Uber CTO Praveen Neppalli Naga confirmed to The Information that the company has already exhausted its planned 2026 AI budget, driven by token-based consumption from Claude Code spreading across 5,000 engineers faster than any budget model could anticipate, and the story it tells about AI as a new class of enterprise cost is one that every CFO with engineers on their payroll needs to understand.
Uber rolled out Claude Code access to its full engineering organisation in December 2025. By February, adoption had jumped from 32% to 63% of engineers using it monthly. By March, 84% were classified as agentic coding users. The tool did not fail. It did not underdeliver on its stated capabilities. Engineers loved it, used it constantly, and put it to work on exactly the kind of tasks it was designed for: parallel agent execution, large-scale codebase refactoring, automated testing, backend code generation. About 70% of committed code at Uber now comes from AI, and roughly 11% of live backend updates are written by AI agents without any human in the loop. From a product and productivity standpoint, the rollout was a success. From a finance standpoint, the budget was incinerated. Naga was direct: "I'm back to the drawing board because the budget I thought I would need is blown away already."
The cost mechanics are what make this story different from a simple overspend. Claude Code is not priced on a per-seat basis the way a traditional enterprise software licence works. It runs on token consumption, meaning the invoice is a function of how many tokens the model processes across all engineer sessions. A developer running a single autocomplete suggestion at the end of a function is consuming a negligible token budget. A developer running Claude Code as an autonomous agent across a monorepo, instructing it to refactor an API layer and generate the associated tests in parallel, can consume thousands of dollars worth of tokens in a single afternoon session. Scale that across 5,000 engineers, many of them running multiple agent loops simultaneously, and the math compounds in ways that no annual software budget cycle was built to absorb. Reported individual engineer costs ranged from $500 to $2,000 per month. Naga himself spent $1,200 in two hours during a personal demo session. Those are not edge cases. They are the natural consequence of using agentic AI tools the way they are meant to be used.
The comparison to cloud cost is the most useful frame for understanding where this is going. In 2010, enterprise software teams started provisioning AWS compute with the same mental model they had used for on-premise servers: a capital expenditure, planned in advance, predictable. AWS bills arrived and were triple what finance had modelled. The pattern repeated in every organisation that adopted cloud at scale. A decade of FinOps tooling, reserved instance strategies, tagging frameworks, and cost anomaly alerts was built to correct for that initial miscalibration. The AI coding cost problem is structurally similar. A usage-based pricing model has been placed in front of a highly motivated user base with a direct incentive to consume as much of it as possible. The tooling to monitor, cap, and allocate that spend at the individual team or engineer level does not yet exist at the maturity of cloud cost tooling. Uber is not unusual for having this problem. It is the first large company to surface it publicly at this level of specificity, which makes Naga's disclosure more valuable than it might appear.
The internal leaderboard detail in Uber's rollout is worth examining separately. The company tracked and ranked engineer usage of Claude Code on internal performance visibility dashboards. That is a management choice designed to drive adoption, and it worked precisely as intended. It also created a cultural dynamic where using more AI tooling was visibly rewarded, and using less was implicitly under-performing. In a token-based billing environment, that incentive structure directly translates into budget acceleration. Engineers competing on leaderboards for AI usage have no obvious reason to be conservative about consumption. The people who designed the leaderboard were almost certainly not the same people responsible for the AI services budget line. That organisational gap, between the teams driving adoption and the teams managing spend, is the root cause of the overrun more than any pricing quirk of Claude Code itself.
The responses large enterprises will take from this situation are already visible in Naga's comments. He mentioned that Uber intends to give engineers access to OpenAI's Codex in the future, suggesting a multi-vendor strategy rather than a single-provider lock-in. That choice is likely motivated partly by competitive pricing leverage and partly by risk diversification. Companies watching Uber's experience will also accelerate conversations with Anthropic, OpenAI, and the other major providers about enterprise framework agreements that replace token-based billing with committed spend deals at negotiated rates. Microsoft has already done this for Copilot: a flat per-seat model that limits the upside for the vendor but gives enterprise finance teams the predictability they need to budget reliably. As Claude Code usage scales across the industry, Anthropic will face increasing pressure from enterprise procurement teams to offer similar structures, or watch finance departments cap usage and throttle the adoption that is driving their revenue growth.
The third option, building internal coding agents on top of open-weight models, is gaining credibility precisely because of stories like this one. A company running Qwen3.6-27B locally on dedicated GPU hardware has predictable per-query costs that are a function of hardware depreciation and electricity, not per-token billing. The setup cost is higher. The ongoing cost is bounded. For organisations with 5,000-plus engineers generating multi-hundred-dollar monthly AI bills per head, the build-versus-buy calculation is no longer theoretical. Uber burned its budget in four months. The next company to do the same will have more options for what happens in month five.
","excerpt":"Uber CTO Praveen Neppalli Naga confirmed the company burned through its entire 2026 AI budget in four months, driven by Claude Code adoption that jumped from 32% to 84% of its 5,000-engineer organisation.
Also read: ChatGPT Got Obsessed With Goblins and OpenAI's Explanation Is More Unsettling Than the Bug Itself • The Tooling Problem in Local AI Is Finally Getting Solved and That Matters as Much as the Models • Tech Giants Are Spending $725 Billion on AI in 2026 and 92,000 Workers Are Paying for It
