Amidst the talk of AI bubbles, perhaps you’ve encountered the concept of the Jevons Effect. Jevons’ argument has to do with coal, hence the title of his work, The Coal Question. The principle is that an increased efficiency in the use of a resource (e.g., coal) might result in an expanded consumption of that resource. It wasn’t “good news.” It was a caution that increasing the efficiency of coal use wasn’t going to result in the conservation of Britain’s coal supply.

However, proponents of AI, or more precisely proponents of investment in AI, have deployed the Jevons Effect differently. Microsoft CEO Satya Nadella, referenced Jevons in response to DeepSeek AI in January of 2025. “As AI gets more efficient and accessible,” he posted, “we will see its use skyrocket, turning it into a commodity we just can’t get enough of.” The problem is that Jevons promises nothing of the sort. In fact, he makes no promises at all. He cautions this may happen with coal. Given the enormous energy demands of AI, Nadella’s assertion ends up paradoxically, closer to demonstrating Jevons concerns than delivering the investment thesis he intends.

Since then, AI proponents have used Jevons to suggest that falling inference costs mean that there will be expanding demand. Then they simply imply that expanding demand will turn into commercial success. But that’s not what Jevons says. At most, the Jevons predicts that efficiency may increase aggregate consumption. It cannot tell us what the resulting environmental, cultural, economic or political effects will be.

Without getting out my calculator, I think there is a reasonable argument that the current scale of AI investment, especially as it is projected to continue by folks like Gartner, would require a Jevons effect transformation on the scale of the Industrial Revolution. And certainly visions on that scale have been described by the frontier AI labs.

But when I think about the Jevons effect, I think of Victorian London and Steven Johnson’s The Ghost Map, which relates London’s 1854 cholera outbreak. Or any one of Dickens novels and so on. A financial success? For some certainly. But not a historical period that would have anyone signing up for a second ride. In some ways it’s not so different from how Silicon Valley likes to read cyberpunk as inspiration.

While AI hype generally stops at the “effect,” there is considerably more to Jevons’s work. His Theory of Political Economy contributed to the mathematical formalization of economics. In Jevons’ account of desire economic value depends not simply on the labor or cost embodied in a commodity but on the changing utility that a consumer attributes to its next, or “final,” unit. [This is his theory of marginal utility. I’ll come back to this point.]

In short, Jevons offers economics a subjective theory of demand. His account establishes the consumer subject as a key determinant of value. Later theories such as revealed preference inherit this problem of consumer choice, though revealed preference moved away from Jevons interest in subjective experience and toward empirical observation.

In A Thousand Plateaus, Deleuze and Guattari describe Jevons as a “Lewis Carroll of economics.” And it is from this reference that this post gets its title. For D&G, Jevons, like Carroll, is a thinker of series, limits, and thresholds. This comes up in the “Apparatus of Capture” plateau. D&G offer this example. There are two tribes, one with seeds, the other with axes. In the exchange of seeds for axes, the rate is determined by the last axe or last quantity of seeds that each tribe could accept without its assemblage being transformed. That is the limit and the threshold (into another state) that establishes the series. E.g., if I take more axes, then I am going to have to find new uses for axes. I will need to change. They also discuss the exchange in terms of the alcoholic.

For example, what does an alcoholic call the last glass? The alcoholic makes a subjective evaluation of how much he or she can tolerate. What can be tolerated is precisely the limit at which, as the alcoholic sees it, he or she will be able to start over again (after a rest, a pause …). But beyond that limit there lies a threshold that would cause the alcoholic to change assemblage: it would change either the nature of the drinks or the customary places and hours of the drinking. Or worse yet, the alcoholic would enter a suicidal assemblage, or a medical, hospital assemblage, etc. (ATP 438).

The Jevons effect intersects with these limits. If a household suddenly has more clean water available, it will use that water. But there is no exchange for the water in that case. In the simple example above, the increase in the efficiency of axe-making wouldn’t only require an increase in axe-use, it would require a reciprocal increase in seed generation and seed use. Otherwise the seed makers would have nothing to trade. Money, specifically capital, accelerates this process by abstracting the process, a la M-C-M.

The Jevons effect does impact the marginal utility of a commodity. First, by making the commodity cheaper, it expands the limit, so people use more. That works with water or electricity, but less so with alcohol as there are other limits to consider. Second, as was the case with coal, new uses were invented, which increases both marginal utility of the commodity by identifying new industries driven by the product. None of this was simply good news. Jevons was aware of the Railway Mania of the 1840s and he would live through the Long Depression that began in 1873. At the time, some called it the Great Depression (until 1929). The Long Depression was global in some respects, but its national effects in the UK included industrial decline on the global stage.

Historically, that’s how the “coal question” played out in the UK in the 19th century. Nothing in there should suggest that the Jevons effect was some kind of technological determinism, and even if there was such claim, would you want this determined outcome?

I could continue here with a critique of whether or not a Jevons effect is occurring with AI, but I think that might be criti-hype at this point. To be clear, technologies do not need to unfold that way to have value. For example, we build aircraft carriers for the navy. AI could be something like that. Not strictly military of course, there would be civilian research-government uses (but obviously military as well). Technologies do end up that way. Aviation runs from fighter jets to prop planes. Indeed, AI could end up like the airlines: a cutthroat race to the bottom. Aviation, like coal, is another example of how the Jevons effect plays out materially and historically.

But I want to end by returning to the connection between Jevons and Carroll. Though this is a brief aside in ATP, it connects back to The Logic of Sense, particularly if we think of Jevons as a theorist of series, limits, and thresholds. The AI wonderland, created around Jevons, imagines that cheaper inference will result in an increased demand for inference, and that this cheap inference will also produce new markets and industries that will rely on inference. It fantasizes causation where there are only fantasies.

For this AI bubble world to make sense, inference would have to be in operation everywhere there is electricity. Inference in operation means AI sensing and operating. What would it be sensing and operating on? If it is going to be similar to electricity then AI will infer to the same extent that electric light illuminates. It is a world of fantastical series, limits, and thresholds. In which industrial series does inference become incorporated? Do those industries encounter inference’s marginal value or do they transform into new assemblages that expand that marginal value? Many industries and corporations are attempting it the other way around. They seem to be arguing that because AI inference will be cheap someday (and is currently subsidized to be cheap), we must organize ourselves around it before our competitors do. However, there’s no real sense of how to do that and not much evidence of success.

The AI-Forward university is engaged in a similar kind of FOMO. They, like many others, were convinced by the inevitability argument around AI (to which their use of Jevons contributes). And who knows what the future of AI will be? What I will say is that if one is convinced that Jevons has a point that’s relevant to AI, then the point is to be cautious.

On the other hand, celebrating the Jevons effect sounds like the theme of a Wonderland Tea Party.