Intensifying global aridification is reshaping the water-carbon balance in dryland ecosystems, yet when and how this coupling reaches critical thresholds remains poorly understood-posing major uncertainties for predicting ecosystem resilience under climate change. Using Central Asia as a representative dryland region, we partitioned ecosystem water-use efficiency (WUE = GPP/ET) into transpiration efficiency (WUEc = GPP/Ec) and the transpiration-to-evapotranspiration ratio (TET = Ec/ET) to probe threshold responses. We identified distinct aridity thresholds associated with nonlinear shifts in water-carbon coupling: WUE and WUEc declined and then rebounded beyond aridity levels of 0.75 and 0.81, respectively, indicating tighter coupling between carbon gain and water loss under severe climatic stress. In contrast, ET, Ec, GPP, and TET declined sharply at thresholds of 0.80, 0.85, 0.86, and 0.88, indicating broader constraints on ecosystem carbon assimilation and hydrological functioning. SHAP-based machine learning revealed threshold-dependent shifts in ecosystem controls, with vegetation traits dominating carbon-water coupling under lower aridity, and hydroclimatic stressors-including soil moisture, precipitation, temperature, and vapor pressure deficit-emerging as dominant regulators beyond critical thresholds. Notably, temperature was positively associated with WUE and WUEc, while VPD exerted persistent negative effects post-threshold. These findings characterize nonlinear, threshold-dependent changes in dryland carbon-water coupling and highlight the need to evaluate apparent efficiency gains in the context of declining ecosystem fluxes under future drought intensification.