I think most people are oversimplifying how much compute even matters.
What really matters is the sensor. Trains for example, know where all other trains are thanks to the modern Positive Train Control network: GPS, electrical connections on rails ("3rd rail" detection), speed sensors, the whole gamut. Positive Train Control was legally enforced starting in 2008 (though it wasn't fully deployed until last year), long before the GPU-boom gave rise to modern AI engines.
You can't "compute" around shitty 1.2MP cameras as your sensor. Step one is having a good enough sensor to do self-driving / crash avoidance / etc. etc. If you don't have a sensor that can accomplish the goal, you're already stranded. On the other hand: knowing the speed, direction, and location of all trains in your area (Positive Train Control network: the safety mechanism behind Amtrak / Freight Trains in USA), gives you all the information such that a simple "dumb" computer can prevent collisions.
---------
That people are trying to solve the self-driving (or crash-avoidance) car problem with such terrible sensors today is the crux of the problem. I get it: computers are cheaper than good sensors. But this route is so much harder and less reliable.