The dynamic changes of soil organic carbon (SOC) are highly sensitive to climate change and changes in land use types, but there is still great uncertainty in the response of the carbon cycle under different land use types to future climate change. In this study, we integrated extensive soil observation data, digital soil mapping (DSM) technology, and machine learning (ML) models with a robust training strategy to develop an optimal spatiotemporal prediction model for assessing carbon stock trends in northeastern China and their response to future climate change. Between 1985 and 2020, the average SOC density in the study area declined by 5.51 MgCha−1, with significant differences in positive and negative SOC density changes, indicating an overall trend toward carbon loss. In the future (2020–2100), carbon source areas will continue to appear in the southeast, especially under the high emission scenario SSP585, the SOC density will decrease by 14.19 MgCha−1, and the carbon source area will spread from the southeast to the northeast over time. Under future climate scenarios, the SOC stocks will continue to be lost (1.8 Pg) in the high emission scenario (SSP585), while the low emission (SSP119) and medium emission (SSP245) scenarios will show dynamic changes with multiple carbon source and carbon sink conversions. Notably, regions with rich SOC stocks, such as forests and grasslands, are more vulnerable to climate change and face a higher risk of carbon depletion. Cultivated land and forests play a dominant role in future carbon stock changes, with cultivated land contributing significantly to carbon stock loss (sink-to-source transition) and forests playing a key role in carbon stock recovery (source-to-sink transition). This study provides important scientific basis for addressing climate change challenges, optimizing land management strategies, and maintaining regional carbon cycle balance in Northeast China.