News · 2026-06-24
A Coding AI Ran Through Uber's Yearly Budget in Four Months
Here is a number that should make any finance chief sit up: Uber handed an AI coding tool to roughly 5,000 of its engineers, and four months into the year the company had already burned through its entire 2026 budget for it. The tool did not break. The engineers did not misuse it. They used it exactly as intended, and the bill still ran away from everyone. The story, reported by Forbes and attributed to Uber's chief technology officer, is the clearest cautionary tale yet about the economics of AI agents.
Let me clear up the eye-catching figure first, because it gets garbled in retelling. Uber's total research-and-development spending was about $3.4 billion last year. That entire sum was not spent on one coding tool -- the budget that actually got exhausted in four months was the dedicated slice set aside for AI, specifically Anthropic's Claude Code. Even with that correction, the story is remarkable, because the overrun was not about scale. It was about a pricing model nobody had learned to forecast.
The background you need is how these tools are billed. Older enterprise software charges per seat: you pay a flat monthly fee per employee, multiply by headcount, and you have a number you can put in a spreadsheet a year ahead. Claude Code does not work that way. It bills by consumption -- you pay for every chunk of text the model reads and writes, every step it takes. And AI agents, the systems that can run many steps on their own, are voracious. The same engineer doing the same job can rack up wildly different bills depending on whether they used the tool for simple autocomplete or set it loose orchestrating dozens of parallel sub-tasks across a giant codebase. Uber's own figures show the spread: a typical engineer cost a few hundred dollars a month, but heavy users ran from $500 to $2,000, and the CTO reported spending $1,200 in a single two-hour session during a demo.
The analogy is a utility bill versus a subscription. A streaming service charges the same whether you watch one hour or a hundred. Your electricity bill charges by how much you actually use -- and if you install a new appliance that quietly runs all day, the bill balloons even though nothing is malfunctioning. Agent coding tools are the appliance that runs all day. The more useful they are, the more they run, and the more they run, the more you pay. Worse, productivity savings show up somewhere else entirely -- in shipped features, in time saved -- so the finance team sees the soaring cost line without an obvious offsetting number to net it against.
There is a human twist that made Uber's case worse, and it is a sharp lesson on its own. The company ranked engineers on internal leaderboards by how much they used the AI tool. That turned heavy consumption into a status game, which is a great way to drive adoption and a terrible way to control spending: the people racking up the tokens were not the people who had to answer for the budget. Adoption climbed from a third of engineers to the great majority in a couple of months, and by spring the large majority of committed code was coming from AI tools, with a slice of live updates written by agents with no human in the loop at all.
Why this matters: Uber is not an outlier, it is a preview. As more companies wire these agents into daily work, the gap between 'this tool is incredible' and 'this tool is unaffordable as priced' is going to become one of the central tensions of enterprise AI. It pairs directly with the bigger argument this week about whether the industry's economics are sustainable, and it is a concrete reason behind the disclosure that AI now writes most of the code at the labs building it -- enormous usage produces enormous bills. The honest caveat cuts toward the optimists: a runaway bill is only a problem if the work is not worth it, and Uber is not abandoning these tools -- it is adding controls, testing rivals, and learning to budget for consumption rather than seats. The lesson is not 'AI is too expensive.' It is that a pilot with a few engineers tells you almost nothing about what the same tool costs once a whole organization leans on it, and the companies that survive the transition will be the ones that put caps and meters in place before the bill arrives, not after. It is also one more reason businesses now treat the ability to swap one model for another -- so they are not trapped by a single vendor's prices, or by a model that could be pulled from the market overnight -- as basic insurance.