Model Predictive Control (MPC) is a well-known control architecture that has encountered an enormous variety of applications since the very beginning until current days. The pros and cons of such control technique are very well known and both of them rely on the embedded model, which is used to determine the control trajectory. Still, the discrepancy between embedded model and real operative conditions can affect the control response due to uncertainties in measurement chain, noise and so on. Still, it is hardly available in literature what would happen in case the plant is operating far from the design condition of the model. This is of particular interest once a linear MPC is governing a non-linear process where the linearization of the target plant must be processed to tune the MPC itself. This paper analyses experimental results from a fuel cell gas turbine hybrid system, namely SOFC/GT emulator test rig, where a linearized MPC was adopted to control stack inlet temperature. The test rig is constituted by a modified Turbec T100 micro gas turbine where a volume of 4 m3 is interposed between compressor and turbine. This emulates the impact of the SOFC on the GT. The system is connected in real-time mode to a model, which runs in parallel and reads what is going on the plant side and simulates the behavior of the associated SOFC stack. The MPC controller governs the plant according to the stack inlet temperature computed by the model in real-time mode. This MPC must be considered as a supervisor of the system, as the gas turbine was still equipped with its original control system. The plant was subject to an on-purpose strong degradation -operated via a constant venting of air from compressor to ambient. This operation strongly influenced the performance of the system, which were no longer able to operate at a level of temperature and power for which the controller was designed for. Still, a ramping down in power and back up was performed and the MPC showed performance which were in agreement with the design performance. Such surprisingly good result is explained with the complexity of the embedded model, which was derived from a completely physical model of the target system and constituted by more than 200 states.