Somewhere at the headquarters of a large corporation, a CFO is sitting and staring at an invoice. It shows a figure that could easily be the annual IT budget of a medium-sized country. More than half a billion dollars. For one month. For artificial intelligence. And most importantly, without anyone setting usage limits for employees. It is a symptom of a trend known as tokenmaxxing, which is now, halfway through the year, beginning to come at a very painful cost.
Over the past two years, companies around the world have been racing to get their employees to use AI tools as much as possible. The result has been astronomical bills, no clear impact on productivity, and growing frustration on all sides.
How tokenmaxxing emerged
Tokens are the basic units in which large language models process text. Roughly one token corresponds to one and a half words in English. And AI companies such as Anthropic and OpenAI charge based on token usage. With that in mind, the logic of tokenmaxxing makes sense at first glance. A company wants to know which employees are actually using AI. How can it measure that? Simple: whoever consumes the most tokens must be using AI the most. And whoever uses AI the most must surely be the most productive. Or are they?
Meta, for example, went full steam ahead in this direction. An employee there built an unofficial leaderboard called Claudeonomics, which ranked all approximately 85,000 employees by token usage. The top performers received titles such as "Token Legend" or "Session Immortal." During a single 30-day window, total usage on this leaderboard exceeded 60 trillion tokens. One user alone burned through 281 billion tokens during that period. At Anthropic's standard prices, that usage would have cost more than $900 million.
Amazon took a similar path. An internal leaderboard called Kirorank tracked employee activity on the Kiro development platform and rewarded those who used AI the most.
Goodhart's law took care of the rest
This is where a principle philosophers and economists understood long before anyone coined the word tokenmaxxing comes into play. Goodhart's law states: when a measure becomes a target, it ceases to be a good measure.
In practice, it looked like this. Amazon employees began running AI agents on pointless tasks simply to boost their scores. One engineer admitted that whenever a project manager annoys him, he dumps the entire Slack conversation history into the internal AI tool and has ten sub-agents search through it for ways to ridicule the manager. The leaderboard ranking goes up, productivity stays where it was, and the bill grows alarmingly.
Amazon's senior executives noticed. Dave Treadwell, vice president of engineering, sent employees a message: "Please do not use AI just for the sake of using it. Use it to solve customer problems, company problems, and to innovate."
Uber burned through its annual budget in four months
The same problem exists at Uber. Chief Technology Officer Praveen Neppalli Naga admitted that the company had exhausted its entire annual budget for AI coding tools by April. This was despite the share of engineers using Claude Code jumping from 32 to 84 percent between February and March 2026.
COO Andrew Macdonald then publicly expressed doubts. He said he was unable to draw a direct line between rising token usage and what the company was actually delivering to users. "The connection simply isn't there," he said, admitting that AI spending was becoming increasingly difficult to justify. Uber's average monthly cost for AI tools per engineer ranged between $150 and $250. For the most active users, it reached as much as $2,000 per month. With approximately 5,000 engineers, the costs are rising quickly.
Uber is far from alone. Microsoft canceled most of its internal Claude Code licenses, primarily because of the cost. It redirected employees to its own GitHub Copilot. Salesforce is facing a bill from Anthropic exceeding $300 million for 2026. Meta removed its Claudeonomics leaderboard shortly after reports about it began to appear. CTO Andrew Bosworth wrote to employees in no uncertain terms: "No one should use AI tools just for the sake of using them."
So what went wrong?
Sophia Velastegui, CEO of Velastegui Ventures and former head of AI at Microsoft, identifies one of the fundamental problems. According to her, most employees automate tasks they do not enjoy rather than those that are most valuable to the company. Handing out licenses blindly and waiting to see what comes of it does not produce results.
Research by Jellyfish provided the numbers. Heavy Claude Code users consumed approximately ten times more tokens than their moderately active colleagues. Yet their productivity increased by only about twofold. In other words, for every additional dollar spent on AI, companies received significantly less than they expected. The reality of AI today is that it works reliably mainly for coding. Everywhere else, the outcome is far less predictable. And it is precisely this gap that is driving IT bills upward without delivering a corresponding return.
An anonymous case reported by Axios is perhaps the clearest illustration of where a complete absence of rules can lead. An AI consultant described a client that spent more than half a billion dollars on Claude in a single month because it simply forgot to set any usage limits for employees. Anthropic offers enterprise customers tools for managing access, tracking usage, and setting caps. But someone has to actively configure these features. And in this case, no one did.
This is not the only such story. One CTO mentioned that employees were using the company's AI subscription to check the weather. Yet corporate AI plans are not unlimited "all-you-can-eat" packages. Even a simple chatbot query uses tokens and incurs costs. And when this adds up across thousands of employees, the numbers rise quickly.
A return to outcomes instead of usage
Amazon responded by replacing its focus on the number of tokens consumed with a metric called "normalized deployment," meaning how much AI-assisted code actually makes it into production. Not how many tokens someone burned, but what those tokens produced.
Meanwhile, no one is stopping investment in AI infrastructure. Amazon is planning approximately $200 billion in capital expenditures for 2026, primarily on AI and data centers. Jefferies points out that spending by US companies on information technology and software reached a record 4.91 percent of GDP in the first quarter of 2026, surpassing the previous peak from the dot-com mania of 2000.
At the same time, banks are warning that a significant amount of capital will be destroyed in this cycle, just as it was in the late 1990s. They also invoke Amara's law: people tend to overestimate the short-term impact of technology and underestimate its long-term impact.
For now, it appears that companies are finally beginning to ask the right question. Not how many tokens did we burn, but what did we get for them.
Sources: businessinsider.com, theconversation.com and finance.yahoo.com
