Uber CTO Praveen Neppalli Naga has confirmed to The Information that the company burned through its entire 2026 AI coding budget in just four months. The cause: Claude Code adoption that jumped from 32% to 84% of Uber's 5,000-strong engineering organisation between December and March.

"I'm back to the drawing board because the budget I thought I would need is blown away already," Naga told The Information.

The numbers are staggering. Individual engineer costs ranged from $500 to $2,000 per month. Naga himself blew through $1,200 in a single two-hour demo session. About 70% of all committed code at Uber now comes from AI, and roughly 11% of live backend updates are written by AI agents with zero human in the loop.

The Token Billing Problem Nobody Budgeted For

This is not a story about a failed product. Claude Code worked exactly as intended. Engineers loved it. Adoption was organic and aggressive. The problem is that token-based billing operates on fundamentally different economics than traditional per-seat software licensing.

A developer running a simple autocomplete uses negligible tokens. A developer running Claude Code as an autonomous agent across a monorepo, instructing it to refactor an entire API layer and generate tests in parallel, can burn thousands of dollars in a single afternoon. Scale that across 5,000 engineers, many running multiple agent loops simultaneously, and the annual budget disappears by April.

Uber's R&D budget is $3.4 billion annually, so the overrun is manageable for a company that size. But the signal matters more than the dollar amount. If Uber, with dedicated FinOps teams and a sophisticated engineering infrastructure, could not forecast the cost trajectory of agentic coding tools, smaller organisations have no chance.

The Cloud Cost Parallel

The closest analogue is the early AWS era. In 2010, engineering teams started provisioning cloud compute with the same mental models they used for on-premise servers: a fixed capital expenditure with predictable annual costs. Then usage-based billing hit, and companies spent years learning to manage cloud cost sprawl. An entire industry of FinOps tooling emerged to solve it.

AI token spending is about to follow the same path, but faster and with higher variance. Cloud compute costs scale roughly linearly with usage. Token costs scale with the complexity and scope of what an agent is asked to do, which means a single engineer's monthly cost can swing from $50 to $5,000 based on the nature of their work.

What This Means for Every CFO With Engineers on Payroll

The uncomfortable truth: there is no established playbook for AI token cost management. Enterprise budgeting cycles are built around predictable per-unit costs. Agentic coding tools break that model completely. Every finance team that approved a 2026 AI tooling budget based on per-seat pricing assumptions is sitting on the same ticking clock Uber just exposed.

Anthropic, the company behind Claude Code, has not commented on Uber's budget overrun specifically. But the dynamics are structural, not unique to any one vendor. OpenAI's Codex, Google's Gemini Code Assist, and every other agentic coding tool that bills on token consumption will produce the same cost surprises at scale.

Uber's story is not a cautionary tale about Claude Code being too expensive. It is a signal that enterprise AI has entered its cloud cost reckoning moment, and most companies have not built the financial infrastructure to handle it.