But the fact is I really, really don't. We are so spoiled now. Spinning up an AWS gpu cluster in 60 seconds from the beach beats spending a day and a half just to upgrade a video card ;)
In a dynamic optimization problem, where we have to take into account the current state of the world, and our ignorance of the future, sometimes the most efficient use decision is going with what we have, since it's good enough.
I'll put this another way: I don't think there's some terrible shortage of fundamental computer science research out there, nor is there a shortage of competing architectures.
One more spin on this. The fact that we have fewer competing architectures than we used to is itself a sign that existing architectures are pretty good. First, they are the survivors, and have out-competed the rest. Second, more importantly, if less research is being done to develop competing platforms, that means researchers are less optimistic about their ability to improve on what we currently have.