Amazon Shuts Down Its Internal Token Consumption Ranking, Warns Workers Against Using AI for the Sake of It
Miles Bennett
Amazon killed an internal AI-usage leaderboard after employees turned token consumption into a performance metric; the episode surfaces an industry-wide problem — more AI spending does not automatically mean more output.
What did Amazon shut down?
Amazon told staff this week that KiroRank, an internal leaderboard, is no longer active. Built by employees, it tracked how many AI tokens each person consumed.
The problem → some employees chased higher rankings by running AI agents indiscriminately, turning token burn into a goal in itself. Volume replaced value.
An Amazon spokesperson confirmed the leaderboard has been "deprecated," saying it "was never intended to encourage using AI for the sake of using AI."
What did leadership say?
Amazon SVP Dave Treadwell told employees this week: "Please do not use AI for the sake of using AI."
He added that AI should help solve customer problems and business problems and drive innovation. In plain terms = AI is a tool, not a timecard.
Amazon says it does track token usage to measure cost, but explicitly rejects "token maximization" — equating AI productivity with the volume of tokens consumed.
Does using more AI produce more output?
Enterprise-management platform Jellyfish offers a clear answer: the top 10% of Claude Code users consumed roughly ten times the tokens of an average developer — but produced only twice the output.
In absolute terms, heavy users burned 225 million tokens per week; average software engineers used 32 million.
This means → marginal returns are collapsing. Five times the token spend buys only one extra unit of output.
Did other tech giants hit the same wall?
Meta ran a similar internal leaderboard called "Claudeonomics." Data showed employees consumed 60.2 trillion tokens in 30 days — roughly $900 million at Anthropic's public API pricing.
An insider said output from the top of the leaderboard was mostly "throwaway junk." Meta subsequently shut the leaderboard down.
Uber COO Andrew Macdonald said the company has not seen productivity gains matching its increased AI investment. Uber's CTO had earlier disclosed that by April the company had already exhausted its annual Claude Code budget.
What does this reflect?
This reflects an awkward phase in enterprise AI adoption: the tools are in place, but usage has gone sideways. When "how much you used" replaces "what you solved" as the metric, AI spending becomes overhead.
Amazon, Meta, and Uber all hit the same wall — this is not a one-off management lapse. This means → the industry needs to shift from "encourage adoption" to "measure outcomes."
Put simply = buying AI is easy. Knowing how to spend on it productively is the real challenge.
Please do not use AI for the sake of using AI.
Dave Treadwell
Amazon Senior Vice President
(internal remarks this week)
Content is for reference only, not financial advice.