Uber exhausted its entire 2026 artificial intelligence budget in just four months, raising hard questions about whether corporate AI adoption is delivering returns that justify rapidly escalating costs.
The ride-hailing giant's president and COO Andrew Macdonald delivered a blunt assessment of the company's AI spending spree during an internal review: "That link is not there yet," he said, referring to the connection between rising AI tool costs and actual improvements in products customers use.
The budget blowout stemmed from an internal gamification strategy gone awry. Uber created an employee leaderboard to encourage adoption of AI coding tools, which succeeded beyond expectations. The problem? The company burned through its allocated AI token budget by April—just one-third of the way through the year.
Macdonald was candid about the ROI problem: "If you're not actually able to draw a direct line to how [many] useful features and functionality you're shipping to your users, that trade becomes harder to justify."
The numbers don't lie, but executives sometimes do. In this case, they're being refreshingly honest. Uber's R&D spending hit $951 million in Q1 2026, up nearly 17% year-over-year. Overall R&D expenses reached 3.4% of revenue in 2025, up 9% from 2024.
CEO Dara Khosrowshahi noted that roughly 10% of committed code now comes from autonomous AI agents rather than human engineers. That sounds impressive until you realize the company can't demonstrate that this AI-generated code is producing features customers actually value.
This is the dirty secret of enterprise AI adoption in 2026: while the per-token cost of AI continues to fall, agentic AI models consume exponentially more tokens per task. Companies thought they were getting a bargain. Instead, they're discovering that cheaper units don't matter when you're consuming 10x or 100x more of them.
Uber isn't alone in this predicament. A Goldman Sachs analysis I reviewed last quarter found that 60% of enterprise AI implementations show no measurable productivity gains after 12 months. The technology works—it's just not clear it works at current adoption rates.
