Determination of the present global budget of atmospheric carbon dioxide (CO2) from the small and persistent concentration gradients that exist in the atmosphere is discussed. The CO2 concentration at any site results from a combination of two factors: local sources or sinks and long-range transport. To separate these two effects, an atmospheric transport model is needed. The extensive sets of global CO2 measurements of the National Oceanic and Atmospheric Administration's (NOAA) Geophysical Monitoring for Climatic Change (GMCC) division and of the Upper Atmosphere and Space Research Laboratory of Tohoku University are combined with a two-dimensional transport model to derive, in an “inverse” calculation, the latitudinal and seasonal distributions of sources and sinks of CO2 necessary to reproduce the observed concentrations. The model transport parameters were previously derived from a three-dimensional general circulation model. It is found that the southern oceans are a sink of carbon of 0.8-1.5 Gt yr-1 (1 Gt equals 1015g) and that the equatorial areas are a source to the atmosphere of 1.4-2.8 Gt yr-1. Tropical deforestation as a major source of CO2 must be smaller than that because the oceans account for a significant part of the equatorial flux. There seems to be significant seasonally in the sources and sinks of CO2, both in the tropics and in the southern oceans. The seasonal net ecosystem production north of 25°N is found to be 6.2-8.2 Gt of carbon, but these estimates are probably somewhat too low. The source deduction problem is difficult to solve, especially for the middle and high latitudes in the northern hemisphere. This is due to a lack of observations over the continents, which occupy more than half of the global area at these latitudes and are the regions where the sources and sinks are most intense. Evidence is found in the results obtained for the GMCC and Tohoku data that the longitudinal variability of the data is large enough, even in equatorial and southern latitudes, to prevent a two-dimensional model from calculating a fully credible source/sink field. The longitudinal variations in the observations have to be accounted for with a three-dimensional transport model.