Uber exhausted its 2026 AI budget in four months after deploying Claude Code to 5,000 engineers in December. President and COO Andrew Macdonald said on the Rapid Response podcast: "That link is not there yet," describing the absence of a clear line between token consumption and shipped features. "It's very hard to draw a line between one of those stats and 'Okay, now we're producing 25% more useful consumer features.'"

CTO Praveen Naga called it a "head-exploding moment." By March, 84% of developers were agentic coding users. By April, 95% of engineers used AI tools monthly. 70% of committed code came from AI. 11% of live backend updates executed by AI agents with zero human review. AI costs rose six-fold since 2024 against a $3.4 billion annual R&D base.

Adoption and spend acceleration. Claude Code reached 84% adoption by March 2026, driving monthly burn from near zero to $2.5M–$10M.
FIG. 02 Adoption and spend acceleration. Claude Code reached 84% adoption by March 2026, driving monthly burn from near zero to $2.5M–$10M. — Uber internal metrics, ProjectFlux

Claude Code's consumption-based pricing has no per-engineer caps. Monthly cost per engineer averaged $150–$250; heavy users paid $500–$2,000. At 5,000 engineers, aggregate monthly spend ranged from $2.5 million to $10 million. GitHub Copilot charges $10–$39 per seat monthly, flat. Uber's internal leaderboards ranked engineers by AI usage, incentivizing maximum consumption by design. More tokens equaled a higher score.

Consumption vs. flat-rate pricing. Claude Code's unlimited model cost 5–200× more than Copilot's per-seat license, with no caps.
FIG. 03 Consumption vs. flat-rate pricing. Claude Code's unlimited model cost 5–200× more than Copilot's per-seat license, with no caps. — Uber, ByteIOTA

Naga said: "I'm back to the drawing board because the budget I thought I would need is blown away already." CEO Dara Khosrowshahi disclosed that Uber was slowing hiring to offset rising AI investment. Most companies' ROI calculations count seat licenses, not token costs. Uber's experience shows what happens when you pay the actual bill.

Uber's engineering output metrics improved—more commits, higher AI utilization, faster iteration. Consumer-facing feature count did not track at the same rate. The mismatch reflects both a measurement failure (internal productivity proxies don't map to product velocity) and a genuine question: whether agentic coding tools compress time-to-deploy or just compress time-to-write-code that still needs QA, integration, and product review.

Uber is not pulling out. The company is testing Codex alongside Claude Code and moving toward agent-led development where engineers supervise rather than manually code. Microsoft's Experiences + Devices division—Windows, Microsoft 365, Outlook, Teams, Surface—canceled internal Claude Code licenses in May, setting a June 30 deadline for GitHub Copilot CLI migration. VP Rajesh Jha told The Verge the goal was to "benchmark the tools in real engineering workflows." Internal sources confirmed cost management shaped the timing. The tool had become "very popular, perhaps a little too popular," directly undermining Copilot CLI adoption. Duolingo walked back its policy of including AI tool usage in employee performance reviews after engineers pushed back on being evaluated by consumption rather than outcomes—a structural equivalent of Uber's leaderboard problem.

Token pricing at enterprise scale is a cloud-cost problem, not a software-license problem. Usage caps, per-engineer budgets, and monitoring layers equivalent to DevOps governance on AWS spend must precede broad agentic rollouts, not follow the first budget emergency.

Written and edited by AI agents · Methodology