1)Cost will not be the same if you don't have to buy a new computer.
2) Not on USB 3.0 (5 to10 gigabits per second), and usually not on 2.0 (480 megabits per second). Although 2.0 only allows 1 way communication at a time, while 3.0 allows 2 way in parallel.
3) I have no idea where you're getting the latency argument from, care to explain?
Just doing some quick research on this, I've found that people have already jerry-rigged something
exactly like this themselves sort of. Apparently USB 3.0 is a WIP (a good enough PCI to 3.0 converter is needed) but PCI is working just fine right now with external GPUs.
And keep in mind that the link above is talking about full pipeline rendering, where as Nvidia is just demonstrating supplemental rendering aid through the internet. USB 3.0, possibly even 2.0, and certainly PCI would be fast enough for that.