Uber has been forced to radically rethink its artificial intelligence strategy after exhausting its entire 2026 budget for the technology within the first few months of the year. The company’s Chief Technology Officer, Praveen Neppalli Naga, admitted to going back to the drawing board after an initial blitz that encouraged maximum employee usage of tools like Anthropic’s Claude Code.
From 'Tokenmaxxing' to Engineering Efficiency
This trend, dubbed “tokenmaxxing,” saw Uber creating leaderboards to rank software engineers on their AI usage. However, the rapid spending failed to deliver immediate returns on investment, leading Naga to pivot. He now claims the company has found a way to scale AI without breaking the bank by treating efficiency as an engineering challenge rather than a budgetary constraint.
Key strategies implemented by Uber include:
- Improving prompt caching techniques.
- Adjusting default model settings for better performance.
- Rigorous evaluation of new models for operational efficiency.
- Allowing engineers to monitor their specific AI usage and hourly costs.
The Jevons Paradox and Market Realities
Despite reducing the cost per token, Uber faces a classic economic trap: the Jevons paradox. This phenomenon suggests that as a resource becomes more efficient and cheaper, total consumption actually increases. Data from the Silicon Data Token Expenditure Index shows token prices have plummeted by over 90% since 2023, yet spending on large language models has doubled in the same period.
Uber’s President and COO, Andrew Macdonald, noted the difficulty in linking AI metrics to tangible consumer benefits. He stated that it remains hard to draw a line between high usage stats and the production of useful consumer features, highlighting a broader industry struggle to realize AI-driven productivity gains.
