As foundation models converge in capability, AI startups face a brutal question: who actually has a defensible moat when raw model performance no longer separates winners from losers?