Uber’s tokenmaxxing reality check

TL;DR: Uber's president said over the weekend that the company is questioning the ROI on its AI spending after burning through its entire token budget in the first four months of 2026. It’s an early sign of a potential reckoning around the aggressive AI adoption we’ve seen in the corporate world. The shift in attitude shows that AI use for the sake of it isn’t necessarily a magic wand for boosting overall company performance.

What happened: The company’s president, Andrew Macdonald, admitted on a podcast that it’s unclear whether its devs’ extensive use of Claude Code is actually leading to useful features for its consumers, saying in an interview that the “link is not there yet.”

It’s a notable change for the company, which has heavily encouraged AI use among its employees. Uber rolled out Claude Code for internal use last December, and now 95% of its engineers use AI tools monthly, with roughly 70% of committed code originating from those tools.

Tokenmaxxing falls flat: Uber is one of several tech companies that set up internal leaderboards pushing employees to compete on using the most AI—the logic apparently being that sending more AI prompts means becoming more efficient or producing more work. CEO Dara Khosrowshahi even said on an earnings call earlier this month that the company would slow hiring to put more money into AI investments (though Macdonald’s recent comments indicate that such a trade-off might be “harder to justify” if the AI payoff isn’t obvious).

Agentic AI tasks like those the Uber engineers are doing are more token-hungry, which in turn is driving up how much (and how quickly) companies spend on AI. The revelation in April that the company had already used up its AI budget for the year was, per Macdonald, a “head-exploding moment” that’s led to a recalibration on whether its all-in AI strategy is justified.

So what’s the deal?: Uber’s retreat from tokenmaxxing might look like an early warning that the corporate world is cooling on AI adoption. But here’s one point of evidence that demand overall remains hot: GPU rental prices have more than doubled since January, according to data compiled by AI compute investment firm AMP PBC. In fact, thanks to high demand, compute costs are now eating up a bigger chunk of major AI labs’ overall expenses. Both OpenAI and Anthropic have introduced higher-priced subscription tiers and separate API pricing for agentic AI this year as a result—which could, in part, be driving up costs for companies such as Uber.

Bottom line: Firms like Uber aren’t abandoning AI as a key workplace tool—but they are starting to take a more measured tack on using it, recognizing that an engineer who keeps their AI agent humming around the clock isn’t necessarily building something anyone wants to use. —WK