From Tokenmaxxing Hype to Hard Questions on Productivity
Tokenmaxxing is a tokenmaxxing strategy in which companies drive up AI token usage across tools and agents in the hope that higher model activity will translate into faster development cycles, more shipped features, and visible productivity gains, even though the business impact of this additional token consumption remains hard to measure in a consistent and reliable way. For a while, this mindset spread through large engineering teams. Internal AI dashboards rewarded usage, and some firms bragged about massive token flows as a badge of modernity. Visa, for example, has publicly highlighted that its monthly token spend nears 2 trillion tokens, while other corporations track how often staff call internal AI tools. Yet AI infrastructure costs are rising much faster than clear output. As Uber’s early internal spend shows, it is easy to blow through an annual AI budget in months without clear evidence of lasting productivity gains or new revenue.
Uber’s Public Break with the Tokenmaxxing Strategy
Uber COO Andrew Macdonald has become the most visible critic of tokenmaxxing so far, giving voice to concerns many engineering leaders share in private. In a widely shared interview, he said he has not seen a direct line from higher AI token use to meaningful productivity improvements. His comment that “that link is not there yet” spread quickly among developers who feel pressured to increase usage. Uber’s situation is a cautionary tale: the company reportedly exhausted its AI budget within the first four months of the year, underscoring how fragile current enterprise AI spending models can be. At the same time, engineers and founders complain that “tokens got burned for millions of dollars without any real significant ROI to show for it.” As AI infrastructure costs mount, executives are less willing to accept token graphs as a proxy for performance and are looking instead for hard measures tied to products and shipping velocity.
Agentic Coding Meets the Reality of AI Infrastructure Costs
While tokenmaxxing loses appeal, many firms are betting on agentic coding: multi-step AI agents that write, review, and refactor code. Salesforce has rolled these agents through its engineering teams, only to discover that its initial token budget was “an almost absurd underestimate.” That gap between forecast and usage reflects a wider pattern: agentic systems often chain many calls to large models, multiplying AI infrastructure costs. The promise is compelling—faster feature delivery, fewer bugs, and better use of senior engineers’ time—but the agentic coding ROI story is still unsettled. Data from Jellyfish shows 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. According to Jellyfish, rewarding raw token consumption is a mistake; companies should instead tie costs to concrete engineering metrics such as pull requests and merged changes.
Redefining ROI: From Token Volume to Business Outcomes
The industry debate is now shifting from how many tokens agents use to when and how to measure their impact. CIOs, according to Google CEO Sundar Pichai, are “so concerned about how much their companies are blowing through budgets,” and he expects the problem to intensify. Enterprise leaders want a clearer model: which projects, teams, and workflows convert token use into durable business value. This means new rules for agentic coding ROI. Instead of celebrating rising token charts, organizations are starting to ask how AI agents affect cycle time, deployment frequency, defect rates, and customer metrics. Jellyfish recommends tying agent costs to specific outputs like pull requests rather than abstract usage. The emerging consensus is that token efficiency matters: higher token usage does not automatically translate into better outcomes, and unchecked tokenmaxxing may be a “crazy, rushed, temporary phase” rather than a stable path to competitive advantage.