Agree those SATA port multiplies are worrisome. In the beginning, our prototype machines used them to squeeze as many drives into a single machine as possible. They have unusually low tolerance for electrical interference and make it possible for one badly malfunctioning drive to take an entire array offline until manually serviced. We've seen occasions where just touching a cable attached to a port multiplier caused the Linux kernel to emit "dazed and confused" NMI events. I am not brave enough to try them again, even in a redundant setup.
How did you compute this "replication equivalent"?
Not if they lose 35% on top of that they are around 41% overhead. But, they are taking a huge it on write speeds, network traffic and reliability for doing so.
Edit: Looks like they have 10,058 TB before partitioning the drives so my guess is ~3-6TB of actual user data.
Backblaze's pods are a data-loss nightmare -- lots of single points of failure which will wipe out many TB of data at a time -- and backblaze has stated that they replicate data across multiple pods. Given that the 10 PB seems to be the amount of raw storage backblaze has, I'm guessing that the amount of actual data stored is much less -- depending on what sort of erasure correction scheme they're using, of course. (They're still much bigger than Tarsnap, of course!)