I think they are doing it wrong then. You do the cheap parts independently. Then you pair down to fewer choices and transfer resources to fewer groups.
You repeat this multiple times depending on your resources. When things get really expensive, you should be down to one idea.
I'd imagine that this can be pretty tricky with more complex systems, where the buy-in for even getting off the ground can be pretty high. In software, that's usually not the case, but I can imagine a number other instances where it wouldn't work. This seems like it would show up the most when the problem isn't a greenfield where getting the first 90% working is the hard part, but on the other end, where the difference comes from optimizing the last 10%, which can be really expensive (i.e. figuring out how to profitably and effectively take production from 1000 parts to 10,000 parts can be much more difficult and costly than moving from 100 to 1000).
You repeat this multiple times depending on your resources. When things get really expensive, you should be down to one idea.