>> You can either pay one person (out of a - growing- pool of tens of millions of willing drivers) ~$15 an hour (depending on geography) in a $20k car or you can pay very little per hour in a 250k for the car + a gigantic engineering team + compute infrastructure. It can make sense on a large scale, but it'll be hard to do.

I believe this $250K figure will end up being off by a multiple of 2. As an example, the cost of LIDAR (the most expensive piece of many SDC's) has dropped by an order of magnitude over the past couple years. (A quick google search for a source on this claim yields: https://qz.com/924212/what-it-really-costs-to-turn-a-car-int...)

The biggest questions around SDC's seems to really be around safety & feasibility. If these criteria are met, I can't imagine the overall cost per mile driven will be higher than manual labor.

And, even if autonomous driving doesn't undercut on price, they will beat human drivers on both quality & convenience. Most people don't enjoy having a human driver who sometimes smells bad and tries to talk to you (not to mention the uptick in assault reports). And, since the cost of under-utilization will be much lower for SDC's, you can bet the average wait time for SDC's will be lower than Uber.

>> There's a general theme here: It's not enough to teach a computer to do something as good as a human, you have to undercut on price as well.

I think the more apt statement is: Consumers care about price, quality, safety, and convenience. A winning solution has to be significantly better in at least one of these categories (price being only one possibility).