Uber's chief technology officer told The Information that the company exhausted its full-year 2026 AI budget by April, four months in, after Anthropic's Claude Code spread through its 5,000-engineer organization faster than anyone modeling the spend had anticipated. Per-engineer costs are running $500 to $2,000 a month. Roughly 70% of committed code at Uber now comes from AI, and 95% of engineers use the tools at least once a month.
What actually happened
Uber rolled out Claude Code to its engineering organization in December 2025. Usage doubled by February. By April, the bill had eaten the year's allocation. "I'm back to the drawing board because the budget I thought I would need is blown away already," said CTO Praveen Neppalli Naga, which is the kind of admission you don't usually hear from a company spending $3.4 billion a year on R&D.
Internal mechanics matter here. Uber, like a few other big shops, ran leaderboards ranking engineers on AI tool usage. That works fine when adoption is the bottleneck. It works less well when the tool is metered per token and a single overnight agent run can consume the equivalent of 200 casual users' worth of inference. Claude Code is not a per-seat license. It is a meter that engineers competing on visibility dashboards have no obvious reason to keep low.
Nvidia, of all places, agreed
Days later, Axios reported Nvidia VP of applied deep learning Bryan Catanzaro saying that for his team the cost of compute had passed the cost of his employees. "For my team, the cost of compute is far beyond the costs of the employees" was the exact line. Coming from the company selling the GPUs underneath the whole boom, that lands differently than it would from a Claude customer.
Set this against the layoff narrative. The tech industry cut roughly 80,000 jobs in Q1 2026, with about half officially attributed to AI replacement, per Nikkei Asia data. The economic case for those replacements assumes the AI doing the replacing costs less than the people getting replaced. At $500 to $2,000 a month per engineer, that math holds for a junior contractor offshore and almost nobody else.
Some founders are bragging about it
Not everyone is panicking. Swan AI CEO Amos Bar-Joseph posted a screenshot of a $113,421.87 Anthropic invoice in a LinkedIn post last month, calling it the proudest invoice of his career. His company has four people. The team optimizes around $10 million in ARR per employee, and the latest bill was more than double the previous month's. The unit economics work as long as model prices keep falling and revenue keeps compounding. Both can stop happening at the same time.
What's next
Anthropic quietly doubled its own published cost estimates on April 15. Business Insider reported the Claude Code documentation now puts the average enterprise developer at $13 per active day, up from $6, with 90% of users staying below $30 per day, up from $12. A spokesperson said this wasn't a pricing change, just a more accurate reflection of how costs scaled once Opus 4.7 became the default model. That distinction is technically true and practically meaningless if you're a finance team running a 2026 plan built on the older numbers.
Uber hasn't said whether it will throttle Claude Code usage for the rest of the year or eat the overrun. Naga is testing OpenAI's Codex as part of his replanning. The next data point lands when other large engineering organizations file Q2 numbers and their CFOs start asking about line items that didn't exist on the planning template.
