From Tokenmaxxing Hype to Hard Questions on ROI
Tokenmaxxing is the practice of aggressively driving up AI token usage across an organization in the belief that more tokens will translate into more productivity, prestige, or innovation, even though many teams lack clear ways to connect that rising consumption to measurable business results or sustained competitive advantage. That disconnect is now public. Uber COO Andrew Macdonald said he has not seen a direct link between higher AI token usage and output, noting that “it’s very hard to draw a line” from internal metrics to “25% more useful consumer features.” At the same time, Uber reportedly burned through its annual AI budget in four months, while Visa has bragged about monthly spend nearing 2 trillion tokens. Behind the leaderboard talk, engineering leaders warn that millions of tokens have been burned without “any real significant ROI” to show for it.
Budgets Blown, CFOs Nervous: The New Cost Reality
The backlash is rooted in AI token costs that are rising faster than proof of value. CIOs are telling Google’s Sundar Pichai they are “so concerned” about how fast AI budgets are being burned, and he expects the problem to worsen through the year. Uber’s leadership now calls those AI bills “harder to justify,” while one consultant highlighted by Axios described a client that spent half a billion dollars in a single month after failing to set usage limits on Claude licenses. The move from flat, subscription-style pricing toward a la carte AI compute has amplified the shock. Some enterprises even discovered employees using powerful models to check the weather, an expensive way to answer routine questions. As sticker shock spreads, finance leaders are pushing teams to move from “as many tokens as possible” to “only the tokens that pay their way.”
Agentic AI and Coding: Where the Bills Are Biggest
For now, AI’s clearest product–market fit inside large companies is in coding. Agentic AI coding tools can generate pull requests, refactor legacy systems, and propose fixes, but they are also responsible for a large share of enterprise AI spending. Salesforce’s early budget for agentic coding turned out to be “an almost absurd underestimate,” and Axios reports that Microsoft canceled most of its Claude Code licenses in part over costs. Even where tools do help, returns are uneven. A Jellyfish report found that the top 10% of Claude Code users consumed about 10 times as many AI tokens as the median developer while producing only about twice the output. That gap undercuts the assumption that letting agents run wild will automatically pay off and highlights why enterprises now scrutinize agentic AI costs instead of celebrating raw consumption.
From Leaderboards to KPIs: Designing ROI-Focused AI Use
The tokenmaxxing backlash is pushing enterprises toward more disciplined AI strategies. Instead of celebrating token leaderboards, teams are tying compute to specific business metrics: revenue impact, cycle-time cuts, defect rates, or ticket volume reductions. Jellyfish recommends that organizations neither reward nor punish raw token use, but rather attach costs to concrete outcomes. Executives are also rethinking which problems deserve AI at all. Many workers default to automating tasks they dislike, not the ones most valuable to the company, and a “thousand flowers bloom” approach to licenses has failed to create reliable productivity gains. Meanwhile, concerns over data access limit how effective agents can be outside coding. For leaders shaping next year’s AI budgets, the lesson is clear: define a small set of high-value use cases, set hard caps on AI token costs, and demand evidence of sustained ROI before scaling.