Two-step instrumental variable estimators for dynamic panel data models are considered that are asymptotically efficient under some auxiliary assumptions, but remain consistent when the assumptions are violated. Asymptotic efficiency is defined in relation to the information bound for the conditional mean specification of the model. Unlike in standard panel GMM, optimal instruments are parameterized using a fixed number of coefficients for any value of T. Thus, the properties of the resulting estimators are not fundamentally affected by the relative dimensions of T and N.