Inference of surface CO2 fluxes from atmospheric CO2 observations requires information about large-scale transport and turbulent mixing in the atmosphere, so transport errors and the statistics of the transport errors have significant impact on surface CO2 flux estimation. In this paper, we assimilate raw meteorological observations every 6 hours into a general circulation model with a prognostic carbon cycle (CAM3.5) using the Local Ensemble Transform Kalman Filter (LETKF) to produce an ensemble of meteorological analyses that represent the best approximation to the atmospheric circulation and its uncertainty. We quantify CO2 transport uncertainties resulting from the uncertainties in meteorological fields by running CO2 ensemble forecasts within the LETKF-CAM3.5 system forced by prescribed surface fluxes. We show that CO2 transport uncertainties are largest over the tropical land and the areas with large fossil fuel emissions, and are between 1.2 and 3.5 ppm at the surface and between 0.8 and 1.8 ppm in the column-integrated CO2 (with OCO-2-like averaging kernel) over these regions. We further show that the current practice of using a single meteorological field to transport CO2 has weaker vertical mixing and stronger CO2 vertical gradient when compared to the mean of the ensemble CO2 forecasts initialized by the ensemble meteorological fields, especially over land areas. The magnitude of the difference at the surface can be up to 1.5 ppm.