Uber’s 2026 AI budget
died in 4 months

When 84% of your engineers use the same coding agent, the budget question stops being “should we” and becomes “how do we account for it.” Uber exhausted its entire 2026 AI tools budget by April — the tool did not fail and engineers did not misuse it; they used it for exactly the workloads it was designed for: parallel agents, large-scale refactoring, automated test generation. Our judgment: the failure mode of coding agents is not rejection — it is ungoverned embrace, and consumption-priced tools need DevOps-grade cost control. Source: Forbes, company-reported.

The numbers behind the burn

Uber rolled out Claude Code in December 2025. Adoption climbed from 32% of engineers in February to 84% classified as agentic coding users by March; by spring, 95% of engineers used AI tools monthly, roughly 70% of committed code originated from those tools, and about 11% of live backend updates were written by agents with no human in the loop. CTO Praveen Neppalli Naga confirmed the overrun to The Information, saying the company was back to the drawing board on its assumptions. Uber’s R&D spend reached $3.4 billion in 2025, up 9% year over year — so this collapse is less about scale and more about a pricing model enterprise finance has not learned to manage.

The distribution of cost is the instructive part: monthly cost per engineer averaged $150–$250, with power users between $500 and $2,000; Naga himself reported spending $1,200 in a two-hour session. Engineers used the tool for exactly the workloads it was designed for — parallel agent execution, large-scale codebase refactoring, automated test generation, backend code production. From a productivity standpoint the rollout was a success; from a finance standpoint it was a runaway.

Why per-seat budgets cannot absorb token bills

Claude Code does not price per seat. It meters tokens consumed across model calls: an engineer running autocomplete consumes a fraction of what an engineer orchestrating parallel agents across a monorepo consumes. 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.

Contrast Microsoft 365 Copilot Enterprise at $30 per user per month: a flat line item finance can multiply by headcount, with the vendor’s upside capped. Anthropic’s consumption model gives the vendor unlimited upside on heavy users and finance almost no forward visibility. Both models are defensible; treating them as interchangeable in a planning cycle is what produced Uber’s outcome. Uber compounded the dynamic by ranking engineers on internal leaderboards based on usage — turning token consumption into a competition. The teams driving adoption were not the teams managing spend, and that organizational gap was the load-bearing flaw.

The structural shift: subscription pricing cannot carry agentic workloads

Behind the single case is a pricing migration. On May 13, Anthropic announced that from June 15, agent tools and third-party harnesses on paid Claude subscriptions would be metered separately at full API rates; GitHub moved Copilot to a credit-based system on June 1. Analysts expect most vendors to introduce separate consumption pools for agents within 12–24 months — the vocabulary will vary (credits, requests, messages, compute units) but the direction is set: flat-rate inference for unbounded agentic workloads never survived the math, and vendors will pass the cost mechanics through to buyers.

Our judgment: the failure mode is ungoverned embrace

The industry’s standard response to consumption-cost stories is that AI pays for itself in productivity. Uber complicates that argument. The marginal productivity gain from a senior engineer running agentic workflows must clear a much higher token-cost hurdle than the gain from autocomplete; 5–20x increases in per-developer consumption are documented in agentic mode, and no public benchmark shows a matching multiplier on output value. Productivity savings also do not appear in the same line item as AI cost, so finance cannot net them out inside a quarterly review.

Our position: pilot economics do not predict scale economics. Pilots run on a few engineers using autocomplete; production runs on whole teams delegating multi-step workflows to agents — an order of magnitude more consumption. Industry data confirms the governance gap: only 43% of organizations have formal AI governance policies and only 21% have mature agentic governance. Most enterprises do not yet apply to AI tooling the spending controls DevOps routinely applies to cloud compute: per-engineer caps, real-time token monitoring, alerts before overrun rather than after. Uber deployed organization-wide without those controls, and the result was visible within a quarter.

Four controls you can copy

Uber is not slowing down: Naga plans to test OpenAI’s Codex alongside Claude Code, with a long-term vision where agent engineers handle coding, testing, and deployment while humans orchestrate. That direction is consistent across major engineering organizations. The open question for boards is not whether to deploy — it is whether finance has any visibility into what these tools will cost once engineers stop holding back.

Actions and the question for buyers

Actions: before any organization-wide coding-agent rollout, install the four controls (token budgets, no leaderboards, workflow-class cost models, committed rates); pilot cost is not scale cost — do not build a full-year budget from pilot numbers. Question for buyers: what is your heavy-user monthly cost per engineer, and can finance see it in real time? If you cannot answer, you are the next company that burns its budget by April.

References: Forbes, Uber Burns Its 2026 AI Budget In Four Months On Claude Code (company-reported; Janakiram MSV, 2026-05-17) https://www.forbes.com/sites/janakirammsv/2026/05/17/uber-burns-its-2026-ai-budget-in-four-months-on-claude-code ｜ The Information interview with CTO Praveen Neppalli Naga (cited by Forbes) https://www.theinformation.com/newsletters/applied-ai/uber-cto-shows-claude-code-can-blow-ai-budgets ｜ Axios on Anthropic’s separate agent credit meter (2026-05-14) https://www.axios.com/2026/05/14/anthropic-claude-price-openai-tokens ｜ SiliconANGLE on GitHub Copilot credits (2026-05-14) https://siliconangle.com/2026/05/14/anthropic-announces-programmatic-credit-pool-agentic-tool-use-rises/ ｜ InfoWorld on industry consumption pooling https://www.infoworld.com/article/4171274/anthropic-puts-claude-agents-on-a-meter-across-its-subscriptions.html