I noticed something odd a few months ago. Several engineers I respect — people building serious AI pipelines, not hobbyists — quietly shifted from API-based inference back toward running models locally. Not because of some principled stance. Not because they read a blog post. Because they hit real problems and local inference solved them faster...