The wrong million tokens
Executive Take
Spending caps treat a measurement failure as a budget problem and leave your best engineers rationing tools instead of fixing the pricing gap. Build cost per task into your AI scoreboard now or you will cap spending blind like Uber did.
Executive Summary
Uber burned its full 2026 AI coding budget in four months after Claude Code adoption rose from 32% to 84% of engineers, then capped spending at $1,500 per employee per tool monthly. The piece argues the real failure was never measuring cost per task, not token prices, and shows prompt caching and smarter retrieval can cut costs while improving accuracy.
Why It Matters
Technology and finance leaders need this because AI coding spend can spike unpredictably without warning, as Uber's case shows. HR leaders should note that blanket caps push rationing decisions onto top engineers rather than fixing the underlying cost model.
Bizquad Perspective
Most leaders will copy Uber's spending cap as a fix, missing that caps are an admission of failed measurement, not a solution, and that fixing retrieval is cheaper and makes AI more accurate, not less.