The computer is the bottleneck here.

Why force all the data to be loaded into one computer before selecting just the data that's actually needed?

Given the falling prices for CPU cores, when will we have a CPU that runs directly on the disk cache?

Then you can finally get to reasonable size disk arrays because the array controller will simply send SQL queries to the disks and only the matching records will use up bandwidth on the way back.

It's almost tempting to do that today with ATOM motherboards that have a few drives plugged into them and connect back through ethernet switches and routers. The result is a database server that's as scalable as the Internet itself. There just needs to be a little bit more software support so that blank units can be plugged into the array and boot into an automated setup process that formats their disks and spreads data onto them. Then put one LED on the front that's starts yellow, turns green and flashes red when the unit needs to be replaced.