A large number of sensors, meters and intelligent electronic devices (IEDs) with the capabilities of sensing, actuation, computation and communication will be deployed in future smart power grids for the purpose of measurement, monitoring, protection, diagnosis, control, optimization and other transactions. The conventional centralized and hierarchical multi-level state estimation is not scalable enough to process the huge amount of data generated all over the grid. In this paper we propose a fully distributed computational network architecture and the associated Message-Passing (MP) algorithms for electric power system state estimation. Unlike the conventional state estimation schemes that centered around algebraic operations on sparse matrices, our approach is based on MP and information fusing on graphs in a totally distributed fashion. The optimality, scalability and other advantages will be demonstrated.