Under multivariant normal linear models, it is proved that usual loss estimators based on the uniformly minimum variance unbiased estimator and the best affinely equivariant estimator of error variance are inadmissible with squared error loss.