Accurate surface air temperature (T2m) data are key to investigating eco-hydrological responses to global warming. Because of sparse in-situ observations, T2m datasets from atmospheric reanalysis or multi-source observation-based land data assimilation system (LDAS) are widely used in research over alpine regions such as the Tibetan Plateau (TP). It has been found that the warming rate of T2m over the TP accelerates during the global warming slowdown period of 1998–2013, which raises the question of whether the reanalysis or LDAS datasets can capture the warming feature. By evaluating two global LDASs, five global atmospheric reanalysis datasets, and a high-resolution dynamical downscaling simulation driven by one of the global reanalysis, we demonstrate that the LDASs and reanalysis datasets underestimate the warming trend over the TP by 27％–86％ during 1998–2013. This is mainly caused by the underestimations of the increasing trends of surface downward radiation and nighttime total cloud amount over the south and north TP, respectively. Although GLDAS2.0, ERA5, and MERRA2 reduce biases of T2m simulation from their previous versions by 12％–94％, they do not show significant improvements in capturing the warming trend. The WRF dynamical downscaling dataset driven by ERA-Interim shows a great improvement, as it corrects the cooling trend in ERA-Interim to an observation-like warming trend over the south TP. Our results indicate that more efforts are needed to reasonably simulate the warming features over the TP during the global warming slowdown period, and the WRF dynamical downscaling dataset provides more accurate T2m estimations than its driven global reanalysis dataset ERA-Interim for producing LDAS products over the TP.