The reasoning is like a choice between buying a single box for 10 million, or buying a hundred box cluster for 2 million, then spend 10 million carefully tuning all your software to run your workload on that 2 million cluster.