Arctic warming has increased vegetation growth and soil respiration during recent decades. The rate of Arctic warming will likely amplify over the 21st century. Previous studies have revealed that the most severe Arctic warming occurred during the cold season (September to May). The cold-season warming has posited significant CO2 emissions to the atmosphere via respiration, possibly offsetting warm-season (June to August) net CO2 uptake. However, prevailing Earth system land models poorly represent cold-season CO2 emissions, making estimates of Arctic tundra annual CO2 budgets highly uncertain. Here, we demonstrate that an improved version of the energy exascale Earth system model (E3SM) land model (ELMv1-ECA) captures the large amount of cold-season CO2 emissions over Alaskan Arctic tundra as reported by two independent, observationally-constrained datasets. We found that the recent seven-decades warming trend of cold-season soil temperature is three times that of the warm-season. The climate sensitivity of warm-season net CO2 uptake, however, is threefold higher than for the cold-season net CO2 loss, mainly due to stronger plant resilience than microbial resilience to hydroclimatic extremes. Consequently, the modeled warm-season net CO2 uptake has a larger positive trend (0.74 ± 0.14 gC m-2 yr-1) than that of cold-season CO2 emissions (0.64 ± 0.11 gC m-2 yr-1) from 1950 to 2017, supported by enhanced plant nutrient uptake and increased light- and water-use efficiency. With continued warming and elevated CO2 concentrations under the representative concentration pathway (RCP) 8.5 scenario, the increasing rate of warm-season net CO2 uptake is more than twice the rate of cold-season emissions (1.33 ± 0.32 gC m-2 yr-1 vs 0.50 ± 0.12 gC m-2 yr-1), making the modeled Alaskan Arctic tundra ecosystem a net CO2 sink by 2100. However, other geomorphological and ecological disturbances (e.g. abrupt permafrost thaw, thermokarst development, landscape-scale hydrological changes, wildfire, and insects) that are not considered here might alter our conclusion.