This study investigates a potential transition in vegetation drought-driving mechanisms within arid and semi-arid regions-from traditional water-supply dominance toward growing atmospheric-demand dominance. Focusing on China's Mu Us Sandy Land (2000-2022), we characterise this transition's spatiotemporal dynamics using multi-source remote sensing data integrated with Theil-Sen trend analysis, Mann-Kendall change detection, moving window regression, and spatial autocorrelation. The key findings are as follows: (1) Significant vegetation greening occurred (NDVI increased at 95.7% of sites), concurrently with stable water supply (SPI) and intensifying atmospheric aridity. This 'greening-while-drying' pattern challenges traditional water-supply-centric models. (2) Moving window regression identified a critical transition circa 2012, during which the dominance of the evaporative demand drought index (EDDI) strengthened markedly. By 2016, a new regional steady-state had emerged, signifying a gradual transition from "water supply limitation" to "water demand limitation". (3) Spatial differentiation was pronounced: atmospheric demand-driven zones (26.4%) formed patchy clusters in centraleastern and southern low-elevation areas, while water supply-driven zones (11.2%) were dispersed in eastern and southern regions. (4) Land use and topography jointly moderated this pattern. Forests exhibited the highest sensitivity to atmospheric drought (55.6% atmosphere-driven), in contrast to the more resilient grasslands. Lowelevation areas proved more vulnerable to atmospheric drought stress. The absence of "Coordinated Improvement" and "Water-Limited Degradation" types provides supporting evidence for the mechanism transition, corroborating that historical vegetation degradation is linked to atmospheric drought rather than precipitation decline. These findings necessitate a strategic reorientation in arid-land management: moving from pursuing vegetation coverage alone toward building climate-resilient ecosystems adapted to the "demand-dominant" paradigm.