You are not wrong, except at scale it gets complicated quickly. For starters, to support large user numbers, you’re going to have to process your own grib2 data for radar and turn them into tiles at zoom levels.

It takes about 24 cores with a GPU to do CONUS, Canada, Alaska, Pacific and Caribbean data. This should be 2x for redundancy. Even being cheap with main processing in my basement (gen power, backup internet) the cloud costs to serve it are $200 month plus data transfer. The standby grib machine spins up should it not see the cheap primary or the NOAAPort receiver is offline.

There is no money to be made without whoring out your user’s privacy. People just won’t pay for a privacy focused weather app. I keep this going as a hobby.