Fortune magazine commented that the once-popular practice of "tokenmaxxing" within enterprises is cooling down. Tokenmaxxing uses the number of tokens used by employees or teams to access AI models as an approximate indicator of innovation and work efficiency. However, with rising costs and increasing invalid calls, more and more companies are tightening their grip on this practice.
The article mentions that companies like Meta, Amazon, and OpenAI have previously used formal or informal token ranking practices to encourage engineers to compete on model usage volume. The problem is that once the metric itself becomes a performance evaluation target, it easily deviates from its original purpose. The Financial Times previously reported that some Amazon employees had AI agents perform tasks with no real substance, solely to maintain performance data.
Cost pressures are beginning to emerge.
As generative AI is deployed on a large scale within enterprises, the cost of model access is rising rapidly. The article states that some companies have begun restricting employees' use of third-party AI agents, especially tools that rely on high-end models. Meta has removed an employee-created token leaderboard; The Verge reports that Microsoft has canceled Claude Code subscriptions for employees in several key product divisions.
Uber also disclosed that it had already used up its entire annual token budget in the first four months of 2026, with some of the expenditure coming from the high-frequency use of Claude Code. Salesforce CEO Marc Benioff stated that the company paid approximately $300 million to Anthropic this year and hopes to have a smarter routing system in the future to allocate different requests to more cost-effective models.
Businesses place more emphasis on business results
The article argues that the core reason companies are tightening their token metrics isn't just cost-cutting, but also the gap between investment and return. Uber's Chief Operating Officer, Andrew Macdonald, recently stated that the company finds it difficult to directly translate improvements in the efficiency of individual employees into new user-facing feature delivery or overall business results. Without clear business outcomes, it's even harder to sustainably justify the cost of the model.
This is why simply tracking token consumption is increasingly less of an effective management tool. It can reflect the scale of calls, but it cannot indicate whether those calls have actually improved the product, processes, or revenue.
The real rewards come from process reengineering
The article cites Azeem Azhar, author of Exponential View, who argues that the current mismatch between AI investment and productivity resembles a "productivity J-curve" common in the early stages of a new general-purpose technology. Companies often increase experimental costs during the exploratory phase, but see no significant benefits in the short term; efficiency improvements only become apparent after business processes are redesigned.
The article uses the example of power system upgrades in factories to illustrate that while companies initially only replaced lighting or power sources, truly significant productivity improvements occurred after the factory layout and individual equipment were restructured around new technologies. Similarly, with AI, many companies are currently still in the stage of partial trials or tool aggregation, and have not yet entered into deeper process transformation.
Commentators argue that the token usage race has subsided because it focused on "how much was used," rather than "what was created." For businesses, the value of AI ultimately lies in product delivery, business models, and revenue performance, not in model usage rankings.
