Well, you only need to get shared infrastructure reliable enough that you can afford to not design it with repair in mind. The cloud servers are already design without unit-level maintenance work in mind, which saves money by eliminating rails and similar. They get populated racks from the factory and just cart them from the dock to their place, plug them in (maybe run some self-test to make sure there are no loose connectors or so), and eventually decommission them after some years.