Just a few PB, and sharding it across servers is fine.

It looks like all the disk-optimized examples on that site (still much more expensive than paying for raw disk, barely 5x cheaper than S3, when a disk-optimized colo solution only has ~3% overhead over the disks themselves) are through some no-name provider "HostKey". I suppose beggars can't be choosers, but in the contexst of storage (where systemic failures should be accounted for in the model) are you aware of more than one provider with reasonably priced storage?