Cost function estimation under production uncertainty is problematic because the relevant cost is conditional on unobservable expected output. If input demand functions are also stochastic, then a nonlinear errors-in-variables model is obtained and standard estimation procedures typically fail to attain consistency. But by exploiting the full implications of the expected profit maximization hypothesis that gives rise to ex-ante cost functions, it is shown that the errors-in-variables problem can be effectively removed, and consistent estimation of the parameters of interest achieved. A Monte Carlo experiment illustrates the advantages of the proposed procedure as well as the pitfalls of other existing estimators.