Claude Code After Uber burns through its annual budget in two months, the COO straightforwardly states: Token consumption and useful output are not proportional.

Uber COO Andrew Macdonald openly admits in a recent interview that the company’s AI spending is becoming increasingly difficult to explain internally. Two months ago, CTO Praveen Naga revealed that the Claude Code budget had already been burned through early, but the more core issue is this: higher token consumption does not translate into a proportional increase in consumer feature output.
(Background: Not just ride-hailing — Uber partners with Expedia to add hotel bookings, moving toward a one-stop travel super app)
(Additional context: Anthropic report: In the race for AI dominance by 2028, if the United States doesn’t hold on to its computing power advantage, China may overtake it)

Table of Contents

Toggle

• A budget crisis
• The broken causal chain: More tokens, not necessarily more features
• Industry divergence: Burn as much as possible, or clarify before you burn

When every engineer at a company spends up to $2,000 per month on AI tools, 70% of submitted code is generated by AI, yet no one can answer the question of “how many features did this actually deliver?”—it is no longer a technical problem, but a management crisis.

Uber Operating Officer Andrew Macdonald recently gave a Rapid Response interview and said something that is an open secret in the tech industry: the money spent on AI is becoming harder and harder to account for to people.

A budget crisis

Earlier, Uber CTO Praveen Neppalli Naga also said in an interview with The Information in April: “I thought the budget I had in mind had already been exhausted early.”

At the time, the context was this: Uber’s 5,000 engineers’ adoption rate of Claude Code surged from 32% to 84% in just a few months. Individual engineers were spending anywhere from $500 to $2,000 per month; and Naga himself, during an internal demo, consumed $1,200 worth of token allotment in just two hours.

Macdonald described how that remark shook things up among executives at Uber and sparked a series of discussions about AI token consumption—whether these expenses are worth it, and the trade-offs they create for staffing.

This month, CEO Dara Khosrowshahi was explicit during the earnings call: Uber is slowing down hiring, partly to offset the spending on AI investments. In other words, the AI tools’ bills are starting to affect real hiring decisions.

The broken causal chain: More tokens, not necessarily more features

In the interview, Macdonald described his findings after communicating with Uber’s senior engineering leadership: higher token usage does not translate into a proportional increase in consumer feature output.

“The link doesn’t exist yet, right?” he said. “Maybe in some faint way more things have been delivered, but it’s very difficult to draw a line between those numbers and ‘we produced 25% more useful consumer features.’”

This problem reveals the core contradiction behind the current AI adoption wave: token consumption is measurable, but what it measures is “degree of usage,” not “value of output.” Salesforce recently dubbed these kinds of metrics “vanity metrics” and explicitly opposes using token consumption volume as a standard for evaluating employee performance.

Notably, Macdonald also pointed out a blind spot in how people think: for individual engineers who don’t pay out of pocket, AI tools “feel like they’re free,” so they can try out all sorts of usage scenarios. But in the end, it’s the company that pays. This misalignment of costs between individuals and the organization is one of the structural reasons why token consumption gets out of control.

Industry divergence: Burn as much as possible, or clarify before you burn

Uber’s confusion isn’t an isolated case—it’s just one of the first to be called out by executives.

At Google I/O 2026, the company strongly promoted “tokenmaxxing,” meaning using AI as extensively as possible, and making it one of the metrics to gauge how much effort engineers put in. The logic behind this approach is: usage itself drives capability evolution, and quantitative change will eventually lead to qualitative change.

But some companies are heading in the opposite direction. Duolingo once included AI usage frequency in performance evaluations; however, after employees raised questions—“Should we use AI just in order to use AI?”—the policy was quietly withdrawn. In an April podcast interview, CEO Luis von Ahn said, “It feels like rather than holding everyone accountable for actual results, we’re pushing something that in many situations simply isn’t applicable.”

A healthcare company represents an even more extreme case: within six months it consumed 1 trillion tokens, generating more than $6 million in unplanned costs, and at the time the finance department even didn’t know what the driving factors were. This isn’t a problem of using AI—it’s that no one knew who was using it, where it was being used, or how much money was being burned.

In the interview, Macdonald did not announce any specific cutback plans, nor did he say Uber would abandon AI tools. He only stated a problem that is common across the industry, but rarely spoken plainly by executives.

When it comes to measuring ROI on AI investments, the industry has no standard answer. But more and more signs suggest that the gap between “how much is used” and “how much is obtained” is still very large.