The advice to "build what OpenAI won't ship" is easy to endorse and hard to execute — because the genuinely defensible niches are the ones that require domain knowledge, compliance infrastructure, or institutional relationships that a general-purpose lab cannot acquire cheaply. The Bluesky post [1] that catalyzed this conversation framed it as a rule: if the one-line pitch could become a ChatGPT feature next week, the company does not exist yet. What that rule demands in practice is that founders stop optimizing for model capability and start optimizing for the…