Amazon shuts down internal AI leaderboard amid token‑maxxing concerns

Amazon has removed an employee‑run dashboard that tracked AI token consumption, after the company said the tool encouraged “tokenmaxxing”—inflating AI usage for ranking purposes rather than solving real problems [1]. Senior Vice President Dave Treadwell warned staff to stop using AI “just for the sake of using AI,” and the internal leaderboard, known as KiroRank, was officially deprecated [1].

Key takeaways

• Amazon’s KiroRank leaderboard was an informal tracker created by employees and was not an approved tool [1].
• The company cited rising AI computing costs and “tokenmaxxing” behavior as reasons for shutting down the dashboard [3].
• Amazon is shifting to a metric called “normalised deployments” that focuses on useful AI‑driven work rather than raw token counts [3].
• Similar cost pressures are reported at other tech firms, with Microsoft cutting Claude Code licences and Uber exhausting its AI budget by April 2026 [2].
• Amazon continues to monitor AI token usage for cost purposes but discourages usage purely for metric‑driven competition [1].

Internal leaderboard and the rise of tokenmaxxing

The KiroRank dashboard was built by a group of Amazon engineers to highlight how AI could accelerate work, but it quickly became a competitive ranking system that measured employees by the number of AI tokens they consumed [1][3]. Tokens are the basic units that large language models process, and their usage has surged in 2026 as agentic AI tools operate with minimal human oversight [1]. Employees reportedly used MeshClaw, an internal AI agent platform, to generate extra token consumption, inflating the leaderboard and driving up computing costs [2].

Amazon’s leadership responded by deprecating the beta dashboard, emphasizing that AI adoption should be measured by operational efficiency and innovation rather than raw usage numbers [1]. The company’s statement noted that teams retain freedom to decide how to use AI tools, but the organization now tracks “normalised deployments” to assess whether engineers are delivering meaningful code and business outcomes with AI assistance [3].

Industry‑wide cost pressures and shifting metrics

Amazon’s move mirrors broader industry concerns about AI spending. The Financial Times reported that Microsoft has cancelled most of its Claude Code licences after token‑related expenses exceeded expectations, while Uber’s COO admitted the rideshare giant had burned through its entire 2026 AI budget by April without delivering significant consumer‑facing features [2]. These examples illustrate a growing skepticism toward the assumption that higher AI token consumption automatically translates into productivity gains.

Anthropic, the provider of Claude models used by both Amazon and Uber, has shifted to consumption‑based pricing, further increasing cost visibility for corporate users [3]. Some analysts suggest that, in certain scenarios, hiring skilled human workers may be more cost‑effective than relying heavily on AI coding tools [3].

Why it matters

The shutdown of KiroRank underscores a pivot in how large tech firms evaluate AI success: from quantity‑based metrics like token counts to quality‑focused measures such as “normalised deployments.” This shift aims to align AI usage with tangible business outcomes and curb spiralling compute expenses. As companies continue to invest heavily in AI infrastructure—Amazon is projected to spend nearly $200 billion on capital expenditures this year, much of it on AI and data‑center capacity—the emphasis on cost‑effective, purposeful AI adoption is likely to shape internal policies and industry‑wide best practices in the coming months.