Taxi
Pixabay
Uber exhausted its entire 2026 artificial intelligence budget by April, four months into the calendar year, after Anthropic's Claude Code spread across roughly 5,000 engineers faster than the company's finance models had anticipated.
Chief Technology Officer Praveen Neppalli Naga confirmed the overrun to The Information, saying the company was back to the drawing board on its assumptions.
Uber’s total research and development spend reached $3.4 billion in 2025, up 9% year over year, which makes the budget collapse less about scale and more about a pricing model that enterprise finance teams have not learned how to manage.
The disclosure landed alongside a structural shift from Anthropic itself. On May 13, the company announced that paying Claude subscribers would soon face a separate monthly credit meter for agent tools and third-party harnesses, billed at full application programming interface rates starting June 15. Read together, the two events describe a single problem. Token-based consumption pricing does not behave like the software line items chief financial officers know how to model, and the gap between what engineers consume and what finance teams expect is no longer hypothetical.
Uber rolled out Claude Code to its engineering organization in December 2025. Adoption climbed from 32% of engineers in February to 84% classified as agentic coding users by March. By spring, 95% of Uber engineers used artificial intelligence tools monthly, and roughly 70% of committed code originated from those tools. About 11% of live backend updates were written by agents with no human in the loop, according to Uber's own disclosures.
The numbers behind the spend are what make the story instructive rather than anecdotal. Monthly cost per engineer ranged from $150 to $250 on average, with power users running between $500 and $2,000. Naga himself reported spending $1,200 in a two-hour session during a personal demo. The tool did not fail, and engineers did not misuse it. They used it for exactly the workloads it was designed to handle — parallel agent execution, large-scale codebase refactoring, automated test generation and backend code production. From a productivity standpoint the rollout was a success. From a finance standpoint it was a runaway.
Uber compounded the dynamic by ranking engineers on internal leaderboards based on Claude Code usage. That created a cultural incentive to consume more tokens, which translated directly into faster budget burn. The teams driving adoption were not the same teams managing the spend, and that organizational gap turned out to be the load-bearing flaw.
Claude Code does not price on a per-seat basis. It meters tokens consumed across model calls, which means an engineer running autocomplete suggestions consumes a fraction of what an engineer orchestrating parallel agents across a monorepo will consume. The same tool, the same engineer, the same workday, can produce wildly different invoices depending on workflow choice. Annual budget cycles built around predictable per-license costs cannot absorb that variance.
Microsoft has taken the opposite approach with Microsoft 365 Copilot Enterprise, which sells at $30 per user per month with an annual commitment. The price caps the vendor's upside and gives finance teams a flat line item they can multiply by headcount. Anthropic's consumption model gives the vendor unlimited upside on heavy users and gives finance teams almost no forward visibility. Both models are defensible, and neither is right for every workload, but treating them as interchangeable in a planning cycle is what produced Uber's outcome.
GitHub is moving Copilot to a credit-based system on June 1, and analysts cited by InfoWorld expect most vendors to introduce separate consumption pools for agents and tool use over the next 12 to 24 months. The vocabulary will vary, credits, requests, messages or compute units, but the direction is set. Flat-rate inference for unbounded agentic workloads was never going to survive the math, and Anthropic's May announcement is the first major confirmation that vendors will pass the cost mechanics through to buyers rather than absorb them.
The industry's standard response to consumption-cost stories is that artificial intelligence pays for itself in productivity gains. Uber's case complicates that argument. The marginal productivity gain from a senior engineer running agentic workflows has to clear a much higher token-cost hurdle than the gain from an engineer running autocomplete. Five-to-twenty-fold increases in per-developer consumption are now documented in agentic mode, and no public benchmark shows a matching multiplier on output value. Productivity savings also do not show up in the same line item as artificial intelligence cost, which means finance teams cannot net them out inside a quarterly review.
There are also operational limits that make the simple cost-versus-output framing incomplete. Only 43% of organizations have formal artificial intelligence governance policies, according to survey data cited in coverage of the Uber overrun, and only 21% have mature agentic governance. Most enterprises do not yet apply to artificial intelligence tooling the spending controls that DevOps teams routinely apply to cloud compute. That includes per-engineer caps, real-time monitoring of token consumption and budgetary alerts before overrun rather than after. Uber deployed Claude Code organization-wide without those controls, and the result was visible within a quarter.
The Uber experience produces a short list of practical implications for finance leaders watching their own engineering organizations adopt agentic coding tools. The first is that pilot economics do not predict scale economics for consumption-priced tools, because pilots run on a few engineers using autocomplete while production runs on whole teams delegating multistep workflows to agents. The second is that incentive structures matter as much as pricing. Leaderboards and adoption targets drive token consumption, and any rollout that rewards usage without capping it should be modeled as an unbounded liability until proven otherwise.
The third is the structural one. Anthropic's June 15 credit-pool change signals that subsidized programmatic usage on subscription plans is ending across the industry. Enterprises that built their forecasts on flat-rate Claude Code economics will see their effective unit costs rise, and the same logic will apply to other vendors as they follow Anthropic's lead. Procurement teams that want predictability will need to negotiate committed-spend agreements at fixed rates rather than ride consumption pricing, and the leverage they bring to those conversations will depend on whether their engineering organizations have any usage caps in place at all.
Uber is not slowing its artificial intelligence push. Naga plans to test OpenAI's Codex alongside Claude Code, and the long-term vision he describes is one where agent engineers handle coding, testing and deployment with humans acting as orchestrators. That direction is consistent across major engineering organizations now adopting these tools. The open question for boards is not whether to deploy them but whether finance functions have any visibility into what they will cost when the engineers stop holding back.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。