The 4am answer

A mother in Mymensingh wakes at 4am with her child burning up. She opens the wallet on a shared phone and asks, in her own Bangla, whether this is something to watch through the night or something to move on now. An answer comes back: the fever pattern, the danger signs to check, whether the pharmacy will do or a hospital is needed now. For that one minute, the most important thing in her house is a model's reply.

Today that reply travels to a server in another country and back. Most nights it works. The question sovereignty asks is not whether it works most nights. It is who decides whether it works at all.

A foreign frontier model is a product, sold under terms the buyer does not write and cannot appeal. The price can move. The rate limits can tighten. Access can be cut over a sanctions reading, a policy change, or a billing dispute between two companies on the other side of the planet. None of those events have anything to do with the mother in Mymensingh, and every one of them can reach into her 4am.

A health system is allowed to depend on electricity, on roads, on the mobile network, because those are things the country can, in principle, govern. It should not depend, at the layer that decides whether a sick child is sent to hospital, on a contract it is not a party to. That is the difference between infrastructure and a subscription.

This is why the intelligence has to come home, even where the foreign model is, today, better. Better is worth a great deal. It is not worth building the country's 4am on a foundation someone else can remove. A model that runs on Bangladeshi servers can be a step behind the frontier and still be the right one, because it answers to the country that depends on it. Availability you control beats capability you rent, at the layer where someone is waiting for an answer in the dark.