Actually, they don’t. CPU power and memory is cheap. No reason for a learning thermostat or controller to need to communicate with the Borg Collective, back and forth.
All data that it needs to develop its program is available locally or via the user’s phone.
IPv6 (practical meaning: static IP ranges for all homes) is coming. There shouldn’t be any need for an intermediary (other than data theft for profit) once every home is routable. Phone can talk directly to a thermostat, possibly via a VPN.
I’d be willing to pay say 2-3x the price of a Nest for a thermostat that doesn’t engage in data rape.