Understanding when additional water consumption ceases to yield proportional carbon gain is essential for irrigation-dominated croplands exposed to strong atmospheric demand. Using the Hetao Irrigation District in arid northern China as a representative system, we developed a frontier-based framework to diagnose how evapotranspiration (ET), vapor pressure deficit (VPD), and deep soil water background jointly constrain gross primary productivity (GPP). Native 8-day PML-V2a ET and GPP estimates for 2019–2024 were combined with ERA5-Land meteorological variables on a common 1 km cropland grid. A generalized additive model was used to reconstruct the marginal effect of VPD on GPP in ET–VPD space, and the zero-isoline of that surface was extracted as the functional VPD threshold frontier. The cropland GPP–ET relationship showed clear nonlinearity and a stable main breakpoint of 12.98 mm/8d (95% CI: 12.49–13.86 mm/8d), indicating a transition from a relatively stronger water-limited stage to a stage of diminishing marginal returns. The tolerable VPD upper bound increased with ET at first, but began to contract once ET approached the breakpoint, revealing an ET-dependent dynamic frontier rather than a fixed VPD threshold. Subgroup fitting further showed that the deep soil water availability index (DSWI) mainly shifted the frontier position: the breakpoint occurred at 7.18 mm/8d under low DSWI and at 9.35 mm/8d under high DSWI, corresponding to a rightward shift of 2.17 mm/8d. Seasonal diagnostics indicated that May–July constituted a high-growth yet high-risk window, while spatial risk clustered in downstream, poorly drained, and salinity-prone parts of the district. Quantile frontiers and a cross-product robustness test using independent GOSIF GPP preserved the same nonlinear saturation pattern. These results suggest that VPD primarily controls frontier shape whereas deep soil water background regulates system vulnerability mainly by shifting the threshold position.