Uber spent itself into efficiency
Uber’s Chief Technology Officer, Praveen Neppalli Naga, admitted in an interview with The Information that the company overspent its AI budget within the first few months of the year. The company pushed employees to use tools like Anthropic’s Claude Code extensively, even creating competition through leaderboards to track usage. This aggressive AI adoption mirrored a trend dubbed 'tokenmaxxing,' where companies prioritized high usage without achieving meaningful returns. However, Uber has now recalibrated its approach, with Naga stating that the era of tokenmaxxing is drawing to a close.
Uber managed to reduce the cost per token by quadrupling the number of employees using advanced AI tools. The company achieved this efficiency by refining prompt caching techniques, adjusting default model settings, testing new models for performance, and giving engineers visibility into their AI usage and costs on an hourly basis. This strategy shifted the focus from budget constraints to solving efficiency as an engineering challenge.
AI cost reductions are not the end of the story
Even though costs have decreased, Uber has yet to connect AI usage directly to increased productivity. Uber President and COO Andrew Macdonald told the Rapid Response podcast that it's difficult to draw a clear line between AI tool use and tangible consumer benefits. He noted that more features may be being developed, but it's unclear if the impact is measurable in terms of productivity gains.
Uber isn’t the only company struggling to show tangible returns. According to Jim Reid, head of macro and thematic research at Deutsche Bank, AI-driven productivity gains are still years away. Between the first quarter of 2023 and 2026, the Magnificent Seven—major tech firms—saw their profit margins rise from 15% to 25%. In contrast, the rest of the S&P 500 index experienced only 10% growth over the same time, indicating limited broader economic returns from AI outside the tech sector.
The AI Jevons paradox kicks in
Economists warn that the drop in AI token costs could lead to unexpected patterns of behavior. This phenomenon, known as the Jevons paradox, was first observed by William Stanley Jevons in 1865 when more efficient coal use led to higher coal consumption. Today, the paradox is playing out in the AI world as well. Despite token costs falling more than 90% since 2023, AI spending has more than doubled by late 2024. This suggests companies are using AI more intensively rather than cutting back, even as the cost per token drops.
A report by Bain and Co. in June supports this view, showing that token costs dropped by half from December 2024 to 2025. However, the amount of tokens used skyrocketed by 450% during the same period as companies upgraded their AI tools. Naga acknowledged this shift and said Uber is now focusing on the quality of AI use rather than the quantity, but he hasn't specified if the company is using more or fewer tokens in its operations.