Uber spent its entire 2026 artificial intelligence budget by April, roughly four months into the fiscal year, after adoption of Anthropic’s Claude Code exploded across the company’s engineering organization faster than anyone in leadership anticipated. Monthly API costs per engineer ranged from $500 to $2,000, and by March, 84% of Uber’s engineers were classified as “agentic coding users,” up from 32% in February. The company is now “back to the drawing board” on AI budgeting, according to its CTO, while COO Andrew Macdonald publicly questioned whether the spending is producing anything consumers will notice.
The Adoption Curve No One Planned For
Uber rolled out Claude Code access to its engineering team in December 2025. Usage doubled by February as developers discovered its multi-step capabilities for code generation, review, and debugging. By March 2026, the tool had gone from a curiosity to the default workflow for most of the engineering organization.
The numbers are staggering. Ninety-five percent of Uber engineers now use AI tools monthly. Seventy percent of committed code originates from AI. The per-engineer token consumption grew so fast that the annual budget, set in late 2025 before anyone knew what agentic coding adoption would look like, was exhausted before the second quarter began.
Fortune reported that Macdonald delivered his skepticism at an internal town hall in May, telling employees that “it’s very hard to draw a line between one of those stats and, ‘Okay, now we’re actually producing 25 percent more useful consumer features.'” That is the core tension: the inputs are measurable (tokens consumed, code committed, engineers activated) but the outputs (features shipped, user experience improvements, revenue impact) remain fuzzy.
The ROI Question Every Enterprise Is Asking
Uber’s situation is not unique. It is the loudest version of a problem every large enterprise adopting AI coding tools is confronting: developer productivity tools are easy to deploy and addictive to use, but connecting that usage to business outcomes requires a measurement framework that most companies have not built.
The traditional way to measure engineering productivity, lines of code, pull requests merged, sprint velocity, breaks down when 70% of the code is machine-generated. A developer who uses Claude Code to generate 500 lines of boilerplate in an hour has produced “more code” but not necessarily more value than a developer who spent that hour designing an architecture that avoids the boilerplate entirely.
Uber’s CTO acknowledged this gap explicitly, saying the company needs to develop new metrics that capture the quality and business impact of AI-assisted development rather than just the volume. That work is underway, but it is happening after the money was already spent.
What the Budget Blowout Tells Us About AI Economics
The speed of Uber’s budget exhaustion reveals a structural problem in how enterprises are pricing AI adoption. Most companies set AI tool budgets based on projected per-seat licensing costs, similar to how they budget for GitHub Copilot or Jira. But agentic coding tools like Claude Code charge by token consumption, and consumption scales with capability: the more powerful the model, the more tokens each interaction burns, and the more interactions developers initiate once they discover what the tool can do.
Uber’s experience suggests that the real cost of enterprise AI coding tools is not the sticker price per seat. It is the emergent behavior that follows adoption: developers who start with simple code completion quickly graduate to multi-step agentic workflows that consume 10 to 50 times more tokens per session. The budget model that works for GitHub Copilot does not work for Claude Code, and most enterprises are going to learn that lesson the same way Uber did.
The Broader Market Implications
For Anthropic, Uber’s budget blowout is both a validation and a warning. Claude Code is clearly sticky, with adoption rates and usage intensity that any enterprise software company would envy. But if the largest customers start publicly questioning ROI, it creates a narrative problem that could slow new enterprise sales.
For the broader AI infrastructure market, Uber’s story is a data point in the emerging debate about whether AI tool spending will follow the cloud computing trajectory (expensive early, indispensable later, budgets adjusted upward) or the blockchain trajectory (expensive early, never proved the ROI, budgets quietly cut). The answer probably depends on whether companies like Uber can build the measurement frameworks that connect AI tool usage to business outcomes before the CFO loses patience.