Developers at Meta, Microsoft, and Salesforce are deliberately burning tokens to inflate AI usage metrics. The behavior, called 'tokenmaxxing,' treats token consumption as a performance target rather than a cost to control. Engineers are spending real money to hit numbers that signal AI adoption, regardless of whether the underlying work requires it.
The economics are already breaking. Anthropic pulled enterprise plan subsidies. Uber burned its entire 2026 AI token budget in three months. The Pragmatic Engineer predicts per-engineer token budgets will become standard across large companies, fast. These are structural corrections, not edge cases.
The full piece also covers Claude's reported quality degradation, Cal.com moving code to a closed repo and whether AI was the real reason, Vercel open-sourcing an agent orchestration tool, and new AI usage guidelines landing in the Linux kernel. The tokenmaxxing section alone is worth the read for what it reveals about how organizations are measuring, and mismanaging, AI ROI.
[READ ORIGINAL →]