I wonder how much cheaper it is to implement software for one, huge machine that „just works”, vs the usual way of implementing for a network of distributed machines. Distributed programming is NOT easy, and, even if the hard parts are supposedly implemented by various protocols/frameworks, these solutions constrain the way you can program and also have a ton of „interesting” failure states - just take a look at logs of say Hadoop cluster that’s been running for a while - even though the cluster is supposedly running fine (or is it?), you’ll find various kinds of exceptions related to the distributed model (socket timeouts and whatnot). Not to mention trying to thoroughly understand various eventual consistency models.
On th plus side, we’ve basically moved complexity from hardware to software (it was pioneered by Google, a software company, so no wonder) which increases salaries of software people, at the expense of hardware people. So yay for us I guess.