Uber burned its entire 2026 AI budget in four months.
The company had been running an internal push to get every engineer using AI tools as heavily as possible, with leaderboards ranking staff by usage. Praveen Neppalli Naga, Uber’s Chief Technology Officer, told Fortune on 7 August that the approach is over.
His phrase for it: “We’re coming to the end of the so-called tokenmaxxing era.”
What makes this worth a post is not that a company overspent. Companies overspend constantly and rarely say so. It is that a named executive at a large, competent, deeply technical organisation has described the specific mistake in public, and the mistake is one that is easy to repeat.
The leaderboard did exactly what it was built to do
Rank people by how much they use a tool and you will get high usage. That is not a malfunction. It is the mechanism working correctly.
The trouble is that usage was never the thing anyone wanted. It was standing in for something harder to see, which is whether the work got better. Uber had four months of rising numbers on a dashboard that measured effort rather than result, and the numbers looked like progress the entire time.
This is an old management failure wearing new clothes. Anyone who has watched a sales team optimise for calls logged, or a support team for tickets closed, recognises it immediately. The novelty here is only the units. Everything else about the pattern is familiar, which is precisely why it caught an organisation that would tell you it knows better.
The number that runs the other way
Here is the part I did not expect, and it complicates the story usefully.
Uber quadrupled the number of employees using frontier AI tools, and reports that its cost per token fell rather than rose. Naga again: “You might expect costs to rise as adoption accelerates. We’ve seen the opposite.”
So the budget did not disappear because each unit got more expensive. It disappeared because the volume went up enormously while the unit price was falling. Four times the people, on a programme explicitly designed to encourage maximum consumption, is enough to exhaust a year’s allocation even against a falling price.
That is worth holding onto if you are watching your own AI spend. Falling unit costs are real and they are not protection. A cheaper thing consumed without limit is not cheaper.
What they did that worked
The more useful half of the story comes from Fortune’s earlier reporting, published on 9 July, and I want to keep the dates separate because they are two different pieces of news.
The July piece describes something Uber calls agentic pods. Thirty of its most AI-proficient engineers were embedded directly into finance, legal and HR for two weeks at a time. Not to run training. To sit inside those functions and work on their actual problems.
Uber’s reported results from that: financial pacing reports went from two days to ten minutes. Allocating capital across the 150 cities the company operates in went from fifteen hours to thirty minutes.
Those figures are from July and they are Uber describing its own programme, so treat them as the company’s account rather than an independent finding. But the method underneath them is the part that transfers, and Naga states it plainly: “You can’t automate them effectively by looking at process diagrams or documentation. You have to understand how the work actually gets done.”
Why that sentence is the whole thing
Read it again with the leaderboard in mind, because the two approaches are opposites.
The leaderboard is a way of driving adoption without ever going and looking. It scales beautifully. It requires no understanding of any particular job. You can run it from a dashboard across thousands of people, and it will produce a number that rises.
The pods do not scale like that at all. Thirty people, two weeks each, sitting inside three functions. That is slow, expensive in senior attention, and impossible to run from a spreadsheet.
It is also the one that produced results anybody could point at.
The gap between those two approaches is not really about AI. It is the difference between distributing a tool and understanding a job, and organisations have been choosing the first because it is legible and cheap for as long as there have been tools to distribute.
The loop back to Monday
On Monday I wrote about OpenAI’s finding that AI use has grown 108 times in legal since February, 41 times in sales and recruiting, and only five times in engineering. The functions everyone assumed would lead are not leading.
Uber found the same thing the expensive way.
The programme aimed at engineers burned the budget. The programme that put capable people next to finance, legal and HR produced the numbers the CTO is willing to quote. Those are the same three kinds of function that came top of OpenAI’s growth table, arrived at independently, by a company that spent a lot of money learning it.
Two organisations pointing at the same place is not proof. It is a strong hint about where to look.
What to check
Look at what your AI rollout actually measures.
If there is a dashboard, find out what is on it. Seats issued, logins, messages sent and licences activated are all measures of distribution. None of them tells you whether a single piece of work improved.
If it counts usage, usage is what it will get you. Uber has now paid to demonstrate that, and said so out loud, which is more than most organisations will do.
Sources: Uber says the “tokenmaxxing” era is over, Fortune, 7 August 2026, and Fortune’s earlier reporting on Uber’s agentic pods, 9 July 2026. All figures are Uber’s own account of its own programme.