As a critical agroforestry crop in Southern China, Moso bamboo, maintains regional timber security and bamboo shoot production, with its net ecosystem productivity (NEP) directly determining dry matter accumulation and economic yield. This study integrates 2024 continuous flux observations with XGBoost and SHAP explanations to characterize the subtropical bamboo forest carbon budget and its nonlinear driving mechanisms. The results show a weak carbon sink in 2024 with an annual cumulative NEP of 120 g C m⁻², as high respiration of 860 g C m⁻² limited organic matter conversion by consuming nearly 88% of the 980 g C m⁻² total primary production. The peak production period during May and June was offset by growth stagnation in August, caused by extreme heat and drought. Net radiation served as the primary driver, with a positive contribution threshold of 75.28 W m⁻², whereas precipitation exceeding 1.85 mm or air temperatures over 17.85 °C hindered carbon accumulation through radiation attenuation and metabolic heat loss. Strong radiation–precipitation interactions confirm that water’s impacts on yield are deeply contingent upon radiation backgrounds. These nonlinear regulatory pathways provide a scientific foundation for stabilizing bamboo forest productivity through synergistic water-radiation management and structural optimization during extreme climate events.

The balance between gross primary productivity (GPP) and ecosystem respiration (ER) defines an ecosystem’s carbon sink-source status. Under global warming, hydrothermal conditions critically shape carbon fluxes, yet their differential impacts on GPP and ER remain insufficiently understood, especially across biomes. Elucidating these differences is essential for reducing uncertainties in terrestrial carbon cycle projections under ongoing climate change. Here, based on flux observations from global terrestrial sites with a focus on forest ecosystems, we selected mean annual temperature (MAT), latent heat flux (LH), vapor pressure deficit (VPD), soil water content (SWC), and annual precipitation as representative indicators of hydrothermal conditions, and employed mixed-effects models to examine how these key environmental drivers influence GPP and ER. After analyzing the fixed effects, LH and MAT promoted GPP more strongly than ER (slope = 0.5 > 0.253, slope = 0.595 > 0.392, respectively), whereas VPD suppressed GPP more than ER (slope = −0.658 < −0.499), yet accounted for a greater proportion of variance in ER than in GPP (R² = 0.14 > 0.07). Although SWC had a significant (p < 0.001) positive effect on GPP, the effect size was minimal, and its impact on ER was insignificant. R² decomposition showed that marginal R² values were similar for the GPP and ER models (0.43 and 0.44), whereas the GPP model exhibited a substantially higher conditional R² (0.82 vs. 0.63), indicating that MAT exerted a stronger influence on GPP than on ER across ecosystem types. The combined analysis of fixed and random effects indicated that MAT affected GPP more variably than ER across ecosystem types, with the strongest responses in mixed forests and savannas, intermediate responses in evergreen needleleaf forests, and the weakest responses in evergreen broadleaf forests. Overall, this study advances our understanding of how environmental factors differently influence GPP and ER, and incorporating these differences can improve predictions of forest carbon fluxes and climate-carbon feedbacks.