The power consumed by a large IT system may be managed by individual servers calculating a set of predictors of the resources of the large IT system that may be required for a future time period, and weighting factors for those predictors. The predictors and weighting factors may be based on historical data showing actual resources used and past values of the predictors and weighting factors. An estimate of the resources required for the large IT system is then produced based on the predictors and weighting factors and the estimate is then used to determine excess resources of the individual server that may be put into a low power state in the future time period to minimize the power used. If too much resource is excess, the server may enter a soft off state.