In December 2025, Uber handed its engineers access to Claude Code, Anthropic's AI coding agent. By April 2026 - four months later - the company had spent its entire annual AI-tools budget. Not overshot by a little. Gone.
The story was reported by Forbes, Fortune and TechCrunch. It is the clearest public example yet of a problem quietly hitting companies of every size: AI costs that compound far faster than anyone is watching.
What actually happened
Uber rolled Claude Code out to roughly 5,000 engineers, and adoption did not creep - it jumped. According to the reporting, the share of engineers using it climbed from 32% in February to 84% by March.
- The average engineer ran $150-250 a month in tokens.
- Power users ran $500-2,000 a month.
- Multiply that across thousands of engineers over four months, and the full-year number was already spent.
Uber even ranked engineers on internal leaderboards by their Claude Code usage - which quietly turned burning tokens into a status game. Following the blowout, the company capped spending at $1,500 a month per employee, per AI coding tool.
Why AI costs compound (and headcount doesn't)
A chatbot is cheap to reason about: one question, one answer, one charge. An agent is different. It runs a loop - reads files, calls tools, makes edits, re-checks its own work - and it re-sends the entire accumulated context on every single step. By step 20 you are paying for the same instructions and history twenty times over.
That is why agentic tools consume 5-30x the tokens of a simple chat. One industry analysis put it starkly: a linear 2023-style task cost about $0.04 per run; the 2026 agentic equivalent runs closer to $1.20 - roughly 30x more for the same-looking task.
The part everyone misses: this wasn't really about Uber
Here is the counterintuitive bit. Token prices actually fell - by around 98% over two years. And yet enterprise AI bills tripled over that same stretch, because consumption grew far faster than prices dropped. Cheaper tokens simply got used a lot more.
Gartner now projects that AI coding costs could rival what companies pay the developers themselves. Uber is just big enough to make the headline - the same curve is hitting startups and mid-size teams right now. They tend to find out on the invoice.
What would have caught it
The failure at Uber wasn't using AI. It was flying blind. No per-engineer limits, no live attribution, no alert when the run rate went vertical. Three things turn that around:
- See it live. Spend per person, per team, per tool - updated daily, not discovered a quarter later.
- Attribute it. Know which projects and people drive the bill, so a runaway agent shows up in hours, not months.
- Cap it. Sensible per-person limits (Uber landed on $1,500) that stop one bad weekend from eating a quarter.
None of this means using AI less. Uber's engineers were getting real value from Claude Code; so are yours. The goal is simply to see the meter while it is running - which is the one thing almost nobody had in place when the bills came due.