Uber’s budget-blowing geeks are making the drive for AI a costlier fare than it planned

A headline last week in Bloomberg caught my eye: “Uber caps usage of AI tools like Claude Code to manage costs.” It turns out that the ride-hailing service blew through its entire AI budget earlier this year and has now capped all employees at $1,500 (£1,100) a month in “token spending” for each AI coding tool. 

So who are these token spendthrifts? None other than the software engineers (programmers, to you and me), who are employed to write and maintain the software for its platform – the thing that links people seeking a ride with the drivers who provide it. 

For this labour, they are richly rewarded. A quick check reveals that they take home between £196,000 and £380,000 a year, depending on seniority.

So how come these dudes have such voracious appetites for tokens? Basically, because these are what AI systems run on. Short words (“the”, “and”, “is”) count as single tokens; longer words usually get split – “tokenised”– into two or three tokens; a rough approximation is that a token is 75% of an average word. 

But punctuation marks, spaces, line breaks and symbols are also tokens. And computer code (which AIs read and write) is more token-hungry than prose because variable names, brackets, indentation and syntax all count.

So when you interact with an AI, you are launching a token-manipulation process. You type in your words, which are tokenised, and the machine assembles different tokens into sentences for your consideration. 

And all of this manipulation has a price. Input tokens (your ChatGPT prompt, for example) are relatively cheap: $3 a million for Anthropic’s Sonnet 4.6 model; output tokens are pricier: $15 a million, because generating text costs more than “reading” it. For most consumer users, these numbers are purely theoretical because they’re not (yet) paying for using the technology, and they’re not heavy users. 

For corporations, though, it’s a very different story, which is where Uber comes in. It employs a lot of software developers, most of whom are heavy users of AI, probably mostly for the purpose of coding. And, unlike you and me, they use the technology to build and launch agents: models that have been given tools (and permission to use them autonomously) on tasks such as reading files, running code, browsing the web and calling external services. And then looping through a cycle of observing, reasoning and acting, until they judge a task to have been completed or are told to stop.

In December 2025, Uber rolled out access to Cursor and Claude Code – two formidable AI programming tools – and its geeks went for them like engineering ostriches going for brass doorknobs (as PG Wodehouse would have put it). Usage of the tools doubled in February and, by April, the company had discovered two awkward facts: one was that running an agent intensively all day across a complex codebase such as Uber’s could burn through tens of millions of tokens; the other was that it had used up the entire AI budget for 2026.

There are all kinds of interesting lessons to be drawn from Uber’s experience. One is that it fell into the trap that many other companies (and governments) are leaping into: incentivising staff to adopt AI using stunts such as internal leaderboards that rank teams by AI tool usage without calculating what the cost of success might be.

Another is that it’s hard to draw a connection between a company’s rising use of Claude Code and innovations that are meant to serve customers. “That link is not there yet,” Uber’s chief operating officer, Andrew Macdonald, admitted in a recent podcast interview. “Maybe implicitly there’s more that is getting shipped, but it’s very hard to draw a line between one of those stats and: ‘OK, now we’re actually producing like 25% more useful consumer features.’”

The bottom line is that AI adoption is not the magic bullet that many organisations (and UK politicians and former prime ministers) seem to think. And tools such as Claude Code and Cursor don’t behave like the apps that companies (and governments) were accustomed to buying in the old days. 

It’s easy to assess and model the costs and benefits of a per-employee licence for an old-style business productivity app such as Slack, for instance. But a “token meter” attached to a software developer who can set up an AI agent to inspect a large corporate codebase, conduct tests, write software and repeat all day is something else entirely. 

And, just for the avoidance of doubt: there’s no such thing as a free token.

Ringing endorsement

AI Models Are Having Their iPhone Moment. What Next? is a very perceptive analysis by Om Malik of where AI fits in the evolution of the tech industry.

On the menu

Theoretical physicist Richard Feynman’s solving of the “restaurant dilemma” 50 years ago is confirmed by the maths in a nice Nature article.

Upper hand

A great essay by Jasmine Sun is The Independent Writer’s Advantage in the Age of AI.

Photograph by Drew Angerer