The successful design and evaluation of autonomous energy optimization techniques requires the availability of a ubiquitous and accurate set of measurement techniques that are cheap and easy to implement. We discuss an approach for mathematically estimating the wall power as well as the power of the principal functional units (like DRAM) in the server platforms without incurring the cost of hardware instrumentation. Support Vector Regression (SVR) has proven to be an effective tool in real value function estimation. In this paper we modify two loss functions, Vapnik’s e-insensitive loss function and an insensitive Huber loss function to be asymmetrical in order to limit underestimates. Our novel approach, asymmetrical support vector regression, provides accurate prediction while maintaining a low number of out of bounds misestimates. We test our approach on two different datasets by predicting the power for the next time interval and achieve accuracy rates of below 6 percent relative percentage error while keeping the number of boundary misestimates below 4 percent.