Active magnetic bearing (AMB) systems are nonlinear, multi-variable and inherently unstable. Conventionally, LTI models are obtained either from first principles or identification experiments and then used to design/synthesize feedback controllers. In general, identification algorithms are focused on obtaining the current and displacement coefficients for a linearized AMB model, which in-turn is based on first principles modeling. In recent years, it is shown that predictor-based subspace identification algorithms (PBSID) give consistent estimates of the parameters even in the presence of feedback. In view of this, we apply the PBSIDopt method to obtain a linear model of an experimental AMB system. Further, a robust controller against uncertainties in a reduced order model (neglecting flexible dynamics) is synthesized.