Amazon has shut down an internal AI usage leaderboard after employees inflated their scores to increase rankings, driving up computing costs.

The move highlights the financial risks associated with unregulated AI adoption within large corporations. As companies rush to integrate generative AI, the cost of compute power can escalate quickly if usage is not tied to productive output.

The internal tool, referred to as “KiroRank” or the “token leaderboard,” tracked the amount of AI tokens employees consumed. However, the company discovered that staff were engaging in “tokenmaxxing” — the act of inflating token usage to climb the leaderboard [1, 3]. This behavior led to a surge in compute costs without a corresponding increase in business value [2].

An Amazon spokesperson said, “Don’t use AI just to use AI” [1].

Internal communications reported by CNET said that the leaderboard existed for only a few weeks [3] before the company decided to retire it [3]. The short lifespan of the project underscores the volatility of incentivizing raw usage metrics in AI development.

Amazon is now shifting its focus toward “normalized deployments” [1]. This new approach aims to measure the amount of useful code produced rather than the raw volume of tokens consumed [1]. By prioritizing utility over activity, the company hopes to curb wasteful spending on cloud infrastructure.

Industry experts suggest this is a cautionary tale for the broader corporate sector. One analyst said, “Companies are being warned about rising AI costs, and this is a clear example of why that matters” [2].

““Don’t use AI just to use AI.””

This incident illustrates a fundamental tension in the current AI boom: the gap between adoption metrics and actual productivity. When companies reward 'usage' as a proxy for 'innovation,' they risk creating perverse incentives that inflate operational costs. Amazon's pivot to measuring useful code suggests a broader industry trend toward 'AI efficiency' rather than raw integration.