Uber’s AI “tokenomics” problem highlights how usage-based AI bills can spike before ROI is clear

Enterprises are increasingly running into a budgeting problem with generative AI that is poorly understood at the planning stage: token costs. In simple terms, AI providers typically price many services based on usage, and each prompt, query, image request, or AI-assisted coding task consumes “tokens,” which are units representing the text and data processed by the model. When more employees and workflows start using those tools, the bill can rise quickly, even if the company’s expectations were based on a more predictable spend model. That dynamic has been grouped under “AI tokenomics,” a phrase that is now circulating beyond AI labs and into corporate finance discussions.

Uber is one of the companies now at the center of that debate. According to reporting cited in a Stocktwits write-up, employees consumed enough AI services for Uber’s annual AI budget allocation to be exhausted within four months, a situation attributed to a report by The Information. The same write-up frames Uber’s experience as a cautionary example of how widespread internal adoption can create unexpectedly large bills, even when the effort is intended to boost productivity.

In response to the cost shock, Uber moved to restrict how much individual staff can spend on certain AI coding tools. TechCrunch reported June 2 that Uber instituted internal usage caps, citing Bloomberg’s reporting that the company set a monthly $1,500 cap per employee and per agentic coding tool, including Anthropic’s Claude Code or Cursor. TechCrunch also said employees’ usage is tracked through an internal dashboard, and that the caps can be exceeded with permission in some cases. Separately, a Bloomberg Law brief said Uber limits all employees to $1,500 in monthly token spending per AI coding tool, again framing the move as a cost-management step after Uber blew through its AI budget earlier this year.

The episode has also raised a question Uber has been careful to address publicly: whether higher AI usage is actually translating into better outcomes. In late May, Tom’s Hardware reported that Uber President and COO Andrew Macdonald warned there was not yet a clear link between “tokenmaxxing” and shipping useful consumer features, suggesting that, so far, management is struggling to draw a straight line from increased token consumption to tangible product wins. In the same coverage, the article describes the issue as partly an attribution problem, with many headlines focusing on usage but fewer results that clearly demonstrate user-facing impact.

The token-cost anxiety is not isolated to Uber. The Stocktwits write-up cited remarks by OpenAI CEO Sam Altman indicating that token costs have become a major concern for customers, including an example of companies saying they spent an entire year’s budget in the first quarter of 2026 and asking for improved efficiency. The same write-up also points to an internal-incentive contrast inside the AI industry, quoting Nvidia CEO Jensen Huang saying he would be alarmed if employees do not consume tokens worth a large fraction of their salaries. The contrast illustrates a broader divide: some leaders treat token consumption as a proxy for effort and learning, while others see it as an expense that can outpace measurable gains.

For Uber and other large employers, the practical implication is that token spending behaves differently from traditional IT. Unlike fixed-fee subscriptions, usage-based pricing can create “runaway adoption” risk when teams are enthusiastic and tools are easy to try. If staff begin using AI tools in exploratory ways, or if “agentic” coding features execute multi-step work, tokens can accumulate faster than internal budget models anticipate. That means AI governance starts looking like cost-control discipline, not only experimentation, with internal dashboards, tool-specific caps, and approvals becoming part of how companies run AI programs.

Several key details remain unclear from the publicly accessible summaries. Uber has not, in these accounts, laid out a full breakdown of its token spending by tool, the precise dollar totals behind the budget blowout, or how it forecasts the relationship between AI usage and product delivery going forward. The Information report itself is paywalled in the materials available here, and Uber’s cap policy may evolve as it learns which categories of usage drive value versus waste. What is still missing is a transparent set of metrics tying token consumption to engineering throughput, quality, and downstream consumer features. Next, investors and industry watchers are likely to look for whether Uber relaxes or tightens its caps over time, and whether future disclosures or earnings commentary offer a clearer account of which AI use cases are producing measurable ROI.