Understanding the spatial distribution of soil organic carbon (SOC) on the Tibetan Plateau (TP) and its response to future climate change is crucial for regional carbon cycling and ecosystem management. This study used 372 soil profiles, including data from a dedicated field survey, to model SOC density (SOCD) at depths of 0-30, 30-50, and 50-100 cm using a recursive feature elimination-random forest (RFE-RF) approach. Based on 10-fold cross-validation, the model explained 56%, 48%, and 35% of the spatial variability in SOCD across these layers. For the baseline period (1990-2025), the total SOC stock in the upper 1 m was estimated at 34.17 Pg (90% prediction interval: 6.29-90.25 Pg). Of this total, 16.29 Pg (4.16-36.48 Pg), 6.85 Pg (0.84-18.75 Pg), and 11.03 Pg (1.29-35.02 Pg) were stored in the 0-30, 30-50, and 50-100 cm layers, respectively. Shapley additive explanations (SHAP) analysis identified the aridity index (AI) as the most influential driver, with higher aridity consistently exerting negative effects on SOCD across all depths. In the topsoil (0-30 cm), additional key drivers included the normalized difference vegetation index (NDVI), soil pH, and mean annual precipitation (MAP). The influence of potential evapotranspiration (PET) increased in the 30-50 cm layer, whereas silt content and shortwave infrared 1 (SWIR1) reflectance became more influential in the deepest layer (50-100 cm). Projections under the SSP5-8.5 scenario suggest a significant decline (p < 0.01) in SOC stocks from 2030 to 2100, particularly in the 50-100 cm layer. While humid to semi-arid regions experienced substantial SOC losses, the ariddesert region showed a significant increase in surface SOC (0-30 cm). These results clarify layer-specific drivers and future vulnerabilities of SOC on the TP and provide useful information for adaptive soil carbon management