Clarifying temporal and spatial variations of carbon exchanges of arid and semi-arid grasslands, the most widespread terrestrial ecosystem in the Northern China, is critical to evaluate regional and national carbon balance. However, estimating the carbon sequestration capability of arid and semi-arid grasslands still remains a challenge due to large interannual variability and high sensitivity to human activities. In this study, based on 696 site-months of long-term eddy covariance observations from 7 sites, we estimated the regional carbon fluxes of the arid and semi-arid grasslands in Inner Mongolia, China with the random forest (RF) approach. The estimated carbon fluxes encompassed net ecosystem productivity (NEP), gross primary productivity (GPP) and ecosystem respiration (ER), at a spatial resolution of 0.1°×0.1°. We found that Inner Mongolia grasslands acted as a weak carbon sink with a mean NEP of 2.43 ± 6.71 Tg C y⁻¹ from 1982 to 2018. Fluctuations of precipitation during the peak growing season and leaf area index were the most important drivers for the interannual variations of regional carbon fluxes. During the dry period of 1999 to 2011, the grasslands switched from carbon sink to source resulting from a great drought-induced reduction in GPP than ER. Comparison of our RF carbon fluxes with the global datasets (FLUXCOM, CMIP6 models, and atmospheric inversion data of JenaCO₂) showed that these products had limited agreement in terms of spatial and temporal variations, particularly for interannual variabilities of carbon fluxes, which highlights the importance of local long-term observations for regional evaluation and simulation of carbon budget.