Over the recent years many advances in the area of nonlinear model predictive control have been made. However, with respect to the output feedback problem and predictive control only a limited number of results are available. Most of the existing approaches do only guarantee local stability. In this note we propose the use of a high gain observer in combination with a sampled nonlinear predictive controller for a special MIMO system class. The resulting output feedback leads to semiglobal practical stability and recovery of performance of the state feedback controller. The results are valid for a wide class of stabilizing NMPC schemes. Thus they can be considered as a special separation principle for NMPC.