The second example is equally shocking. Uber introduced Claude Code and for approximately 5,000 engineers created an internal ranking of teams, based solely on token consumption. As a result, the company burned its entire annual budget for AI tools in April because employees… were racing to use artificial intelligence. However, the “best” users generated costs of “only” PLN 2,000. hole. monthly. COO Andrew Macdonald later publicly admitted that it was becoming increasingly difficult for him to associate this astronomical expense with real new features for users.
These two stories are not exceptions. They are a symptom of a deeper problem that affects more and more corporations. In 2026, companies are drowning not in data or competition, but in artificial intelligence bills. Just a year ago, AI was supposed to be a cheap way to replace humans. Today, it has become one of the most expensive items in operating budgets, not because of the technology itself, but because of the pathological way in which companies implement it.
Initially, AI was supposed to be a tool to increase employee efficiency and results, but companies provided it with separate metrics that were completely disconnected from whether an individual employee actually achieves better results.
Pathology of KPI and tokenmaxxing in companies
Corporations live by growth rates, and each new system must have a number of its own KPIs. Therefore, companies did not stop at just changing the settlement model. They went further and created an entire internal system in which the measure of an employee's success was no longer what he actually delivered to the company, but instead the use of AI, most easily measured by the consumption of tokens.
Instead of holding people accountable for business effects, such as goods produced, satisfied customers or real savings, they are held accountable for how much they “burned” artificial intelligence. So a mechanism was created that experts are starting to call “tokenmaxxing”: the art of maximizing the use of tokens to appear as productive as possible.
- Read also: The CEO's golden dream is ending. AI has become more expensive than humans
The result is internal rankings and dashboards that have turned work into a points game. A special panel called “Claudeonomics” was created in Meta. It tracks who uses how much AI, and those with the lowest scores end up in the spotlight and, in extreme cases, face dismissal. Engineers quickly learned how to cheat the system. They left agents running non-stop, for example counting the digits of π to a billion after the decimal point, just to avoid falling in the rankings. No one asked if it made business sense. The only thing that mattered was maintaining the position in the ranking.
“AI first” is taking its toll
|
Collagers / Shutterstock
Employees, feeling the pressure of KPIs, began to come up with more and more creative, but completely worthless ways of using up tokens. Instead of focusing on tasks, they generated dozens of versions of one email asking to “sound a little warmer.” They asked about the weather or asked for an analysis that had no impact on business. Agents working 24 hours a day on a daily basis, they have become not so much a tool to replace humans, but a tool to cheat the birth certificate.
This is a classic example of Goodhart's Law: once a measure becomes a goal, it is no longer a good measure. Management wanted to show investors the “deep transformation of AI,” so they promoted consumption as evidence of progress. Employees, fearing for their jobs, played this game all the way. Model suppliers only gained and companies lost control. Instead of efficiency, a race has been created in which the winner is the one who spends the most money on pointless queries.
The most extreme cases
The culmination of this pathology is the already mentioned case of the anonymous AI consultant. The company has not placed any limits on the use of Claude. Employees were given unlimited access, and the pay-per-token settlement system worked without any brakes. Effect? A bill of $500 million. in just one month. This story immediately went viral and symbolizes the risk of uncontrolled implementation of AI tools on a large corporate scale.
- Read also: We will pay more for AI. The end of the market where you fight for the user
The second well-documented case is Uber. The company introduced Claude Code in late 2025 and quickly created internal rankings of engineering teams that worked not by how much code was shipped or new features for users, but solely by how many tokens each “burned.” By March 2026, the adoption of the tool reached 84%. among 5 thousand engineers.
Uber logo
|
Alex Millauer / Shutterstock
As a result, Uber's management admitted that the budget planned for the entire 2026 had been exhausted in April. COO Andrew Macdonald in an interview with “Fortune” and “Business Insider” emphasized that it is increasingly difficult to justify these expenses because there is no direct impact on new consumer functions. As a result, the company had to slow down its hiring process to cover rising token bills.
The absurdity of the system and the consequences of the AI boom
Boards inflate these metrics for a very specific reason. They want to show “deep digital transformation” and being “AI ready” in quarterly reports and presentations for investors. Billions of investments in AI infrastructure, data centers and licenses must be justified somehow. It's easier to boast that “90 percent of engineers use Claude Code every day” than to admit that the return on these investments is still unclear. This creates an artificial increase in results: dashboards glow green, presentations look impressive, but in reality tokens are burned on things that do not translate into any real profit.
Tokenomics
|
Created using AI / Copilot
The consequences of this absurdity are already visible and painfully concrete. Companies don't fire people because AI took over their jobs. The layoffs come as the AI bill has begun to replace jobs in the budget. As Mark Ajzenstadt, founder of Limestone, aptly put it: “Companies are laying off people to pay the AI bill. Not because AI has replaced their jobs. Because the bill has replaced their wages.” Instead of the savings that were dreamed of at the beginning of the boom, there is a new, very expensive item in operating costs.
The bigger picture is even more disturbing. Even at Nvidia, the company that makes the most money from this entire AI ecosystem, the vice president openly admitted that in his team, computing costs have already exceeded the costs of employing people. Big Tech, which not long ago was throwing billions at AI infrastructure, is starting to slow down.
- Read also: AI drives investment. However, there are two new risk factors
The first companies to implement the tools at full scale are now withdrawing licenses, slowing down hiring and looking for ways to reduce expenses. It turns out that the more AI we use, the more expensive it costs us and the more difficult it is to find economic sense in the whole process.
How and why AI costs are skyrocketing
The change to the AI pricing model was a turning point that many companies didn't anticipate. Back in 2024 and early 2025, most enterprises paid for AI tools in a simple, predictable way. A monthly subscription gave access to a specific pool of queries or unlimited use within a set limit. It was convenient and easy to budget for, just like paying for office software or cloud computing. Today, this model is becoming a thing of the past.
Claude Code
|
aileenchik / Shutterstock
It was replaced by a pay-per-token system, i.e. payment for each individual token, i.e. the smallest unit of text that AI processes or generates. In practice, this means that the company no longer pays a fixed amount once a month, but is billed for each prompt sent to the model, for each agent's response and for each tool call. Even a seemingly trivial task, such as “improve this e-mail to make it sound more friendly” or “analyze this report and propose three conclusions”, generates a specific cost. The longer the text and the more complex the query, the higher the bill.
- Read also: China has started a price war for AI. It's time for the West to move
In addition, the scale of consumption is growing exponentially. Companies are no longer limited to simple chatbots used several times a day. They are massively deploying autonomous agents – AI-based programs that work 24 hours a day. around the clock, seven days a week. Such an agent independently searches for data, generates reports, responds to e-mails and conducts business simulations. Each such process consumes thousands of tokens per minute. Even if the price of a single token drops by 90% in the coming years, the total costs will still increase because the amount of tokens consumed by systems is growing much faster than the price drop.
As a result, a pathological system was created in which the measure of success is not the business value, but the consumption of tokens. Employees are held accountable for KPIs based on how much they “burned” artificial intelligence. Management boards pump up usage statistics to show “deep digital transformation” in reports and justify billions of dollars in infrastructure investments. And in the end, everyone pays the bill – first the company's budget and then, unfortunately, the employees in the form of layoffs. This is a classic case where the tool no longer serves the purpose, and the purpose becomes a tool to increase metrics.
