In this paper, we deal with the estimation of the ergodic capacity of large MIMO systems, using training sequences whose lengths are of the same order of magnitude than the number of antennas. In this context, the traditional estimator becomes inconsistent. Following the ideas developed by Girko in the context of the so-called theory of G-estimation, we propose a new estimator. We analyze its asymptotic behaviour and show, using numerical simulations, that it tends to improve significantly the performance of the standard estimate for a realistic number of antennas.