THIS ANALYSIS IS POOR.

Let me spend 10 min and try to do better:

If Robo Taxis win then all the fixed costs (like training models) end up being a rounding error as the scale of the industry is 10,000x larger 1,000,000x than is today.

Now of course the variable costs associated with each car will be very relevant. But your discussion of the cost of the physical car is besides the point. Yes both human driven and machine driven cars will keep getting safer, but any costs here are a wash, and can be discarded from our reasoning.

The leaves the sensor suite and compute costs. This is a very real cost per car, and could sink the industry if its amortized costs was greater than a human drivers cost. I don't have those numbers, but I have to believe VCs that are dropping BILLIONS into self driving cars DO have those numbers. So my default assumption would be that the price comparison for the segment of the costs (which presently are not the dominant cost) look favorable.

And you refer to "diminishing returns" in savings. But nothing could be further from the truth in the crucial variable cost human-machine differentiation section. Those costs are dominated by compute which is still on a Moore's law progression of exponentially cheaper (though a lower exponent these days). And the dominant cost sensor (the lidar) has made like a >>100x price reduction in the last decade and is still dropping like a ROCK!

So this arm-chair blogger is calling bullshit on your arm-chair assessment of the industry.

(All this said, I do expect that the MASSIVE MULTI-BILLION dollar losses will continue for some years, but this will only heighten the moat that Waymo etc. will have when it is profitable.)