The more machines used to parallel process a task, the more this kind of cost rounding hurts the consumer. Say I can process a queue in 3 hours with 1 machine, throwing 3 instances at the problem maintains the same cost, but cuts the time to 1 hour. But throwing 6 machines at the problem doubles costs for the same total work and data center usage. Is the per-hour-billing a mechanism to throttle peak usage?

The sleep/wake could allow for deploying or archiving a role such that it would no longer incur running and deployment costs. Instead, it would simply incur storage costs. This would remove the single point of failure of having to monitor queues from an outside system and use the management api to deploy and remove roles.