It's neither. They make the assumption that their combined cloud partners aren't all in cahoots, and distribute their data such that the probability that a) one can intercept enough chunks and b) one can infer meaning from the chunks is very low.

The cynical part of me wants to say that it's very similar to a standard mapreduce, and the security properties came for free/very little.