The land surface water balance components simulated by 20 atmospheric global circulation models (AGCMs) participating in phase II of the Atmospheric Model Intercomparison Project (AMIP II) are analyzed globally and over seven Global Energy and Water Cycle Experiment Coordinated Enhanced Observing Period basins. In contrast to the conclusions from analysis of AMIP I, the results presented here suggest that the group average of available AGCMs does not outperform all individual AGCMs in simulating the surface water balance components. Analysis shows that the available reanalysis products are not appropriate for evaluation of AGCMs’ simulated land surface water components. The worst simulation of the surface water budget is in the Murray–Darling, the most arid basin, where all the reanalyses and seven of the AGCMs produce a negative surface water budget, with evaporation alone exceeding precipitation and soil moisture decreasing over the whole AMIP II period in this basin. The spatiotemporal correlation coefficients between observed and AGCM-simulated runoff are smaller than those for precipitation. In almost all basins (except for the two most arid basins), the spatiotemporal variations of the AGCMs’ simulated evaporation are more coherent and agree better with observations, compared to those of simulated precipitation. This suggests that differences among the AGCMs’ surface water budget predictions are not solely due to model-generated precipitation differences. Specifically, it is shown that different land surface parameterization schemes partition precipitation between evaporation and runoff differently and that this, in addition to the predicted differences in atmospheric forcings, is responsible for different predictions of basin-scale water budgets. The authors conclude that the selection of a land surface scheme for an atmospheric model has significant impacts on the predicted continental and basin-scale surface hydrology.