This gets especially critical once you stop assuming that data can be centralised. Once your required transactional context needs to scale large areas (geographically or logically) then the overhead of this is mighty. 2PC itself does not scale indefinitely, and while interesting research is being done in alternative transactional approaches, it would be tough to argue that any are acceptable solutions to the problem.
Also your example of banks is perhaps not an ideal one. While banks may indeed run some very large databases, they don't run one HUGE database which is expected to be always consistent. Things like ATM systems etc. run based on eventual consistency and tolerance favouring availability precisely because a simple horespower approach to a transactional database (whether SQL or not) wouldn't currently work in this domain.