Jevons' paradox and the future of AI infrastructure: a misapplied economic theory?

Jevons' paradox, first described by 19th-century economist William Stanley Jevons, states that as technology enhances efficiency, the overall consumption of a resource may increase rather than decrease. Jevons witnessed this during the Industrial Revolution: more efficient steam engines lowered the cost of coal-based energy and led to greater overall demand, rather than conservation. Today, this is the concept taken to AI by some industry leaders when they claim that the efficiency gains of AI, as expressed in DeepSeek's breakthrough performance, will drive exponential growth in demand for chips, GPUs, and, in general, AI infrastructure.

But is Jevons' paradox an ironclad economic law? In some cases, yes, it holds, but it is very far from being a deterministic rule. The efficiency-consumption relationship is rather more nuanced than Jevons' framework suggests, especially in the high-tech sector, which often faces market saturation, limited capital expenditure, and shifting economic priorities that dampen or even eliminate the rebound effect. AI infrastructure may not follow the same trajectory as coal or energy consumption in Jevons' original context, and historical examples do suggest that efficiency gains have not necessarily begotten runaway demand.

Jevons' paradox is often invoked in discussions of energy efficiency, especially in transportation and manufacturing industries. The automobile industry conventionally shows elements of the paradox. As technology in engines improved, the fuel efficiency rose, a factor that reduced the price of driving per mile. This encouraged greater car use, suburban expansion, and even the proliferation of larger vehicles such as SUVs, ultimately leading to higher overall fuel consumption.

The gains in electrical efficiency sometimes result in more overall use of the resource - electricity. As innovative technologies such as LED bulbs made the cost of lighting plummet, the overall energy consumption of lighting did not shrink commensurately because people started using lighting more widely. These cases corroborate Javons’  contention that increased efficiency leads to more overall consumption.

However, other historical examples run counter to this pattern. What this says is that, when compact fluorescent and LED lighting came onto the market, that did not result in a runaway increase in the total energy demand-the market eventually reached saturation whereby further efficiency gains meant lower electricity bills, not exponentially higher consumption. Similarly, while industrial automation greatly enhanced productivity in manufacturing, it did not lead to an unbounded increase in the number of factories or production facilities. Instead, automation has led to the restructuring of jobs and changes in industry focus rather than an endless expansion of manufacturing output.

AI infrastructure: efficiency does not guarantee exponential growth

Application of Jevons' paradox to AI infrastructure assumes that there will be explosive demand for chips, GPUs, and data centers as a result of improvements in AI efficiency. There are several important economic and technological constraints that this hypothesis doesn't take into consideration.

1. Market maturity and diminishing returns

AI infrastructure is not like coal or fuel, an unlimited resource. Technology companies are bound by economic and logistical realities: CAPEX for AI hardware is subject to budgetary limits, competitive pressures, and diminishing returns. While demand for AI computing power may continue to increase in the near term, companies will eventually reach a point where adding more GPUs or data centers does not yield proportional improvements.

2. Regulatory and environmental factors

Unlike coal in Jevons' time, AI infrastructure lives in a heavily regulated environment. Already, governments are starting to look at the environmental impact of giant data centers based on their energy use and carbon output. That could put limits on how much infrastructure AI companies can deploy, which would hamstring the exponential growth that Jevons' paradox would suggest.

3. Shifting technological development

History is rife with the fact that technology industries do not always grow infinitely into one direction. As microprocessors got more powerful, the answer was not an infinite proliferation of personal computers but a shift to efficient small devices such as smartphones. This might be so with AI: the uncontrolled growth of GPU and chip consumption may give over to the focus on more economical AI models and hardware and a search for novel computational architectures.

4. Economic and strategic decision-making

Companies do not invest in infrastructure based on pure efficiency gains. Business strategy, cost-benefit analysis, and external economic conditions all come into play. If AI processing becomes much more efficient, for example, firms may choose to shift resources elsewhere rather than continue scaling up computing power. For example, gains in AI efficiency might increase the adoption of AI by small and medium-sized enterprises rather than simply driving more GPU purchases by tech giants.

A flawed prediction?

Jevons' paradox is a useful concept in some economic contexts; applying it to AI infrastructure is less direct. Inasmuch as breakthroughs in AI would power demands for more computing - say, DeepSeek - the assumption of efficiency gains leading to an unchecked infrastructure expansion simply remains overly simplistic. History is riddled with efficiency improvements that did not lead to exponential growth: market saturation, regulatory concerns, shifting priorities of the industry.

Instead of blindly applying Jevons' paradox, industry leaders should consider a broader economic and strategic landscape. Growth in AI infrastructure will likely be tempered by a combination of financial constraints, environmental concerns, and evolving AI itself. Efficiency can indeed drive adoption and new applications, but it does not guarantee runaway demand for hardware. In the case of AI, reality is likely to be much more nuanced than the paradox would suggest.

illuminem Voices is a democratic space presenting the thoughts and opinions of leading Sustainability & Energy writers, their opinions do not necessarily represent those of illuminem.