a situation where making a resource more efficient to use — so that less of it is needed per unit of output — leads to an increase, not a decrease, in total consumption of that resource, because the efficiency gain lowers the effective cost of use enough to expand demand by more than the saving. It is a claim about aggregate, economy-wide behavior under specific demand conditions, not a claim that efficiency improvements are pointless or that they always backfire.
Origin
William Stanley Jevons, an English economist, made the argument in his 1865 book The Coal Question, addressing British anxiety over the country's finite coal reserves. He observed that James Watt's improvements to the steam engine, which used coal far more efficiently than earlier Newcomen engines, had not reduced Britain's coal consumption — instead, cheaper effective power opened up so many new industrial uses for steam engines that total coal consumption rose sharply, the opposite of what efficiency advocates of the day had predicted.
History and context
The argument was largely confined to historical and energy-policy economics for over a century until Daniel Khazzoom and Leonard Brookes revived and formalized it in the 1980s, in direct response to energy-conservation policy that assumed efficiency mandates would straightforwardly cut national energy consumption — a claim that became known as the Khazzoom-Brookes postulate and fed into ongoing debates over vehicle fuel-economy standards and building insulation codes. The reasoning resurfaced with new visibility in the mid-2020s in commentary on AI infrastructure, after episodes such as the market reaction to DeepSeek's efficiency claims in early 2025, when commentators — including figures at major AI labs — invoked Jevons's original argument to explain why cheaper inference was more plausibly a driver of higher total compute demand than a reason to expect data-center buildout to slow.
Main ideas
Efficiency lowers the effective price of use
The mechanism runs through price, not virtue: making a resource cheaper to use per unit of output lowers its effective cost, and lower effective cost increases the quantity demanded — sometimes enough to outweigh the per-unit saving and raise total consumption.
Jevons's coal engine observation
Jevons's original 1865 case was Watt's more fuel-efficient steam engine: rather than shrinking Britain's coal consumption, the engine made coal power viable for far more applications, and total coal use rose as adoption spread faster than efficiency reduced use per engine.
Rebound effect vs. Jevons paradox (backfire)
Economists distinguish a spectrum: a partial rebound effect, where some but not all of an efficiency saving is offset by increased use, from the full Jevons paradox or 'backfire,' where increased use more than offsets the saving and total consumption rises — the paradox is the extreme end of the spectrum, not every instance of rebound.
The Khazzoom-Brookes postulate
Economists Daniel Khazzoom and Leonard Brookes formalized and extended the argument in the 1980s energy-economics literature, arguing that improved energy efficiency at the macroeconomic level tends to increase, not decrease, aggregate energy consumption once economy-wide adoption and growth effects are accounted for.
It's an empirical claim, not a universal law
Whether backfire actually occurs, and how large it is, depends on the resource's demand elasticity in the specific market — some efficiency gains produce only partial rebound with a net reduction in total use, which is why the paradox is contested case by case rather than assumed automatically.
The 2020s AI-compute reframing
Commentators invoke the paradox to argue that cheaper inference per token, rather than shrinking total AI energy and compute demand, instead expands the range of tasks for which running a model is worthwhile — producing more total inference calls and more aggregate compute and energy use even as per-query cost falls.
Critique
- Backfire is the exception, not the default. Empirical estimates of rebound effects across energy-efficiency studies mostly find partial rebound well short of full backfire — treating every efficiency gain as certain to increase total consumption overstates how often the extreme end of the spectrum actually occurs.
- Conflated with ordinary rebound in casual use. Popular invocations of “Jevons Paradox” frequently describe any rebound effect at all, rather than the specific claim that total consumption rises — collapsing a spectrum of outcomes into a single dramatic label.
- Demand elasticity is doing all the real work. Whether backfire occurs depends entirely on how responsive demand for the resource is to its effective price in that specific market — a fact that gets lost when the paradox is treated as a general law rather than a conditional prediction.
- The AI-compute version is still mostly an analogy. Applying Jevons's 19th-century coal argument to 2020s AI inference is a structurally plausible parallel, but it is argued from historical precedent and directional trends rather than measured demand-elasticity data specific to AI workloads.
Impact
The paradox is now a standard reference point in energy policy and, increasingly, in AI-industry commentary for why efficiency gains in model inference are argued to expand rather than shrink total compute buildout — a framing that sits alongside Scaling Laws, which describes the demand side of the same story: predictable returns to more compute are exactly what gives falling per-unit cost somewhere new to go rather than simply lowering total spend.