关于运行 AI 基础设施的笔记，以及搞错它的代价。
Cache multipliers vary 20x across models, and most cost tools treat them as a constant. The number stays plausible, which is why nobody catches it.