Over the past decades, land ecosystems removed from the atmosphere approximately one-third of anthropogenic carbon emissions, highlighting the importance of the evolution of the land carbon sink for projected climate change. Nevertheless, the latest cumulative land carbon sink projections from 11 Earth system models participating in the Coupled Model Intercomparison Project phase 6 (CMIP6) show an intermodel spread of 150 Pg C (i.e., ∼ 15 years of current anthropogenic emissions) for a policy-relevant scenario, with mean global warming by the end of the century below 2 ∘C relative to preindustrial conditions. We hypothesize that this intermodel uncertainty originates from model differences in the sensitivities of net biome production (NBP) to atmospheric CO2 concentration (i), to air temperature (ii), and to soil moisture (iii), as well as model differences in average conditions of air temperature (iv) and soil moisture (v). Using multiple linear regression and a resampling technique, we quantify the individual contributions of these five drivers for explaining the cumulative NBP anomaly of each model relative to the multi-model mean. We find that the intermodel variability of the contributions of each driver relative to the total NBP intermodel variability is 52.4 % for the sensitivity to temperature, 44.2 % for the sensitivity to soil moisture, 44 % for the sensitivity to CO2, 26.2 % for the average temperature, and 21.9 % for the average soil moisture. Furthermore, the sensitivities of NBP to temperature and soil moisture, particularly at tropical regions, contribute to explain 34 % to 65 % of the cumulative NBP deviations from the ensemble mean of the two models with the lowest carbon sink (ACCESS-ESM1-5 and UKESM1-0-LL) and of the two models with the highest sink (CESM2 and NorESM2-LM), highlighting the primary role of the response of NBP to interannual climate variability. Overall, this study provides insights on why each Earth system model projects either a low or high land carbon sink globally and across regions relative to the ensemble mean, which can focalize efforts to identify the representation of processes that lead to intermodel uncertainty.