Jevons Paradox
The Jevons paradox, first articulated in The Coal Question (1865), is the empirical observation that efficiency improvements in the use of a resource often produce increased rather than decreased total consumption. Jevons noticed that steam engines had become dramatically more efficient over the preceding century, yet British coal consumption had risen, not fallen. His explanation: cheaper energy per unit of work made new applications economically viable, and the total demand generated by the new applications exceeded the savings from the efficiency improvement.
The Substrate Dependency Trap — Contrarian ^ Opus
There is a parallel reading that begins from the material foundations of computation itself. The Jevons paradox, when applied to AI, isn't merely an economic phenomenon that could be redirected through institutional redesign — it's rooted in the thermodynamic realities of information processing. Every bit manipulated requires energy; every model trained demands rare earth minerals; every datacenter needs water for cooling. These are not efficiency problems but existence problems. The substrate that makes AI possible — from lithium in batteries to gallium in semiconductors — comes from extractive processes that cannot be made "regenerative" in any meaningful sense. Mining cobalt in the Congo or lithium in Chile doesn't become ecologically benign because we've reorganized our economic institutions.
The deeper trap is that AI efficiency gains don't just enable more queries; they enable qualitatively new forms of computational dependency that lock in future consumption. When efficiency makes real-time video analysis cheap enough for every security camera, or makes language models accessible enough for every email client, we're not just expanding usage — we're building infrastructures that assume and require these computational intensities going forward. The "doughnut economy" framework assumes we could choose to capture efficiency as sufficiency, but the political economy of AI development suggests otherwise. The companies building AI infrastructure have already committed trillions in capital expenditure based on exponential growth assumptions. The geopolitical competition frames compute as a strategic resource. The venture capital model that funds AI research requires returns that only hypergrowth can deliver. These aren't design choices we can simply reverse; they're path dependencies we've already traveled far down.
In the AI Story
The paradox operates across domains with remarkable consistency. Computing hardware has become vastly more efficient per operation since the 1960s; total computing energy consumption has risen, not fallen. LED lighting is dramatically more efficient than incandescent; total lighting-related electricity consumption has risen, not fallen. The efficiency gains are captured by growth logic and converted into more throughput rather than less resource use.
Applied to AI, the dynamic is direct. Each generation of AI hardware does more computation per watt. Each generation of model architecture achieves more performance per parameter. These efficiency gains are real. They are also being overwhelmed by the expansion of total computation: total energy consumption is rising, total water consumption is rising, total material extraction for hardware is rising. The AI industry's efficiency improvements are not producing ecological space; they are producing cheaper queries, which drive adoption, which drive total consumption.
In Raworth's framework, the Jevons paradox is the mechanism by which growth-addicted economies convert efficiency into expansion rather than into ecological relief. A doughnut economy would capture the efficiency differently — using the reduced energy per query not to process more queries but to reduce total resource consumption, creating room within the ecological ceiling for other essential activities. The same hardware, running the same models, with the same efficiency, produces radically different ecological outcomes depending on whether the governing economic logic is growth-oriented or doughnut-oriented.
The paradox is therefore not a technology problem but an economics problem. Engineering efficiency cannot, by itself, deliver ecological space within a growth-addicted system. The system must be redesigned to convert efficiency into sufficiency rather than into throughput.
Origin
William Stanley Jevons (1835–1882) was a British economist and logician who made foundational contributions to marginal utility theory. The Coal Question (1865) argued against contemporary complacency about British coal reserves and, in the process, articulated the efficiency paradox that now bears his name.
Key Ideas
Efficiency expands consumption. In a growth-oriented system, efficiency gains make new applications viable, driving total consumption upward.
Cross-domain consistency. The pattern holds across coal, electricity, computing, lighting, and transport.
AI instance. AI efficiency gains are being converted into query volume expansion, not into reduced total energy or water consumption.
System-level problem. The paradox cannot be solved by more efficient engineering; it requires institutional redesign of how efficiency gains are captured.
Appears in the Orange Pill Cycle
The Efficiency-Sufficiency Bridge — Arbitrator ^ Opus
The tension between these views resolves differently depending on which timescale and system boundary we examine. For immediate material impacts (next 5-10 years), the contrarian reading dominates — perhaps 80% right. The substrate dependencies are real, the capital commitments are locked in, and the infrastructural momentum is massive. No amount of institutional redesign will make lithium extraction regenerative or prevent the next generation of datacenters from being built. Here, the Jevons paradox operates almost mechanically.
But shift the frame to institutional possibilities (10-30 years), and the weighting changes to roughly 60/40 in favor of Edo's view. While we cannot unmake the material realities of computation, we absolutely can design systems that capture efficiency gains differently. Public utilities don't maximize consumption the way private markets do. The internet's underlying protocols were designed for resilience, not profit. The precedent exists for treating computational resources as public goods with usage patterns oriented toward sufficiency. The contrarian is right that current political economy makes this unlikely, but wrong that it's impossible.
The synthetic frame that holds both views is this: the Jevons paradox in AI is simultaneously a thermodynamic reality and an institutional choice. The material facts of computation create hard limits that no economic reorganization can overcome — every query will always require energy, cooling, and hardware. But within those limits, the difference between a 10x expansion and a 1000x expansion of AI computation is very much a function of institutional design. The real work isn't choosing between efficiency and sufficiency, but recognizing that we need new institutions precisely because the material constraints are so severe. The efficiency gains are coming regardless; the question is whether we build systems that treat them as room to breathe or as fuel for acceleration.
Further reading
- William Stanley Jevons, The Coal Question (1865)
- Blake Alcott, "Jevons' Paradox," Ecological Economics (2005)
- John Polimeni et al., The Jevons Paradox and the Myth of Resource Efficiency Improvements (2008)
- Mario Giampietro et al., Energy Analysis for a Sustainable Future (2013)
- International Energy Agency, Electricity 2024 (2024)