Jevons Paradox
Good news for consumers. Bad news for the companies spending trillions to sell them intelligence.
I didn’t know what Jevons Paradox was until about a year ago and I realize I am joining a chorus by writing about it, but I find it an interesting concept and felt I needed to learn more.
From Wikipedia:
The Jevons paradox or Jevons effect is an economic phenomenon said to occur when technological improvements that increase the efficiency of a resource’s use lead to a rise, rather than a fall, in total consumption of that resource. Greater efficiency reduces the amount of the resource needed per application, lowering its effective cost; if demand is sufficiently price elastic, this induces demand, frequently resulting in a net increase of total resource consumption.
The classic example used to describe the paradox is fuel consumption. William Stanley Jevons himself observed the phenomenon from Britain rise during the industrial revolution as steam engines become more efficient. The assumption was that greater efficiency in steam engines use of coal would decrease consumption, however the total amount of coal demand had actually increased instead of declining.
A more modern example: the increased fuel efficiency of cars on the road, i.e. more miles per gallon, has actually increased the total amount of miles a person drives therefore increasing total demand / consumption, on average. Economists call this the rebound effect, and to be precise about the distinction driving did increase, but not by enough to raise total fuel consumption. The efficiency gain was partially offset, not reversed. The full paradox, where consumption actually rises, is the extreme case and in the 160 years since Jevons first wrote about it, it has been rarely observed.
The theory is having a resurgence because it is being used to describe what AI might do to the economy. The argument goes: compute is getting radically cheaper per unit of intelligence, cheap intelligence will find unlimited uses, therefore total demand for compute, and for the energy, chips, land, and grid capacity behind it, explodes. Buy everything upstream of the datacenter.
I think the people making that argument are right about the mechanism and have not noticed what it implies.
The condition nobody prices
For the paradox to occur, three things must be true. (1) There has to be a technological change that increases efficiency, (2) the increased efficiency has to result in a lower price for the good or service, and (3) that lower price has to drastically increase demand.
Conditions two and three are critical in understanding what type of product it is, one where the provider does not have pricing power and which the consumer has very low switching costs. Efficiency gains only pass through to price when the provider cannot hold onto them (the consumer has the bargaining power), a company with pricing power keeps efficiency gains as margin. A company in a commodity market hands it to the customer, otherwise their competitors will (high elasticity of demand). Their product is not differentiated so competition is fierce.
Every time someone invokes Jevons to justify infinite compute demand, they are also asserting, whether they notice or not, that intelligence is a commodity and the model layer has no pricing power. The two claims travel together. You cannot have exploding demand driven by collapsing prices and durable margins at the layer where the prices are collapsing. The same premise that makes the paradox bullish for the inputs makes it bearish for the companies spending trillions to win the market.
If Jevons paradox holds for AI, that is great for the consumer of intelligence, but terrible for the big tech companies spending trillions of dollars to compete in the market of artificial intelligence.
Testing the conditions
The first condition of the phenomenon is overwhelmingly met. The cost of a unit of model output has fallen significantly in the past three years, through better chips, better models, and better inference.
The second condition has been met so far. Efficiency gains have flowed to price rather than margin. Frontier model pricing has fallen relentlessly, open-weight alternatives sit underneath every closed model as a ceiling on what anyone can charge, and switching costs at the model layer remain low. Today, the switching costs for a developer are low. The model layer is behaving like a commodity. A clean example is OpenAI’s recent price cuts to take back market share from Anthropic. That is not the behavior of a company with a competitive advantage.
Condition three is an open question. Consumption has increased as prices fall (though there have been some signs of stagnating demand). But the historical cases where the full paradox held involved goods where demand was effectively bottomless at the right price: mechanical work in an industrializing economy, artificial light, miles of personal mobility. Whether demand for intelligence behaves the same way, whether every drop in the price of a token durably creates more than a proportionate rise in tokens consumed, is not something three years of data can settle. It is the load-bearing assumption, and it is unproven.
What the companies are doing about it
I don’t think the people running these companies are dumb. My argument is not that these companies don’t see what is happening. I think they all want to avoid the current path of destructive pricing wars on tokens so they are building products to break the condition. They want high switching costs, they want pricing power. That is easier said than done, and it could be unavoidable. I am not making any such call, just sharing my perspective on where we are and where this might be going.
What I am watching
First, the variable costs. To avoid the paradox, there needs to be meaningful increases in token margins. Does cost of tokens continue to decline relative to the price of/revenue from tokens or is it the opposite? This is critical, otherwise it’s a commodity. Which, by the way, seems to be the route Apple and Microsoft are taking. I think they see the writing on the wall and are heading towards model agnostic products. They both have a gigantic installed base (the iPhone, Microsoft 365) and they intend to allow all models on their platforms, even the open-weight ones. Microsoft is doing both at once, funding a frontier lab while installing other models into their products. That is a hedge, and it tells you something that the company closest to the spending is the one least willing to bet on it.
Second, the fixed costs. Can pricing power show up before the money runs out? Less of a problem from Google, Microsoft, and Meta, as they have their core profit engines to continue funding the investment, even though trillions in wasted capital will not go over well. Anthropic, OpenAI, and, to a lesser but still significant extent, SpaceX (xAI) will need to both differentiate their product offerings while stabilizing the costs of infrastructure. I am not so sure the capital sources they have had so much success tapping into will run forever.
Third, whether usage keeps growing faster than prices fall. That is condition three, and it is the one I said the whole thing rests on. If that gap ever closes, the paradox was never here, and a lot of capacity got built for demand that does not show up.
I am not making a call on how this ends. But the two halves of the Jevons argument go together. If intelligence gets cheap enough that we use it without thinking about the cost, that is because nobody managed to hold the price. Jevons was writing about the coal, not about the people selling engines.
—
Michele Chiacchio
August 18, 2026
Important Disclosures
All publications are provided for informational and educational purposes only and do not constitute financial, investment, or economic advice. Investing involves risk, including the possible loss of principal. The views expressed are solely those of the author and do not necessarily reflect those of any affiliated organizations.