That's a claim of the durability of the data, or the odds that a chunk of data will be lost in a year.
They calculate this through the odds that a single node fails, and then multiplying that odd out through all replicas. This covers the most easily quantifiable failure mode.
Obviously the real odds are somewhat higher when you consider that a rogue admin, malicious actor, or buggy code could delete multiple instances of replicated data at once. There's no way to estimate these odds though, and really they don't matter - they're big enough events that they could spell the end of Dropbox if they happened.
Another major class of data loss in this kind of system is operator error when dealing with rare, unrehearsed events requiring operator intervention. Often combined with confusing messages or behaviour from the software.
They calculate this through the odds that a single node fails, and then multiplying that odd out through all replicas. This covers the most easily quantifiable failure mode.
Obviously the real odds are somewhat higher when you consider that a rogue admin, malicious actor, or buggy code could delete multiple instances of replicated data at once. There's no way to estimate these odds though, and really they don't matter - they're big enough events that they could spell the end of Dropbox if they happened.