In December 2025, Uber rolled out an AI coding tool to its engineering team.
By April 2026, the company's chief technology officer made a quiet admission. The entire annual AI budget was gone. Not most of it. All of it. A twelve-month allocation, spent in four.
The mechanism was not a procurement failure. It was a leaderboard.
Uber built internal rankings that tracked how much AI tooling each engineering team consumed. The more you used, the higher you ranked. Adoption jumped from 32% to 84% of the company's engineers in three months. Individual engineers were generating between $500 and $2,000 per month in AI usage costs on top of their salaries. The company built a system that rewarded spending, and the system worked exactly as designed.
Meta ran the same experiment at larger scale.
An internal leaderboard called Claudonomics tracked AI token consumption across more than 85,000 employees. It ranked the top 250 power users and awarded gamified titles: token legend for the highest consumers, session immortal for extraordinary session duration. In a single thirty-day period, Meta employees collectively consumed sixty trillion tokens. The top individual user averaged 9.36 billion tokens per day. Some employees were leaving AI agents running idle for hours, doing nothing, to inflate their position on the leaderboard.
The leaderboard was taken down after press coverage. Meta's CTO had publicly endorsed the underlying logic before that.
These are not isolated cases. They are the same incentive structure appearing in different buildings.
The unit cost of running an AI model is falling fast. A recent Gartner analysis found that by 2030, the cost of running a frontier AI model will drop by over 90% compared to 2025. That sounds like good news. It is not, because AI agents, the kind that operate autonomously across multiple steps rather than answering a single question, require between five and thirty times more tokens per task than a standard chatbot. The unit price drops. The number of units explodes. Goldman Sachs projects a twenty-four-fold increase in token consumption by 2030. The economist William Stanley Jevons identified this pattern in the nineteenth century. When a resource becomes cheaper to use, consumption does not fall. It accelerates.
The AI industry is not solving a cost problem. It is reorganizing one.
Microsoft ran a six-month experiment in which its own engineers chose Anthropic's Claude Code over its own GitHub Copilot CLI. In May, Microsoft began cancelling Claude Code licenses across key divisions. The official reason was toolchain unification. The timing, the last day of the fiscal year, suggests something more practical. Microsoft consolidated onto the platform it owns, at the exact moment that platform switched to usage-based billing. The cost problem did not go away. It moved inside the Microsoft ecosystem.
Nvidia's chief executive said in March 2026 that he would be deeply concerned if a $500,000 engineer at his company was not consuming at least $250,000 in AI tokens. Nvidia's vice president of applied deep learning said publicly that for his team, the cost of compute now exceeds the cost of the employees.
Both statements are accurate. They reflect different incentive structures within the same company. The CEO sells the chips that process every token. More consumption means more revenue. The executive managing an engineering budget measures whether the output justifies the input. Neither is wrong. The problem is that the signal to consume originates from the people who profit most from consumption, and it cascades downward through the industry until it becomes a leaderboard rewarding employees for running agents idle.
The combined capital expenditure for 2026 from Amazon, Microsoft, Alphabet, and Meta is approaching $740 billion, a 69% increase from 2025. There have been over 92,000 tech layoffs in 2026 so far.
The spending and the headcount reduction are not separate stories.
For builders developing products for enterprise clients, the current moment has a specific shape. The companies buying AI tools are not asking whether consumption produces proportional value. They are asking whether they are consuming enough. That is a procurement environment with a known failure mode, and the failure mode is already visible in Uber's quarterly numbers.
The question worth asking before the next enterprise sales conversation is not whether the client has an AI budget. It is whether they have a measurement framework. Most do not. The ones who build one next will be the ones who can tell the difference between a leaderboard and a strategy.
404 Found covers AI developments from a European Insider, three times a week.
