Burning a year’s AI budget in four months isn’t a scandal. It’s Jevons paradox — and a discipline problem.
Uber just burned its entire 2026 AI budget in four months. Gone by April, mostly on Claude Code and Cursor. Engineers were running $500 to $2,000 a month each, 70% of committed code now originates from AI, and the CTO says they’re “back to the drawing board” on how to pay for it.
Here’s the part most people are misreading. The tools didn’t fail. They worked so well that usage exploded, and the bill went with it. That’s not a scandal — it’s Jevons paradox. Make something more efficient to use, and people use far more of it. Cheaper per task, bigger in total.
So the lesson isn’t “AI is too expensive.” It’s that we keep treating AI spend as a special category that gets to skip the basics.
Now, the fair pushback: a tool that didn’t exist last year has no adoption curve to forecast, and a cap set too early could have throttled the very gains worth paying for. True. But that’s the case for instrumenting the spend, not for flying blind.
Do this instead. AI doesn’t suspend the fundamentals of running technology: cost discipline, observability, unit economics. What does one unit of this cost? How does that scale as adoption climbs? Where’s the meter, and who’s watching it? You don’t need a perfect forecast to put a ceiling and an owner on the line. Right?
A tool too good to afford is a fine problem to have. Not seeing the bill coming is a different one.
