Amazon and Uber Reassess AI Token Usage

Amazon deprecated an employee-built internal leaderboard called KiroRank that tracked AI token usage, after employees used the company's MeshClaw agent tool to run trivial, unnecessary tasks purely to inflate their rankings, according to the Financial Times, Fortune, and subsequent coverage in Business Insider, TheStreet, and CNET. Amazon SVP of engineering Dave Treadwell told staff, "Please don't use AI just for the sake of using AI," and the company has restricted visibility of team-wide usage statistics and shifted to a "normalized deployments" metric that measures whether AI-generated code produces real outcomes, such as successful commits, rather than counting tokens. Fortune reports Meta ran a comparable internal leaderboard nicknamed "Claudeonomics" that ranked roughly 85,000 employees by token consumption, and pulled it within days of the ranking becoming public. Separately, Fortune and Investing.com report Uber exhausted its entire 2026 AI token budget within the first four months of the year, driven heavily by use of Anthropic's Claude Code at reported per-engineer costs of $500 to $2,000 a month; Uber has since imposed a $1,500-per-person monthly usage cap per tool.

Tokens are the billing unit for large language model usage, and agentic workflows, where an AI agent chains multiple actions together, can consume dramatically more tokens than a single prompt-and-response interaction; reporting describes agentic workflows using up to roughly 1,000 times more tokens than standard interactions. Salesforce CEO Marc Benioff is quoted in Fortune describing a "smart router" concept that would direct queries to cheaper models when a task doesn't require the most capable, and most expensive, model, an approach several companies are now pursuing to control costs.

Analyst Gil Luria of D.A. Davidson has warned that gamified usage metrics like KiroRank cast doubt on the demand signals used to justify AI infrastructure spending, invoking Goodhart's Law: once token usage became a target, it stopped being a reliable measure of productivity. With combined 2026 AI capital expenditure from Amazon, Microsoft, Alphabet, and Meta already approaching $700 billion, the practical lesson for practitioners and finance teams is to instrument outcome metrics, such as time saved, error rates, and completed-task throughput, rather than raw token or usage counts, and to expect vendor-side model-routing tools to become a standard cost-control lever.

• •Whether other large AI spenders, reporting already names Walmart, Cisco, and Meta, formalize per-team token budgets or usage caps similar to Uber's.
• •Adoption of model-routing or capability-routing tools that steer low-complexity queries to cheaper models.
• •Whether analysts revise AI capex demand estimates if more companies disclose that reported "usage" partly reflected gamed internal metrics rather than organic productivity gains.

This is a textbook Goodhart's Law failure: incentive systems built to encourage AI adoption instead optimized for the metric itself. The bigger open question is how much of the AI industry's broader usage and capex narrative rests on similarly ungoverned internal metrics, a concern serious enough that a Wall Street analyst is now raising it publicly rather than treating it as an internal HR curiosity.