An optimal GMM estimator is shown to be asymptotically equivalent, in an i.i.d. setting, to a maximum likelihood estimator for a certain parametric problem. This helps to explain the nonparametric efficiency of GMM. A new efficiency result is also proven.