Self-driving cars don't really need to talk to "the cloud" much, if at all. They need some map data, and they might contribute to map data. They can use some traffic data, and can contribute to traffic data, which needs about as much data as Waze. None of this is real-time; it can be seconds or minutes behind.

Transmitting all the car's imagery to servers is an R&D activity. Google downloads all that from their vehicles to train their machine learning systems, but training is an offline process. Production vehicles might occasionally upload "interesting" imagery they didn't recognize, or data from situations where there was trouble, but there's no need to upload it all.