I think you'll want to create a watchdog to run on these servers and monitor processor loads, in order to contrast these with the usage logs and assess the size of your required stack. Many cloud hosting services include APIs that allow scaling through a SOA API. Some (I know Azure does) include automatic scaling as an option though.

Beyond that this energy initiative is mostly dependent on your application logic itself. I don't know what kind of volume and complexity we're dealing with as far as the data is concerned, but you'll need to resort to ETL architecture where possible, i.e. where data doesn't need to be actual, in order to cut down processor cycles.