Dryland ecosystems play a vital role in global carbon cycling, yet their carbon dynamics remain poorly understood, particularly under climate change. Most studies in dryland ecosystems have concentrated on soil organic carbon (SOC) stock, with limited attention to aboveground biomass carbon (AGBC) and belowground biomass carbon (BGBC). Here, we conducted a comprehensive assessment of carbon stocks in the dryland ecosystems of Northwestern China. By integrating 2304 samples points from 2000 to 2020, we developed predictive models with machine learning (ML) techniques for AGBC, BGBC, and SOC, respectively. Our results demonstrate that ML models substantially outperformed multiple linear regression (MLR). Incorporating climate variables (CL), CO2 concentration (CC), and land use (LU) as environmental covariates into random forest (RF) model, further enhancing the predictive accuracy. By shapley additive explanations (SHAP), LU emerged as the most influential predictor for AGBC, contributing 34% of the explained variance, while mean annual maximum temperature (TMX) and elevation (ELE) were identified as the dominant drivers for BGBC and SOC. Nonlinear relationships were evident between these key drivers and the three carbon pools. The structural equation model (SEM) further revealed that TMX exerted both direct and negative indirect effects on BGBC through its influence on soil properties, whereas ELE consistently exerted dominant positive effects on SOC through its influence on vegetation. From 2000 to 2020, the dryland ecosystems of Northwest China functioned as a net carbon sink, with ecosystem carbon (AGBC+BGBC+SOC) increasing at a rate of 32.87 +/- 8.33 Tg C yr(-1). This increase was primarily driven by contributions from mountainous regions, whereas SOC in desert ecosystems accounted for the majority of carbon losses. These results highlight the spatiotemporal heterogeneity and the difference of key drivers of AGBC, BGBC, and SOC. Our work offers valuable insights and methodological frameworks for evaluating carbon dynamics in dryland ecosystems.