Jevons Paradox: why are AI bills rising as costs fall?

Just one year ago, companies across various fields were racing to prove their leadership in the AI revolution. Executives rushed to encourage employees to use new tech tools for almost every task, even offering incentives, believing that heavy usage was a sign of innovation and not falling behind.

But with the arrival of massive bills, the situation changed and the question became: what is the return on using AI tools? And here, the 19th-century 'Jevons Paradox' comes to mind, providing an explanation for this phenomenon.

History repeats itself

In 1865, economist William Stanley Jevons observed that improving the efficiency of steam engines did not lead to a reduction in coal consumption, but rather to its doubling, because lower costs encouraged increased usage.

The same scene repeats itself today with AI: the cost of a 'token' or digital processing unit has fallen by more than 90% since 2023, while spending on AI models has doubled.

Torsten Slok, chief economist at Apollo, explains the phenomenon: a lower token price does not push companies to reduce spending, but rather to run more agents and expand automation, raising the total despite the lower unit cost.

Bain & Company analysis confirms this: the cost per token halved between the end of 2024 and the end of 2025, while the number of tokens used rose 450% in the same period, driven by companies' desire to keep up with the latest technologies and the increasing complexity of tasks.

From innovation showcase to bill shock

At the start of the AI wave, adopting these tools was a message to investors that the company would not miss the next technological revolution. Some companies even rewarded employees for increased usage, leading some to use the most expensive models even for simple tasks to demonstrate technical leadership.

But this illusion quickly dissipated when AI tool developers (such as OpenAI and Anthropic) switched from fixed subscription models to consumption-based billing or 'digital tokens'. Finance departments then discovered that impressive productivity tools turned into hefty bills.

It became clear that running AI agents for long periods could cost hundreds of dollars per session, and in some cases, computing costs exceeded the cost of human labor itself.

This led Kosti Perikos, an official at Deloitte, to say that the prevalent belief that AI is 'cheap or free' is no longer true, while Sam Altman, CEO of OpenAI, admitted that cost has become a 'big problem' for its clients this year, unlike last year.

Moreover, productivity did not rise at the same pace as spending. According to data from Intelligence AI, only 18% of spending on digital tokens in software companies translates into products that actually reach users, while the rest is wasted on experiments and unnecessary usage.

Corporate responses

Companies vary in their spending control methods. Some have turned to 'small models' costing as little as $0.05 per million tokens versus $15 for leading models. Others have had to break down complex tasks and route each part to the most cost-effective model, impose specific token budgets per user, or resort to open-source models.

Most importantly, management has started sending a message completely opposite to the one sent months ago: do not use AI just for the sake of using it.

For example, Uber Technologies, which exhausted its AI budget in just four months with costs ranging from $500 to $2,000 per engineer per month, resorted to imposing a $1,500 monthly cap per employee, with individual dashboards and a mechanism to request over-limit.

Microsoft also restricted some employees' access to Anthropic tools, Salesforce developed a system to link token usage to actual business outcomes, and Amazon, Walmart, Cisco, and Meta joined the list of companies that have started curbing excessive usage.

Concerns

Thus, AI is moving from a phase of 'collective experimentation' to a phase of 'strict accountability', where every query and every task is measured by return indicators, not by number of uses.

Adding to this debate is the superiority of Chinese developing companies in offering lower-cost models, giving them a competitive advantage despite American companies' reservations about using them.

Critics also see that this shift towards rationalization may indicate a slowdown in the sector's growth, at a sensitive time when Anthropic and OpenAI are preparing for their initial public offerings.

In the end, just as improving steam engine efficiency more than a century ago led to more coal consumption, the lower cost of AI has led to increased usage and higher bills. Between these two scenes, the same lesson repeats: technology is not measured by its ability to impress, but by its ability to enhance productivity to justify its cost.

Sources: Argaam – Fortune – Wall Street Journal – Axios – Business Insider – Financial Times – TechCrunch

Economic investigations

AI news

Artificial intelligence

{{displayname}}

{{profession}}

{{followercount}}

{{aboutme}}