← all posts

uber burned its entire 2026 AI coding budget in four months. the $3.4B number everyone's spreading is wrong.

Empty corporate war room with blank budget screens and a calendar circled in red
Image: fortune.com

Uber’s CTO Praveen Neppalli Naga had a budget for AI coding tools in 2026. By April — four months in — it was gone. “I’m back to the drawing board because the budget I thought I would need is blown away already,” he told The Information.

Here’s what actually happened: Uber rolled out Anthropic’s Claude Code to its engineering org in December 2025. Adoption went from 32% of engineers in February to 84% classified as agentic users by March. By spring, 95% of Uber engineers were using AI tools monthly, ~70% of committed code came from those tools, and about 11% of live backend updates were written by agents with zero humans in the loop.

Monthly costs per engineer ranged from $500 to $2,000 for power users. Uber also ranked engineers on internal leaderboards by Claude Code usage — essentially paying people in social status to burn more tokens. Predictably, they burned more tokens.

Now, the fact-check the internet skipped: that $3.4B figure everyone’s citing as “Uber’s AI budget”? That’s Uber’s total R&D spend for 2025, up 9% year over year. The actual AI coding tools budget that got torched was a subset of that — its real dollar amount hasn’t been publicly disclosed. The viral framing inflated the number and misattributed it. The story is still wild without the exaggeration.

Satya Nadella’s name got pulled in too. In a June 21 WSJ interview — and a June 14 essay on X titled “A frontier without an ecosystem is not stable” — he argued that concentration of AI value in a few models won’t survive political scrutiny. Multiple outlets connected his argument to the Uber case as a concrete illustration. He didn’t originate the $3.4B claim.

Uber’s fix, per TechCrunch: a hard $1,500/month cap per employee per agentic coding tool.

The actual lesson here isn’t the number. It’s the pricing model: metered by token, the more useful the tool becomes, the more it costs. Enterprise finance teams built budgets assuming linear adoption curves. Claude Code gave them an exponential one. Every company rolling out agentic tools right now is one leaderboard away from the same call.