In this paper, we propose a server provisioning strategy for energy conservation in large scale VOD systems, which dynamically turns on/off servers in order to adaptively tailor active servers to dynamic user load. By defining quality-of-service (QoS) in terms of system overload probability, we reinterpret the energy conservation problem into minimizing the number of active servers subject to a QoS requirement. Relying on designing a recursive least square (RLS) based online user number predictor and constructing a large deviation based overload probability estimation model, we derive a practical dynamic server provisioning strategy which does not require any prior knowledge about the future workload. Finally, the experiments are carried out based on synthetic and real workload respectively to investigate the achievable performance of the proposed strategy.