Atmospheric water demand is practically characterized as vapor pressure deficit (VPD) and has been identified as a critical driver of ecosystem function, by affecting plant mortality, wildfires, and carbon loss. In this study, we used daily eddy covariance data across Chinese forest, grassland and shrubland ecosystems, in combination with remote sensing data, to investigate the impacts of VPD on gross primary productivity (GPP). We found divergent VPD impacts on GPP among grassland, shrubland and forest ecosystems. The VPD yielded substantial inhibitory impacts on GPP in grassland ecosystems and this suppressing impact was regulated by soil water content (SWC), showing that GPP declined with VPD under dry conditions but increased with VPD under wet conditions. This GPP variance was attributed to VPD, SWC and their interactions. More than 50% of the variability in GPP was explained by SWC in grassland ecosystems and by VPD in forest and shrubland ecosystems. Partial correlation analysis, random forest, and multiple linear regression revealed similar results when temperature, radiation and SWC were considered. Compared with shrubland and forest ecosystems, grassland has drier environmental conditions and poorer soil water-holding capacity that led to lower SWC and stronger negative impacts on VPD. Thus, GPP was susceptible to the negative impacts of higher VPD, especially under dryness stress. Our results highlight the need to comprehensively consider divergent VPD impacts for different ecosystems to more accurately assess climate impacts on ecosystem function.