A startup is choosing between GPT-4 (closed API) and LLaMA-3-70B (open weights) for their healthcare analytics product. Their CTO lists four requirements: (1) patient data must never leave their private cloud, (2) they need to fine-tune on proprietary clinical notes, (3) they need guaranteed API uptime SLAs, (4) inference cost must be predictable. Which requirements favor open weights and which favor closed API, and what is the decisive factor for healthcare?