I have a feeling that weather data compresses together well — especially if stations 1. are ordered along a space-filling curve geographically, with read-outs of geographically-neighbouring stations in sequence; and 2. are stored column-oriented where each future time is its own column. Basically the same optimizations as you would use to store any other IoT time-series sample-cloud data.

I suspect that, encoded this way, weather for 4000 locations wouldn’t be much bigger than weather for 1 location.

Also keep in mind that half the reason the Weather app needs to use as much bandwidth as it already does, is that whenever you move even slightly, it can no longer give you accurate info without grabbing the point-forecast again for your new location. It uses far less data if you just stay at home all day every day. If it had a pre-cache of forecasts for the entire local area, it wouldn’t need to do that; it could just refresh once every few hours. Maybe even wait for you to be on wi-fi before doing so, if it has modelled you as usually connecting to wi-fi several times daily.