When you put together a bunch of equipment with a small error rate, the time between errors climbs very fast. Build a RAID array with 33 disks and you’d better have a vendor picked out for replacements because you’ll be doing replacements fairly often, instead of every four to ten years with a single disk.
And they don’t understand dependent statistics either. Every failing rocket engine can potentially damage its neighbors. Every failing hard drive requires a stressful operation on the remaining disks (resilvering) that may push the next drive to failure.