With global warming, precipitation variability is increasing, affecting water availability and consequently ecosystem functioning and carbon uptake. Semi-arid ecosystems are very sensitive to changes in water availability since both gross primary productivity (GPP) and ecosystem respiration (Reco) are strongly linked to it, contributing most to the interannual variability and trend of the global land carbon sink. However, their carbon-dioxide (CO2) flux dynamics and specifically Reco are not well understood and poorly represented in models. Here we compiled a long-term global dataset of eight eddy-covariance sites located in semi-arid savannas to analyze the effects of precipitation variability on their net ecosystem exchange (NEE, with NEE = Reco-GPP). We assessed the effects of precipitation variability, represented by precipitation amount, frequency, intensity, and dry spell length, on both annual and seasonal CO2 fluxes. We used linear mixed effect models to identify relationships between the precipitation metrics, environmental factors, and CO2 fluxes. We further employed structural equation models to identify direct and indirect effects of precipitation variability on ecosystem CO2 fluxes in different phenological seasons. The results indicate that precipitation amount and intensity mostly lead to increasing GPP and Reco on the annual scale, with different net effects on NEE depending on the sites. On the seasonal scale, both precipitation frequency and intensity explained more NEE variations than precipitation amount only, while the contribution of dry spell length was negligible. The three precipitation metrics together lead to highest model improvement. Across seasons, soil water content was the main mediator of precipitation impacts on CO2 fluxes, with air temperatures co-dominating in the drydown season. Overall, this study shows that precipitation intensity, frequency, and their seasonal timing are crucial for explaining ecosystem CO2 flux dynamics of semi-arid savannas. This can help improve predictions of the effects of increasing precipitation variability on the global carbon cycle.