You're always going to require a GPS connection or the next-generation equivalent sometime anyway.

No, that's exactly my point: a self-driving car should be a human replacement, not your GPS that happens to also control the wheel.

GPS'es are loaded with data, too, whether it is pre-loaded or daily updated. The way this system seems to work is that it requires more detailed geographical data, which becomes ever more feasible in the near future. It's just a box, like a GPS, that uses a database of locational data to know where it is and make assumptions based on that data. Only now the assumptions aren't 'where should I go' but 'where do I steer this thing'.

I understand that the way they do it now is a good stepping stone, but Google being Google, I am afraid that when as soon as they can they will mass-produce these cars and just make money off that. It's a reasonable business case. It is not, however, a real human-AI-like self-driving car.

and I can tell you that housing all of the spatial data, regardless of format, is a LOT of disk space.

Which is exactly the problem. Normal humans don't depend on a colossal GIS/CAD database to find their way. A normal driver should be able to find its way reading signs and having a general idea of the route to take. I don't see why a self-driving car shouldn't do the same. If self-driving cars can operate without needing preloaded data, you don't need all that disk space at all: just lots of realtime calculations. All the car needs to know for more than a couple of seconds is where to go and the implied road speed. The rest should be handled by the car's AI. That's how I envision self-driving cars.

Of course it is difficult, and I'm not saying it should be available now, but I don't see Google going down that path if this trick will put them far ahead of competitors and make them tons of money.