We address online scheduling for servers of Cloud service providers. Each server is composed of several variable speed processors whose power function is convex. The servers may be busy, idle or switched off. The objective of our scheduling is to minimize the energy consumed by a Cloud computing platform. To achieve this goal, we try to anticipate computing demands by predicting a workload, then we modify the set of available servers to fit this prediction and finally we schedule our jobs on the available servers. To schedule jobs we have developed the POD (Predict Optimize Dispatch) algorithm. We evaluate its performance for real-life traces in the presence of different types of prediction. The analysis shows that our scheduling reduces energy consumption considerably.