Theory #1 – Because deleting stuff at scale is harder than you think. If you've got structured data (think "relational database, much of which is pretty much guaranteed to be append-only") and you've optimised it's on-disk structure for reporting (think "highly indexed, possibly even with both the index data and the transaction data being stored in ways that take advantage of known query patterns and physical disk geometry) – removing data from it is likely to be _much_ more effort than just flipping a "deleted" flag and continuing to expand your storage pool. This is even more true if you'd also need to consider replicated copies/archives/snapshots/backups. Facebook/Google/Twitter et al store everything, not just because they think it'll make them more valuable to advertisers, but because deleting data from distributed/sharded/backedup/archived databases is more expensive than just marking it "deleted" and leaving it in place.

Theory #2 – The NSA (or it's predecessor or a related agency) have been paying them to store whatever they can since whenever it became possible. If you're prepared to ignore the privacy implications, it's obvious that some _tiny_ percentage of that data will become useful for law enforcement purposes. Unfortunately – when the "privacy implications" are considered in terms of "Is _your_ privacy more important than _my_ career?" – it's pretty clear that a very powerful cohort of law enforcement and intelligence agency decision makers say "Hell no! I just need one or two more big successes and I'll get that promotion I want! Of course it's worth monitoring every single person on the planet to give me a shot at the executive bathroom and an office with a door!" (or, less cynically but equal in consequence "Should I listen to every phone conversation in America, if it might possibly mean we can stop the next 9-11? Yeah, I think it might be…").