Techniques are provided for predictively activating physical servers. Embodiments determine a plurality of physical servers that are available within one or more data centers. A plurality of virtual workload deployment requests that were processed within the one or more data centers during a historical window of time is determined. Embodiments classify each of the plurality of virtual workload deployment requests into one of a plurality of categories. A respective neural network prediction model is generated for each of the plurality of categories, based on the virtual workload deployment requests classified into the respective category. Embodiments then determine a number of physical servers to have active at a future moment in time, using the generated neural network prediction models.