A CFO built and shipped a batch job in 2 days with AI coding—then burned a month's server budget in one day
一天花掉一个月服务器费:AI 编程民主化与消失的生产门禁
A CFO used Claude Code to build a batch job in two days and deployed it straight to production. A deployment ordering mistake, automatic retries on a deterministic failure, and non-idempotent LLM calls caused the job to rerun 21 times in a single day—each time paying for the same model calls again. The API cost spike exceeded the company's entire monthly server bill. The post argues this isn't a reason to reject Vibe Coding; it's a sign that infrastructure hasn't yet built circuit breakers for non-engineers. Cost must become a first-class safety metric, paid API calls need default idempotency caching, and risk-based review gates are overdue.
Why it matters: A finance lead's AI-written batch job went straight to production, hit a missing-column error, and the queue retried it 21 times, each retry calling the paid model API and dropping the result on write.