At the beginning of this year, employees were still being told by their managers to use artificial intelligence as much as possible. Employees turned it into a sport. On internal leaderboards, they competed to see who could consume the most tokens. A token is the basic unit of work used by AI tools, roughly equivalent to one word. But the bills for AI tools began piling up enormously.
Meta, Uber, Amazon, Walmart, Accenture. One name after another has appeared in headlines in recent weeks as companies abruptly change their views on tokenmaxxing. From unlimited encouragement to strict rationing. A new term making its way through the tech world is tokenminimizing, the deliberate and targeted reduction of AI usage.
What tokenmaxxing looked like in practice
One OpenAI engineer processed 210 billion tokens in a single week. That is enough text to fill the entirety of Wikipedia thirty-three times. One Anthropic user spent more than $150,000 in a single month on the Claude Code tool alone. And these are not exceptions; this was, and still is, a trend introduced by the companies themselves.
Meta introduced an internal leaderboard nicknamed Claudeonomics, where employees competed over token consumption. In just over thirty days, they racked up more than 73 trillion tokens. Amazon launched a similar system called Kirorank, which rated employees based on their activity with the Kiro AI tool. Employees began creating AI agents that performed pointless tasks solely to improve their scores. Amazon's management took the leaderboard down.
Meta discontinued its leaderboard in April, and Amazon followed a month later. Both companies discovered the same thing: measuring the value of AI by the number of tokens consumed is like measuring productivity by the number of hours spent in the office.
Uber burned through its annual AI budget in four months
Uber deployed Claude Code for its engineers in December 2025. By February 2026, use of the tool had more than doubled, from one-third of employees to more than four-fifths. In April, the company's chief technology officer, Praveen Neppalli Naga, admitted something that sounds like a warning to the entire industry: the entire planned annual AI budget was gone. Spent in four months.
Costs per engineer ranged from $500 to $2,000 per month for tokens alone. Uber has now imposed a monthly cap of $1,500 per employee for each coding tool. Company president Andrew Macdonald admitted on the Rapid Response podcast that he cannot connect all the spending on tokens to anything Uber customers actually see or experience.
Meta is building a monitoring platform
An internal Meta report described the situation as an “exponential increase" in AI consumption. According to the document, the company is on track to spend billions of dollars a year on AI usage. Yet individual employees and their teams have so far had virtually no visibility into how many tokens they consume. CTO Andrew Bosworth responded with his own internal statement: “No one should use AI tools just for the sake of using them. Not every movement is progress, and token consumption in itself does not measure value in any way."
Meta is now building a central dashboard called AI Gateway, which tracks usage and spending in real time. Starting in 2027, it plans to transition to a structured system with budgets, limits, and automatic alerts for unusual spending spikes. At the same time, the company is encouraging employees to move away from third-party tools, particularly Claude, and adopt its own MetaCode coding assistant.
Engineers are not the problem
One thing surprisingly turned the entire discussion on its head: programmers are not primarily responsible for the massive token consumption.
Accenture, the consulting giant, is struggling internally with how to stop non-engineering employees from spending the company's AI budget on trivial tasks, such as converting PDF documents into presentations. According to an audio recording from an internal meeting obtained by 404 Media, Justice Kwak, Accenture's head of agentic AI strategy, said: “The data shows us that our technical team is not driving token consumption. It is largely non-engineers."
Until recently, Accenture had been pressuring executives to start using AI or prepare to accept that they had no future at the company. The classic “use it, but not too much" is now becoming the standard corporate approach.
Costs no one expected
Why did all of this escalate so quickly? The answer lies in how AI tools charge their customers. OpenAI and Anthropic offer subscriptions ranging from $10 to $200 per month for individuals. But the real revenue comes from corporate customers such as Meta, Shopify, and Amazon, which pay not only for subscriptions but also for every token consumed by their tens of thousands of employees. The more tokens, the higher the bill.
And the models themselves are becoming more expensive. Anthropic's latest model, Fable, is twice as expensive as its predecessor, Opus. At the same time, many employees get used to using the most powerful model available for everything, regardless of whether the task actually requires it. The highest-spending companies now pay up to $7,500 per employee per month for AI. According to available data, deploying AI agents that run models repeatedly in automated loops has tripled corporate bills on average, even though the price per token itself has fallen.
Cut back, or slow productivity?
Not all companies are approaching the situation in the same way, however. Databricks has not yet imposed limits on its engineers. The company's management says its employees use AI effectively, so rationing makes no sense. This is precisely the tension facing the entire sector: limits control costs, but they may also hinder the very benefits that justified the deployment of AI in the first place.
The solution companies are gradually embracing is technical: instead of imposing blanket restrictions, move routine tasks to cheaper or open-source models and reserve expensive, powerful models for genuinely complex tasks. Microsoft and Databricks have already launched their own tools for monitoring and limiting spending. The startup Factory, backed by Nvidia among others and valued at more than $1.5 billion, has launched a system that automatically redirects simpler tasks to cheaper models.
One American technology executive told The Economist: “It will be an absolute nightmare." He was talking about a scenario in which each of hundreds of corporate software tools offers its own AI agent. Costs would then rise on their own, without anyone consciously ordering it.
Sources: techcrunch.com and theinformation.com
