Likely it has more to do with the types of tasks they're doing. If you think about facebook and how it's used, it's mostly a lot of little bits of data. An awful lot of little bits of data. Sure there are lots of images - but ignoring the image data its self (that's only really useful to end users) it's still just a bit of data, owner, dates and users tagged.
What they have to do is crawl a lot of social graphs to update users with a small bit of data, the things their friends are doing. Locality of reference is a good thing, but probably very hard with the type of operations they're performing.
And HTTP needs a mode for reassembling out-of-order snippets. Which use snippet independent data compression. Oh, and everybody gets a pony.