My read is that it's via something very similar to Reed-Solomon erasure coding. Think of the old Usenet parity file approach. "Here are 100 files. You need 80 o

My read is that it's via something very similar to Reed-Solomon erasure coding. Think of the old Usenet parity file approach. "Here are 100 files. You need 80 of them in order to reconstruct the file." Spread those throughout multiple facilities and you're basically there.