Uber Put a Leaderboard on Token Spend. The Budget Ran Out in April.

The first big public AI cost blowout was never about a tool that failed. It is about what happens when a company measures the input and calls it output.

In December 2025, Uber gave its 5,000 engineers access to Claude Code, Anthropic’s agentic coding tool. By February, 32 percent of them were classified as agentic coding users. By March, 84 percent. By April, according to CTO Praveen Neppalli Naga, the company had spent its entire 2026 AI budget. Four months into the year.

The figures Naga shared with The Information in May are the most specific any large company has put on the record about what AI-assisted engineering actually costs. Average spend ran $150 to $250 per engineer per month. Heavy users ran $500 to $2,000. Naga said he personally burned $1,200 in a two-hour demo. Roughly 70 percent of committed code at Uber now originates from AI tools, and about 11 percent of live backend changes ship with no human in the loop at all.

None of that is a failure. The tool worked. The engineers used it. Uber spends $3.4 billion a year on R&D, and the AI line was never going to sink the company. What makes the story worth reading twice is a detail that got less attention than the dollar figures: Uber ran internal leaderboards ranking engineers and teams by how much Claude Code they used.

Consider the incentive that creates. Agentic tools are metered per token, not per seat. A single overnight agent run can consume what a couple hundred casual users would in the same period. Put a scoreboard on top of that meter, tell people that being high on the board is a good thing, and hand the bill to a different team than the one running the contest. You do not need a finance background to guess how it ends.

We have seen this metric before

Every generation of engineering management finds a number that is easy to count and starts treating it as productivity. Lines of code came first. Commits, story points, and pull request volume followed. Each one measures something real. Each one gets gamed the moment it becomes a target, and each one eventually gets dropped once leadership notices that the number going up and the product getting better are two separate events.

Token consumption is the newest entry on that list, and it may be the most dangerous, because it is the first one that converts directly to cash. An engineer who pads a commit count costs you nothing extra. An engineer who leaves an agent looping on a poorly specified task overnight costs you real money, and the leaderboard tells them they had a great day.

“Token consumption may be the most dangerous productivity metric yet, because it is the first one that converts directly to cash.”

 

Around the same time as Uber’s disclosure, Nvidia’s VP of applied deep learning, Bryan Catanzaro, told Axios that on his team the cost of compute now exceeds the cost of the people. That is Nvidia, where compute is the product, so take it with some salt. But it signals where engineering budgets are heading. Inference is becoming a line item that sits beside salaries rather than beneath them. Any organization that treats it as a rounding error on the tooling budget is going to have Uber’s spring.

What Uber did next is the useful part

Fortune reported on August 7 that Uber did not respond by pulling the tool back. It added prompt caching, changed default model selection so routine tasks stopped defaulting to the most expensive option, and built usage dashboards so teams could see their own spend. Usage has since roughly quadrupled from January levels while cost per token has fallen. Whether the productivity is there remains an open question. COO Andrew Macdonald said in July that it was getting harder to justify token costs without evidence of more useful features reaching customers. But the cost curve, at least, bent.

What Uber did is what cloud teams learned to do a decade ago. When AWS bills first surprised CFOs, nobody stopped using the cloud. They built FinOps: tagging, reserved capacity, anomaly alerts, showback to the teams doing the spending. AI coding tools are roughly where cloud was in 2014. The controls exist in principle. Almost nobody has them wired up. Survey data cited in Forbes’s coverage of the Uber overrun found that only 43 percent of organizations have formal AI governance policies at all, and 21 percent have anything resembling mature governance for agents.

The cheaper fix is upstream

Caching and dashboards manage the bill. They do not address why an agent burned $1,200 in two hours in the first place.

Talk to engineers who have worked with these tools at scale and a pattern shows up quickly. The expensive runs are almost never the ones where the task was clear. They are the ones where the agent was handed an ambiguous ask, made a reasonable guess, built the wrong thing, got corrected, tried again, and looped. Every loop costs tokens. Every loop also produces code that someone has to review and probably discard. The token bill and the review burden grow from the same root: nobody wrote down what “done” looked like before the agent started.

This is why the organizations getting predictable results from agentic coding tend to look old-fashioned on the front end. They write specifications. They define acceptance criteria before the first prompt. They keep senior engineers at the decision points rather than only at the end of the line, because a senior engineer who spends ten minutes tightening a spec saves an agent an hour of wandering, and saves the company the tokens that wandering consumes. None of this discipline is new. What is new is that skipping it now has a price you can read off a dashboard.

Uber’s engineers, to their credit, appear to have worked a lot of this out on their own once they could see their own numbers. That is the argument for dashboards. The argument for specs and senior oversight is that you would rather not learn it after the budget is gone.