That's weird....I believe that when the dataset is bigger than ram, you just break it into different servers (or just get more ram). This isn't anything new, this is actually old OLD technology, and you can find cases of this everywhere (google, microsoft uses it for hotmail, etc), called the Brick architecture or something like that.
Actually, a ton of research was done on these types of architectures by Jim Gray (who tragically was lost at sea a few years ago) at Microsoft. Not sure how many of these links are still active but:
http://research.microsoft.com/en-us/um/people/gray/
From my own experience, at a job a few years ago, we were dealing with a little over 1 petabyte of data, and the system was engineered to NEVER hit platter, and to always have data in ram.