Uber Exhausted Its Entire 2026 AI Budget in Four Months — Then Capped Claude Code

Uber set its 2026 AI budget in 2025. By April 2026, it was gone. The culprit wasn’t waste — it was the unexpected token density of agentic coding tools.

What Happened

Uber is capping every employee’s monthly token spending on AI coding software — tools like Cursor and Claude Code — at $1,500 per tool, with separate budgets per platform rather than a shared pool. The cap translates to roughly $36,000 per engineer annually across two tools, about 11% of median Uber software engineer compensation.

The immediate cause was simple: budgets set in 2025 assumed usage patterns that predated the surge in token-intensive coding agents. When engineers adopted agentic tools in earnest, actual consumption outpaced projections by a wide enough margin to exhaust the full annual allocation within four months.

Why It Matters

The cap isn’t a retreat from AI adoption. It’s the first generation of enterprise-scale AI cost governance. This is what mature AI adoption looks like: usage policies, per-person budgets, and the acknowledgement that uncapped agentic tool access is not a sustainable default.

The critical dynamic at play is enterprise vs. individual pricing. Individual developer plans are often subsidized to drive adoption. Enterprise contracts at Uber’s scale face a fundamentally different cost structure — one where monthly per-engineer spend at $1,500 is a deliberate ceiling, not a floor.

Two things made this inevitable. First, coding agents are uniquely token-hungry — agentic loops consume far more than conversational use. A developer running autonomous refactors, test generation, and multi-file edits generates orders of magnitude more tokens than a developer writing prompts manually. Second, enterprise AI budgets in 2025 were largely estimates built on 2024 usage data, which did not include production-grade coding agents.

What This Means for Builders

Any engineering organization planning 2027 AI budgets should treat Uber’s numbers as a calibration point. $1,500/engineer/month per agentic coding tool is a ceiling that a hyperscaler with cost pressure set after hitting actual limits — not a theoretical maximum.

For platform teams evaluating AI tooling rollouts: model usage governance should be part of the rollout plan, not a retrofit. The tools to watch are those with the highest token density per task — autonomous agents, long-context file analysis, and multi-step refactors all compound spend faster than conversational tools.

Source: Bloomberg via Simon Willison

Content created with AI assistance and reviewed for accuracy.