Given the bare-bones API and semantics, probably because it describes absolutely no fault tolerance at all.
Lose a processor or a message and it's just lost. Try to address that with copy-put-take and you now have duplicates. Try to address that and you now have byzantine failures with no resolution or supporting tooling. So it's wonderfully simple until you need to do something outside of a single process and then it's useless.
Internet-friendly stuff generally has a hard requirement on being fault tolerant.
near the beginning of the Scientific Computing Associates work (?) they implemented an extended model that supported multiple spaces - to scope the number of machines involved in a given update.
