The "tokenmaxxing" wave that kept artificial intelligence (AI) developers awake at night has hit the brakes. Big Tech companies that had pressured developers to consume more tokens in order to win the AI war are now scrapping their leaderboards, saying they lack substance. Developers that had evaluated employee performance based on token usage expressed regret, saying that token consumption had become a mere display and was setting productivity back.
The Financial Times (FT) reported on the 28th (local time) that Amazon has halted "Kiro Rank," which tracked employees' AI tool usage. Kiro is an AI code generation tool developed in-house by Amazon. The system that assigned scores and ranked employees based on how much they used Kiro has been abolished.
The FT explained that the decision was made because employees engaged in unnecessary activities to boost their scores, increasing the company's computing costs. Employees were forcibly consuming tokens to raise their evaluation rankings, even when it was not essential for AI development or their work. Instead, Amazon began using a metric that measures not just AI usage but how regularly an employee uses AI to develop useful code.
A token is the minimum unit of data that an AI model processes during learning and inference. The AI breaks down information into characters or punctuation marks and then recognizes them as tokens. When inference is carried out with AI agents, thousands of tokens are used per task unit. Tokenmaxxing refers to the phenomenon of maximizing token usage to produce results in AI development or work processing.
The management leader who most emphasizes tokenmaxxing is Jensen Huang, the chief executive officer (CEO) who leads Nvidia, the world's largest company by market capitalization. Appearing on a podcast in March, he stressed that a capable engineer should spend half their salary on purchasing tokens, saying: "Not using AI is like using paper and pencil to design chips. If an engineer earning $500,000 (about 752 million won) a year hasn't spent more than $250,000 on tokens by year-end, I would be very worried." He even proposed giving every engineer an annual "token budget" on top of their base salary. Some interpret the background of CEO Huang's advocacy of token usage as ultimately a strategic move to expand Nvidia's revenue.
Although Big Tech companies, including Nvidia, led the way in tokenmaxxing, it produced side effects contrary to their intentions. The latest Amazon case is a representative example. Amazon pressured developers by setting a target that more than 80% should use AI every week, but employees instructed AI agents (assistants) to perform unnecessary tasks to use up tokens and raise their own evaluation scores. The FT, citing sources, reported that Dave Treadwell, Amazon's senior vice president of engineering, said the leaderboard was created with good intentions but that tokenmaxxing, which inflates token usage, increased the cost burden, and that he urged employees not to use tokens simply for the sake of using AI.
For the same reason, a growing number of companies are eliminating token ranking competitions. Meta, the operator of Facebook, tracked the token usage of 85,000 people through a leaderboard called "Claudenomics" and singled out the top 250. Employees competed to earn titles such as "Immortal" or "Token Legend," but as token consumption itself degenerated into the goal, voices complaining of fatigue erupted. As waste worsened to the point where monthly token usage exceeded 60 trillion and the top-ranked person alone used 281 billion tokens, the internal employee who developed Claudenomics shut down its operation.
Skepticism about tokenmaxxing is also growing at ride-hailing company Uber. After pouring tokens into the use of Anthropic's coding tool "Claude Code" this year, Uber used up its annual AI budget in just four months, and warning lights turned on for return on investment (ROI). Enterprise software company Salesforce moved away from the method of simply counting the number of tokens and introduced the concept of an agent work unit (AWU), which evaluates how many tasks an agent has handled.
Andrew Macdonald, Uber's chief operating officer (COO), appeared on a podcast and expressed concern, saying: "The technical research team is using AI coding services, but the productivity gains have not been clearly apparent," and "Management has begun discussing how to manage the cost of token consumption."
