The real cost wins of self-hosted are that anything using new hardware becomes an ordeal, and engineers won't use high-cost, value-added services. I agree that there's often too little restraint in cloud architectures, but if a business truly believes in a project, it shouldn't be held up for six months waiting for server budget with engineers spending doing ops work to get three nines of DB reliability.

There is a size where self-hosting makes sense, but it's much larger than you think.