Uber burned through its entire $3.4 billion AI R&D budget for 2026 in a few months. CTO Praveen Neppalli Naga admitted the company is "back to the drawing board." The cause was engineers going hard on AI coding tools like Anthropic's Claude Code and Cursor, pushed along by internal leaderboards that tracked who used them most. About 11% of Uber's backend code updates, including systems that handle ride-matching and pricing, are now written by AI agents. It happened fast and it wasn't cheap.
Those leaderboards created what critics call "token maxxing," developers chasing usage volume instead of actual value. And the company isn't tracking whether AI-generated code is any good or what it'll cost to maintain. Quality got lost in the volume race. Meanwhile, heavy spending on internal engineering tools like Cursor drove up costs with little to show where customers would notice.
The blowout also reveals a structural problem. Cloud giants like Amazon and Microsoft run internal AI tools through their own infrastructure at cost. Uber pays third-party vendors at market rates and never controlled the unit economics. Now they're testing OpenAI's Codex. But swapping one pricey vendor for another doesn't solve the core problem.