Uber burned through its entire 2026 AI budget in four months. The company’s R&D line ran $3.4B in 2025. The AI tooling slice of that budget, which went mostly to Claude Code and Cursor for 5,000 engineers, was gone by April. Their CTO, Praveen Neppalli Naga, went back to square one telling The Information: “I’m back to the drawing board, because the budget I thought I would need is blown away already.”
Two weeks later Microsoft pulled internal Claude Code licenses across the division that builds Windows, Office, Teams and Surface. The tool was too popular, with MS engineers preferring it to Microsoft’s own Copilot, and that was the problem. Usage spiraled beyond all expectations.
Two of the best-equipped companies in the world to absorb a surprise AI cost got surprised big time. An estimated quarter-billion dollars spent in four months.
The mechanics here are finally getting some scrutiny. Most of these tools are priced per token rather than per seat, which means an engineer running autocomplete costs a few dollars a month while an engineer running parallel agents against a large codebase can run thousands in the same period. To be clear, those engineers are usually getting real value for the spend, and a developer who knows how to use these tools well is materially more productive than one who doesn’t.
The problem isn’t whether the tools work. It’s that finance plans software spend on per-license assumptions, and that model just doesn’t hold when usage and cost move together that aggressively. Without caps, monitoring, or someone whose job it is to own the number, the bill at the end of the month is whatever engineers decided to do during it.
Uber didn’t have caps. No token monitoring, no spend owner at the team level, none of the discipline that IT and finance teams have spent fifteen years building around cloud. AI tooling was rolled out as if it were Office 365, and it isn’t.
If companies with multibillion-dollar R&D budgets get blindsided, a company between $200M and $2B in revenue is in a tougher spot, with less room to absorb the surprise when it shows up. The Uber and Microsoft numbers aren’t just their stories. Most companies have a version of it on a smaller scale and need to start looking for it now if they haven’t already.
The three questions I’d put to the CFO and the CIO this quarter are straightforward. What is the monthly AI tooling spend, broken out by team and by tool? Who owns the cap? What happens if usage doubles next quarter? Because at the current rate of adoption inside engineering, it will. If any of these takes more than a day to answer, that’s the trigger for the next joint Finance/IT project.
