Uber burns year's AI budget in 4 months as tech cuts back on agents

For the past three years, tech executives couldn't stop promising that AI would soon replace software engineers. But as enterprise billing statements roll in, a far more awkward reality has emerged: hiring human developers is often significantly cheaper.

Reports surfaced showing that Uber had already blown through its entire 2026 budget for AI coding tools in just four months. The company had created an internal leaderboard to incentivize AI adoption by ranking teams on total tool usage. The result was not a miraculous leap in productivity, but an unmitigated budget blowout.

The leaderboard that broke the budget

Uber's President and Chief Operating Officer, Andrew Macdonald, sounded the alarm on the Rapid Response podcast. He admitted it has become nearly impossible to draw a clear line between the heavy internal use of tools like Claude Code and real, customer-facing innovation.

Macdonald noted that while more raw code might be shipping under the surface, there is simply no evidence that it is producing 25% more useful features for app users. Without that return, escalating spend becomes impossible to justify.

Yet just weeks ago, Uber CEO Dara Khosrowshahi told investors that about 10% of the firm's committed code was being generated by autonomous agents. Uber spent $951 million on R&D in the first quarter of 2026 alone, up nearly 17% year-over-year. The tools spread fast across engineering, legal, and marketing, but the financial return failed to keep up.

Cutting licenses as agents burn through cash

Uber is far from alone in hitting the brakes. Microsoft recently moved to cancel most of its direct Claude Code licenses, steering engineers back to GitHub Copilot CLI to rein in costs.

Nvidia has acknowledged that its internal AI expenses now exceed what it pays the teams using them. Other companies, including Starbucks, are quietly scaling back their AI bets. Even Duolingo CEO Luis von Ahn previously walked back his bullish stance, admitting AI is not replacing his staff's core tasks.

Counterintuitively, tumbling per-unit model prices have made the problem worse rather than better. A study from Gartner found that while inference costs will drop by 90% by 2030, total enterprise AI software spend is projected to surge to nearly $207 billion in 2026.

Standard chatbots take a single prompt, but autonomous agents run looped tasks that consume massive volumes of tokens behind the scenes. With Anthropic shifting to purely usage-based pricing and OpenAI CEO Sam Altman comparing future intelligence to metered utilities like water and electricity, uncontrolled agents turn into runaway financial drains.

Why it matters

Corporate tech has officially entered its AI hangover phase. The blind FOMO of the last two years is crashing headfirst into brutal enterprise procurement accounting.

For engineering managers and developers, the takeaway is clear: stop using vanity metrics to evaluate AI tooling. Uber's internal leaderboard rewarded activity rather than impact, incentivizing teams to generate tokens instead of solving structural business problems. If you are deploying coding agents today, hard token caps and feature-level ROI tracking are no longer optional.

For the broader industry, this punctures the myth of friction-free labor replacement. Repairing AI-generated bugs and financing agentic reasoning loops carries enormous hidden overhead. In many workflows, competent human engineers remain far more cost-effective than autonomous code generators.

We see this pushback as a necessary correction. When even Silicon Valley bellwethers refuse to sign blank checks for utility-metered AI, the burden shifts back onto frontier model providers to prove actual efficiency rather than celebrating runaway compute usage.