Can We Test Jevons Paradox in AI Inference? A Longitudinal Audit of Token Prices, Demand, and Energy (2022–2026)
This paper constructs a longitudinal dataset and a new analytical framework to demonstrate that current public data lacks the necessary granularity and causal identification to verify whether AI inference efficiency is driving a Jevons Paradox, as observed price stability and unverified volume growth prevent a definitive conclusion on whether efficiency will ultimately limit or increase total energy demand.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine a world where making a machine smarter and cheaper to run actually causes us to use it so much more that the total energy it consumes skyrockets. This is the heart of a question that has puzzled economists for over a century, known as the Jevons paradox. It suggests that when we become more efficient at using a resource, we don't necessarily save that resource; instead, the lower cost encourages us to use it so aggressively that our total consumption grows. Today, this question has moved from coal mines to the digital realm, centering on artificial intelligence. As AI models become faster and cheaper to operate, there is a fierce debate about whether this efficiency will cap the industry's massive energy appetite or, conversely, fuel an explosion in demand that outpaces our ability to generate power. To settle this, we need to see three things happening at once: the price of using the AI, the volume of how much it is being used, and the actual energy it consumes.
A researcher named Eugen Spivak set out to find these three pieces of the puzzle for the years 2022 through 2026. The goal was to build a single, unified record that tracks the price of AI text, the amount of text generated, and the energy required to create it, all for the same services and time periods. Spivak spent years gathering data from public sources, including archived versions of pricing pages from eleven major AI developers, statements about how many tokens (the basic units of text) were processed, and published figures on energy usage. The result was a massive ledger containing over a thousand price points, dozens of volume statements, and more than a hundred energy figures. However, the most significant finding of this work is not a confirmation of the paradox, but a revelation about the data itself: the public record simply does not contain the information needed to answer the question.
The investigation revealed that while we have plenty of numbers, they rarely fit together in a way that allows for a fair comparison. For instance, when researchers looked at the prices of specific AI models over time, they found that the cost for a single model rarely changed. In nearly 700 instances where the same model was checked at different times, its price remained exactly the same. The dramatic price drops often cited in the news were not due to developers lowering the cost of existing tools, but rather because new, cheaper models were introduced to the market while the old ones stayed expensive. This distinction is crucial because it means the "price" of AI is shifting due to product changes, not just efficiency gains.
When the researcher tried to match these prices with how much people were actually using the AI, the data became even more fragmented. For the two largest providers, Google and OpenAI, there are clear statements showing that the amount of text generated on their platforms grew incredibly fast, multiplying by more than four or five times every year between 2025 and 2026. At the same time, the prices for their standard models did not fall; in fact, for many categories, they rose. This creates a confusing picture where usage is exploding even as the cost per unit is not dropping. However, a critical gap remains: the volume statements do not specify exactly what kind of text they are counting. They might include input text, output text, or cached data, while the price lists only cover specific types of output. Without knowing if the volume numbers and the price numbers are talking about the exact same thing, it is impossible to say if the growth in usage was a direct response to a price drop.
The missing piece that makes the whole picture impossible to complete is the energy data. While there are many claims about how much energy an AI uses, most of them do not state the boundaries of their measurement. Some figures count only the computer chips, while others include the entire building, the cooling systems, and the electricity used to keep the servers running even when they are idle. Because these measurements are taken on different scales, they cannot be reliably compared to one another, let alone matched to the price and volume data. The study found that in over half of the energy figures reviewed, it was impossible to tell exactly what was being measured. Without a clear, consistent measurement of energy per unit of work, and without a clear link between price and volume, the data cannot show whether efficiency is driving higher consumption.
Ultimately, this rigorous audit concludes that we cannot yet say if AI is suffering from the Jevons paradox. The public record is too patchy to prove that efficiency gains are causing a surge in total energy use, nor can it prove that efficiency will eventually limit that use. The data shows that AI usage is growing rapidly, but it does not show that this growth is caused by lower prices or better efficiency. The researcher argues that to solve this mystery, the companies building these systems must start publishing clearer data. They need to state exactly when a price takes effect, specify exactly what their volume numbers count, and define the boundaries of their energy measurements. Until these details are provided, the debate over AI's energy future will remain a matter of theory rather than fact, leaving us unable to know if the machine is becoming a sustainable tool or an insatiable consumer of power.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.