1) they're not all using it at the same time 2) you must likely CAN run a GPT-4 equivalent model locally 3) still requires a lot of engineering but it's a field that has grown a lot since the cloud era