We consider a multi-user MIMO downlink where the transmitter has only estimates of the channel while the receivers have perfect channel information. The impact of channel estimation error on the sum rate is studied. It is shown that the sum rate saturates due to the inaccurate channel information. In order to maintain full multiplexing gain, the channel estimation quality has to increase with at least the square root of data SNR but there is no need to increase more than linearly with the SNR. With user scheduling, the sum rate scales at least M/2 log log K, where K is the number of users and M is the number of transmit antennas.