Abstract Terrestrial evapotranspiration (ET) is a major component of the surface hydrological cycle and controls land-atmosphere feedbacks by modulating the surface energy budget. Accurate ET quantification at global or regional scales is crucial for understanding variations in carbon and water cycling in a changing environment. Although various grid-based ET data models have been developed using multiple approaches, these vary in concept and physical scheme, leading to differences in performance. We examine uncertainties associated with the limitations of the physics used to assist in model selection and improvement. We evaluate multiple ET data models, including estimates derived from a variety of land surface models (LSMs) based on the operational North American Land Data Assimilation System (NLDAS) phase 2 (NLDAS-2) and the experimental NASA LIS-based NLDAS Testbed (NLDAS-Testbed) drivers, and satellite retrievals, compared to water budget-derived ET and tower observations. Overall, all models are able to capture the spatial variability of mean annual water balance-based ET (ETwb) and monthly seasonal cycles of tower ET measurements, although there is a large range of estimates. NOAH28, FLUXNET, SSEBop, LandFlux, and GLEAM perform best, as demonstrated by their higher correlation and smaller bias and RMSE values. Simple relative uncertainty analysis shows that the NLDAS-Testbed ensemble mean has a slightly lower uncertainty than that of the NLDAS-2 ensemble. Our study indicates that NLDAS-Testbed/VIC412 (NLDAS version/LSM version) is improving and NLDAS-Testbed /CLSM is deteriorating relative to NLDAS-2/VIC403 and NLDAS-2/Mosaic. NLDAS-Testbed /NOAH36 and NLDAS-Testbed /NOAHMP36 are comparable to NLDAS-2/NOAH28, although biases between models and ETwb exhibit opposite trends. These findings will help further improvement of these models and support future NLDAS development.