Meta's 73 trillion token bill and the quota problem managers already know how to solve

Meta 的 73 万亿 token 账单，和一个管理者早就会解的问题

Meta's internal leaderboard Claudeonomics tracked ~85,000 employees' token usage, hitting 73.7 trillion tokens in 30 days—billions of dollars. Uber burned its full-year AI coding budget in four months after giving 5,000 engineers Claude Code. The subsidy cycle is ending: Claude Code's $200/month subscription masks heavy-user costs of ~$5,000/month, roughly 25x the subscription price. Meta's June memo set 2027 as the year for structured token budgets and allocation tools. The article maps AI cost management to four management moves: model routing instead of tiered staffing, context engineering instead of bounded scope for new hires, prompt caching instead of codifying SOPs, and measuring output instead of token count. Jellyfish's analysis of 12,000 developers found the heaviest users burned 10x tokens per PR with only 2x throughput; per-PR cost jumped from $0.28 to $89.32 with no quality gain. Bosworth championed unlimited token burning in April, then wrote in June that token usage alone is not a measure of impact of any kind.

Why it matters: 73.7T tokens, 25x subsidy multiplier, Uber blowing its annual budget in four months — three concrete numbers that nail the end of the AI tool subsidy cycle. Not scoring higher because the article body is truncated mid-argument, and some figures come from third-party estimates ...