AI Spending Surges Despite Cheaper Tokens, Economists Cite Jevons Paradox
Despite a more than 90% drop in the price of individual AI tokens since 2023, corporate spending on large language models has doubled since late last year, according to the Silicon Data Token Expenditure Index. This trend, where increased efficiency leads to greater consumption, is being linked to the 19th-century economic observation known as Jevons paradox. Economists warn that as tokens become cheaper, companies tend to increase their use of AI agents, automate more workflows, and generate more code, leading to higher overall expenditure.
A phenomenon observed in the 19th century, known as Jevons paradox, is being cited to explain current trends in artificial intelligence (AI) spending. In 1865, William Stanley Jevons noted that increased efficiency in coal use by the Watt steam engine led to a significant rise in overall coal consumption, not a decrease.
Today, economists are applying this paradox to AI. Despite a reduction of over 90% in the price of a single AI token since 2023, aggregate spending on large language models has doubled since late last year, as reported by the Silicon Data Token Expenditure Index. Torsten Slok, chief economist at Apollo, identifies this as a prime example of Jevons paradox in action. He explained that as tokens become more affordable, companies expand their AI applications, running more agents, automating additional workflows, and generating more code, which ultimately increases total expenditure even as the unit cost of intelligence declines.
Companies are grappling with these rising AI costs. Uber's president and chief operating officer, Andrew Macdonald, stated that the rideshare company exhausted its entire AI budget within the first four months of the year due to increased use of its Claude Clode model. Bloomberg reported that Uber has since imposed a monthly AI spending cap of $1,500 per employee. Bryan Catanzaro, vice president of applied deep learning at Nvidia, also noted that for his team, compute costs for AI far exceed employee costs.
Analysts at Bain and Co. corroborated Slok's observations, finding that while token costs were halved between December 2024 and December 2025, token consumption grew by 450% during the same period. They attribute this to companies feeling compelled to upgrade to newer, more efficient AI models rather than saving costs by sticking with older versions. Additionally, the number of tokens required per query has increased as AI agents become capable of more complex tasks.
This paradox also extends to employment figures. Slok found that despite AI's ability to automate 86% of tasks for customer service workers, employment for call center workers in the Philippines has nearly doubled over the past decade. Similarly, the number of radiologists in the U.S., a profession once thought to be endangered by AI, has increased by 10% in the last 10 years. Slok suggests that lower costs per interaction lead to more interactions, more customers served, and an expansion of markets. Bain and Co. projects a future where a company’s operating expenses are split 70% from human headcount and 30% from tokens.
According to Fortune, these trends suggest that the era of "tokenmaxxing," or employees rapidly increasing their AI use, may be ending, but it will not necessarily resolve companies' AI budget challenges.
