The alpine grasslands of the Three-River Source Region (TRSR), located in the hinterland of the Qinghai-Tibet Plateau, serve as the ecological security barrier of China and Asia's vital water tower. However, long-term overgrazing, climate variability, and fragile ecosystem foundations have led to widespread degradation. Since 2005, two major phases of national ecological protection projects have been implemented to restore grassland function and control grazing pressure. Yet, the extent to which different grassland types and regions responded to these interventions over time remains unclear.In this study, we integrated multi-source datasets, including remote sensing-derived grassland productivity, field surveys, and livestock statistics, to evaluate the spatiotemporal dynamics of theoretical and actual livestock carrying capacity and grazing pressure in the TRSR from 1990 to 2020. We classified major alpine grassland types and constructed a climate response model using the XGBoost algorithm. This model, trained on pixel-level annual yield data and key meteorological variables, was used to simulate future carrying capacity under CMIP6 climate projections (2021-2030).Our findings reveal that both ecological project phases improved carrying capacity, but with clear differences among grassland types and regions. In Phase I (2005-2012), theoretical carrying capacity increased significantly for alpine meadow (up to 0.646 Sheep Unit /hm²), alpine steppe (0.319 Sheep Unit /hm²), and alpine desert (0.220 Sheep Unit /hm²). Phase II (2013-2020) emphasized zoning and ecological economy, resulting in a sharp increase in alpine desert capacity (+15.58%), moderate gains in meadow-steppe (+4.36%), but a slight decline in alpine meadow (-0.58%). This indicates that fragile grassland ecosystems such as alpine desert responded most sensitively to ecological intervention.Despite ecological recovery, actual livestock numbers remained above sustainable limits throughout both phases. The average grazing pressure index declined to 1.42 during Phase I, suggesting reduced ecological stress, but rebounded to 1.47 in Phase II, reflecting a mismatch between ecological restoration and grazing intensity. Regionally, eastern counties (Henan, Zeku, Banma) consistently exhibited higher grazing pressure than western counties (e.g., Zhidoi), aligned with economic development levels and grazing dependence.The XGBoost-based prediction model achieved high performance (R²=0.729 on the test set) and strong spatial transferability. Its integration with CMIP6 scenarios allows high-resolution forecasts of future carrying capacity, providing valuable tools for adaptive management.This study offers the first regionally detailed analysis of the dynamic response of alpine grassland carrying capacity to national ecological policies. It highlights the need for differentiated management strategies that account for grassland type, regional pressures, and climate trends. Our results emphasize the importance of coupling ecological restoration with socioeconomic coordination to ensure long-term sustainability and ecological security in the TRSR under climate change.