Persistent storage to which access may be significantly delayed on a human time frame probably tends to loose utility rapidly.

And if one has a Terabyte of data, how is putting it in the cloud and accessing it at web speeds better than a hard disk at bus speeds?

With a processor cache, populating the cache predicatively is far easier due to the limited ___domain of alternatives, the logical structure of instructions, and trillions of cycles per core per hour available for testing alternative predictive schemes.

On the other hand, a user may ask for a rarely requested file once every ten to thirty fortnights, or never. And predicting that request would require parsing a joke told on WJMZ's morning show fourteen minutes ago.

I don't see him mentioning CPU caches anywhere.

The point is that a cache, even if it is a fat cache, will remove files eventually. Files the user may be expecting to find.