Quantifying soil organic carbon (SOC) dynamics in rangelands is challenging due to sparse observations, high spatial heterogeneity, and nonlinear responses to climate and management. Using a hybrid framework that integrates process-based modelling (FullCAM), machine learning, and spatio-temporal Bayesian upscaling, we assessed SOC trends and drivers across Australian rangelands from 1990 to 2021 at 0.05 degrees (similar to 5.5 km) resolution. Model calibration with SOC fractions data from 166 sites revealed that while grass biomass influences short-term SOC variability, climate, particularly rainfall, temperature, and evaporative demand,dominates system-level dynamics. Spatio-temporal modelling indicates that rangeland SOC stocks (similar to 10-12.5 Gt C) have remained near a quasi-steady state over three decades, with only a minor net decline (<1%) despite pronounced seasonal fluctuations. Explicit application of the Area of Applicability (AoA) and uncertainty analysis ensures predictions are constrained to ecologically valid regions, enhancing transparency for national greenhouse gas accounting. These findings highlight the limited potential for large-scale SOC sequestration in arid and semi-arid rangelands and underscore the need for climate-adaptive strategies rather than reliance on management interventions alone. Future work should focus on higher temporal resolution, integration of remote sensing vegetation indices, and expanded observational datasets to refine SOC monitoring and inform policy under changing climate regimes.