Savanna ecosystems cover ∼20% of the total land surface and account for ∼30% of the terrestrial global net primary production. They are also highly sensitive to climate change, since their carbon (C) sink capacity may decline under rising temperatures and irregular rainfall. These responses, which will define the future climate role of the savanna ecosystems, are currently not well understood. The Colombian Orinoco River basin (“Llanos”) natural savannas are being rapidly converted to agriculture. The impact of this transformation on C fluxes and accumulation is not clear. It is thus urgent to understand the Llanos natural savanna ecosystem services, including their C cycle and underlying mechanisms. Here we report and analyze 2 years of measurements of carbon dioxide fluxes from a naturally-restored (secondary) Llanos High Plains savanna ecosystem, using eddy covariance. Meteorological conditions, particularly rainfall, were quite variable during the measurement period. During the first year of measurements, the savanna was a weak carbon source (35 gC m−2), while during the second year, the system was a comparatively strong carbon sink (−273 gC m−2), despite receiving less rainfall than during the first year. As expected, the savanna net ecosystem exchange (NEE) was highly dependent on global solar radiation, soil water content, and ecosystem respiration. We found that after ∼10 days of nominal drought, i.e., with less than ∼5 mm/day of precipitation, the NEE became highly dependent on drought duration. The ecosystem reached a critical condition of low photosynthetic activity after ∼60 days of nominal drought. Based on these findings, we developed and applied a simple standard meteorology-based model that properly reproduced the observations. Our results indicate that a shift to a climate with similar total precipitation but split into extreme dry and wet seasons might eventually suppress the savanna C uptake capacity.