We examined 1) the causes of inter-model variations of surface energy partitioning (SEP) and 2) the effects of model grid size on the simulated SEP. In particular, we focus on the nonlinear effect of spatial heterogeneity in atmospheric conditions on the simulation of surface fluxes in the mesoscale model by testing their scale-invariance from a tower footprint to regional scales. The test domain was a homogeneous shortgrass prairie in the central part of the Tibetan Plateau with an eddy-covariance flux tower at the center. We found that 1) soil evaporation controls the model differences of the SEP and 2) the spatial variability resulting from changing distribution of clouds and precipitation in the model domain affected radiative forcing at the ground surface, thereby altering the partitioning of surface fluxes. Consequently, due to increasing spatial variability in atmospheric conditions, the results of the mesoscale model did not produce convergent estimates of surface fluxes with increasing grid sizes. Our finding demonstrates that an atmospheric model can underestimate surface fluxes in regional scale not necessarily due to intrinsic model inaccuracy (e.g., inaccurate parameterization) but due to scale-dependent nonlinear effect of spatial variability in atmospheric conditions.