Uber Exhausts 2026 AI Budget in Four Months

Uber burned through its entire 2026 AI budget in just four months due to the massive popularity of Claude Code among its engineers. With 95% of staff using AI tools, monthly costs per developer have reached as high as $2,000. The company is now forced to re-evaluate its R&D spending to sustain this high level of AI-driven productivity.

Key Points

• Uber depleted its full 2026 AI budget by April due to high API costs from Claude Code and Cursor.
• Engineering adoption reached 95%, with AI tools now generating 70% of all committed code at the company.
• Individual API costs for the tools are unexpectedly high, reaching up to $2,000 per engineer per month.
• Claude Code has become the dominant tool in Uber's workflow, while usage of the competitor Cursor has plateaued.
• The company is forced to rethink its AI financial strategy as productivity gains outpace original budget forecasts.

Sentiment

The community is deeply ambivalent. While most commenters acknowledge AI coding tools provide real value, there is widespread skepticism about the efficiency of current corporate adoption patterns. The dominant sentiment is that the technology works but is being deployed with poor incentive structures, wasteful practices, and insufficient oversight. Many commenters are concerned about perverse corporate mandates around token usage and the quality implications of unchecked AI-generated code.

In Agreement

• Token costs are trivial compared to engineer salaries, so even high spending is justified if it produces meaningful productivity gains
• AI coding tools enable individual developers to accomplish work that would previously require entire teams, with one commenter reporting building a Postgres-compatible database in Rust in four weeks
• Power users running multiple parallel agent sessions can legitimately burn through large token budgets on real, valuable work like API integrations and complex system development
• The speed of delivery enabled by AI tools has intrinsic value beyond raw cost comparisons, as getting to market sooner matters
• Companies are right to encourage broad experimentation with AI tools as an R&D investment, even if some spending is wasteful during the exploration phase

Opposed

• Most high token spend does not translate to proportional business value, with much of it being wasteful loops, excessive context windows, and agents spinning unattended
• Corporate tokenmaxxing leaderboards and minimum usage mandates create perverse incentives that reward spending over productivity, a classic Goodhart's Law failure
• Engineers who fully offload coding to AI agents lose the ability to understand and maintain their own code, creating hidden technical debt
• The massive volume of AI-generated code is producing 'slop' with declining quality in production, and human brains cannot properly review thousands of lines of machine-generated changes
• Non-technical staff using AI to produce code creates fragile systems that inevitably break and burden engineering teams with cleanup
• The pipeline for training junior engineers into senior engineers is being destroyed, creating a looming talent crisis