One of the main selling points of the cloud is dynamic scaling which necessitates that Amazon have enough servers for some multiple of their customers’ base load. As long as internal Amazon resources are given a lower priority and booted in favor of customers during leak load, the opportunity cost is basically zero leaving just the marginal cost of electricity and hardware maintenance.
I am fairly certain this does not happen. An internal AWS customer account can have numerous flags associated with it but "boot me out first" isn't one of them (aside from spot instances that everyone has access to).
