Well, you clearly wouldn't lease dedicated hardware with this model. You just redistribute load amongst data pods when there is a failure, and pull the whole pod from service for refurbishment once it passes your total failure threshold.