Uber’s AI Conundrum: A Cautionary Tale of Tokenmaxxing

Uber’s Chief Technology Officer, Praveen Neppalli Naga, recently announced that his company has found a solution to its AI spending problem. Having blown through its 2026 budget in just the first few months, Uber is not an isolated case. Many companies have fallen victim to “tokenmaxxing,” a trend where businesses incentivize employees to use AI tools extensively without seeing a corresponding return on investment.

Tokenmaxxing has been likened to a game, with companies competing to see who can consume the most tokens – essentially, units of computing power used by AI systems. Naga noted in his X post that “We’re coming to the end of the so-called ‘tokenmaxxing’ era.” This statement raises questions about the future of AI adoption and the companies investing heavily in it.

One key takeaway from Uber’s experience is that simply throwing more tokens at a problem doesn’t necessarily lead to better results. In fact, costs can decrease even as adoption accelerates. However, this raises a crucial question: what happens when tokens become cheaper? Will companies continue to spend more on AI agents and automation, effectively canceling out any potential cost savings?

The concept of Jevons paradox seems particularly relevant in this context. This phenomenon, named after 19th-century economist William Stanley Jevons, states that as a resource becomes more efficient, its consumption often increases rather than decreases. The price of tokens has dropped by over 90% since 2023, but large language model spending has doubled since late last year.

This paradox is not unique to AI; it’s a classic example of how human behavior can be at odds with economic logic. As Torsten Slok, Chief Economist at Apollo, noted, “As tokens get cheaper, companies don’t spend less but instead run more AI agents, automate more workflows and generate more code, pushing aggregate expenditure higher even as the unit cost of intelligence collapses.”

Uber’s solution to this problem involves improving prompt caching, adjusting default model settings, and evaluating new models for efficiency. Naga emphasized treating efficiency as an engineering problem rather than a budget problem. This approach suggests that companies need to rethink their approach to AI adoption, moving from a quantity-based mindset to one focused on quality.

The stakes are high; with profit margins swelling for the Magnificent Seven while remaining stagnant for the broader S&P 500 index, companies must deliver on their massive AI investments. As Jim Reid, global head of macro and thematic research at Deutsche Bank Research Institute, warned last month, “AI productivity gains were still years away.”

The future of applied AI at enterprise scale will not be characterized by who spends the most tokens but about how people use them as efficiently as possible. Uber’s experience serves as a cautionary tale for companies considering investing heavily in AI. While the promise of AI is clear, its implementation requires careful consideration and strategic planning to avoid falling into the trap of Jevons paradox.

The next phase of AI adoption will be marked by a shift from quantity to quality, where companies focus on using tokens efficiently rather than simply consuming them. As Naga concluded, “This is the future of applied AI at enterprise scale.” The question remains: are other companies ready to follow suit?