The Tibet Plateau is highly sensitive to climate change, yet the spatial patterns, stocks, and climatic thresholds of soil inorganic carbon (SIC) versus soil organic carbon (SOC) remain poorly quantified. Here, this study compiled SIC and SOC measurements from field surveys (2016-2024), ISRIC datasets (1980-2012), and published literature, and integrated them with multi-sensor remote-sensing covariates to map SIC and SOC across the Tibet Plateau using six machine-learning models. The modeling was based on 865 and 329 SOC samples at 0-20 cm and 0-100 cm, and 140 and 136 SIC samples at 0-20 cm and 0-100 cm, respectively. To ensure robust evaluation, we used an 80%/20% train-test split and applied 10-fold cross-validation within the training set for hyperparameter tuning and best-model selection, performed separately for 0-20 and 0-100 cm prior to mapping. The results demonstrated that: (1) Random Forest (RF) algorithm achieved the highest predictive accuracy for SOC and SIC at 0-20 cm depth, whereas Boosted Regression Trees performed best for SIC at 0-100 cm (R2 = 0.79). (2) Estimated total soil carbon stocks were 20.61 Pg(0-20 cm) and 82.32 Pg (0-100 cm). Spatially, SIC was enriched in the arid northwest and deeper layers, while SOC dominated the humid southeastern plateau and surface soils, consistent with their contrasting formation mechanisms. (3)Using generalized additive models with SHAP, we identified nonlinear climatic controls with distinct thresholds: SIC peaked at 331 mm MAP, whereas SOC peaked at 691 mm, and SOC showed stronger temperature sensitivity. SOC estimates from this study were broadly consistent with SoilGrids and GSOCmap (R2up to 0.78), while our plateau-focused framework with depth-specific model selection (0-20 and 0-100 cm) provides depth-consistent SIC and SOC maps for the Tibet Plateau. These results highlight divergent climate sensitivities of SIC and SOC and improve regional carbon-budget assessments for the Tibet Plateau.