Major technology companies including Amazon, Meta, and Uber are dismantling internal AI usage leaderboards after discovering that employees were artificially inflating token consumption to improve their rankings, a phenomenon known as "tokenmaxxing" that cost companies billions in unnecessary computing expenses. The shift signals a critical reckoning in how enterprises measure AI productivity and value.

What Is Tokenmaxxing and Why Did It Become a Problem?

A token is the smallest unit of data that an artificial intelligence model processes. When an AI system reads text, it breaks information down into individual characters, words, or punctuation marks and treats each as a token. During inference, when an AI agent performs a task, it can consume thousands of tokens per request.

Tokenmaxxing refers to the practice of maximizing token usage simply to produce higher numbers on performance metrics, regardless of whether that consumption actually contributes to meaningful work. The trend gained momentum after high-profile executives, most notably Nvidia CEO Jensen Huang, publicly advocated for aggressive token spending. In a March podcast appearance, Huang suggested that a capable engineer earning $500,000 annually should spend at least $250,000 on tokens by year-end, comparing the avoidance of AI to "using paper and pencil to design chips".

Huang, publicly advocated for aggressive token spending

How Did Amazon's Kiro Rank Become a Case Study in Misaligned Incentives?

Amazon developed an internal AI code generation tool called Kiro and created a leaderboard system called "Kiro Rank" that tracked and scored employees based on how frequently they used the tool. The company set an ambitious target that more than 80 percent of developers should use AI every week.

The system backfired spectacularly. Employees, motivated to improve their rankings and evaluation scores, began instructing AI agents to perform unnecessary tasks simply to consume more tokens and climb the leaderboard. This artificial demand inflated Amazon's computing costs without generating proportional business value. Dave Treadwell, Amazon's senior vice president of engineering, acknowledged the problem, stating that while the leaderboard was created with good intentions, tokenmaxxing had become counterproductive. He urged employees to stop using tokens merely for the sake of appearing productive.

Amazon ultimately scrapped Kiro Rank and shifted to a new metric that measures not just how much employees use AI, but how consistently they use it to develop genuinely useful code.

Which Other Tech Giants Abandoned Their Token Leaderboards?

Amazon was not alone in recognizing the problem. Multiple enterprise technology leaders have dismantled similar systems:

• Meta's Claudenomics: The company tracked token usage across 85,000 employees through a leaderboard that awarded titles like "Immortal" and "Token Legend" to top performers. As token consumption became the goal rather than a means to an end, employee fatigue mounted. Monthly token usage spiraled to 60 trillion tokens, with the top-ranked individual alone consuming 281 billion tokens. The employee who developed Claudenomics ultimately shut down the program.
• Uber's Budget Crisis: After investing heavily in Anthropic's Claude Code tool, Uber exhausted its entire annual AI budget in just four months. The company discovered that despite high token consumption, productivity gains were not clearly apparent, prompting management to reassess how it measures AI tool value.
• Salesforce's Shift to Task-Based Metrics: Rather than counting tokens, Salesforce introduced a new measurement called an agent work unit (AWU), which evaluates how many actual tasks an AI agent completes rather than how many tokens it consumes.

How to Measure AI Productivity Without Gaming the System

The collapse of token-based leaderboards has forced enterprises to rethink how they evaluate AI tool effectiveness. Rather than counting inputs, companies are now focusing on outputs and business outcomes:

• Task Completion Metrics: Measuring the number of meaningful tasks completed by an AI agent, rather than the volume of tokens consumed in the process, provides a clearer picture of actual productivity gains.
• Code Quality and Utility: Amazon's new approach evaluates whether employees consistently use AI to generate useful code, not simply whether they use it frequently, aligning incentives with business value.
• Return on Investment Tracking: Companies like Uber are now scrutinizing whether AI tool spending translates to measurable productivity improvements, preventing budget overruns and wasteful consumption.
• Behavioral Consistency Over Volume: Focusing on regular, purposeful AI usage patterns rather than peak consumption numbers encourages sustainable integration of AI into workflows.

What Does This Mean for the Future of AI Tool Adoption?

The tokenmaxxing backlash reveals a fundamental challenge in enterprise AI adoption: measuring productivity in an AI-driven environment is far more complex than simply tracking tool usage. Executives like Uber's Chief Operating Officer Andrew Macdonald have begun publicly questioning whether token-heavy approaches actually deliver value.

The shift away from leaderboards and token-counting metrics suggests that mature AI adoption will require more sophisticated measurement frameworks. Companies are learning that aggressive token consumption can mask inefficiency and waste, while meaningful AI integration depends on aligning tool usage with genuine business outcomes. As enterprises continue deploying AI code editors and extensions like Amazon Q Developer, Tabnine, JetBrains AI Assistant, and Zed Editor's built-in AI features, the lesson from Amazon, Meta, and Uber is clear: how you measure success matters as much as the tools themselves.