Uber's CTO Accidentally Spent $1,200 in Two Hours on Claude Code

This is the enterprise AI spending story I've been waiting for someone to tell honestly.

Yahoo Finance reports that Uber's aggressive rollout of Anthropic's Claude Code blew past internal budget expectations, with AI-related costs up 6x since 2024 despite a $3.4B R&D spend. CTO Sundeep Gupta accidentally burned $1,200 in a two-hour coding session. Not on some exotic fine-tuning job. On regular agentic coding.

The detail that made me pause was the leaderboard. Uber created internal rankings of engineers by AI tool usage. The intent was to encourage adoption. The effect was to create a perverse incentive where engineers competed to use AI more, not to use it better. When you measure consumption instead of output, consumption goes up. That shouldn't surprise anyone who's ever worked in a large organization, but it's apparently surprising enough that Uber's leadership is now "back to the drawing board" on AI spend strategy.

I use Claude Code every day. I know exactly how this happens. You start a complex refactor. The agent reads your codebase, proposes changes, runs tests, iterates. Each cycle burns tokens. An aggressive multi-turn session with extended thinking on Opus 4.6 or 4.7 can easily hit hundreds of dollars if you're not watching the meter. For a solo builder like me, that's self-limiting. I feel it in my wallet. For an enterprise with 10,000 engineers and a "use more AI" directive, there's no natural brake.

This is the first time a major enterprise has publicly admitted that AI coding tool costs are spiraling beyond projections. I suspect dozens of others are having the same conversation internally. The usage-based pricing model that works great for OpenAI and Anthropic's revenue creates genuinely unpredictable costs for buyers. When you can't tell your CFO what next quarter's AI bill will be, that's a procurement problem, not a technology problem.

The RTK story above becomes much more interesting in this context. If a single Rust binary can cut token costs 60-90%, that's not a nice-to-have for enterprises. It's a finance team requirement.

What to do now: If you run an engineering team using AI coding tools, track cost per engineer per week, not just total spend. Set per-session budgets. And look hard at RTK or similar compression tools before your CFO has the same conversation Uber's is having now.