A carbonate (CO3) prediction model was developed for soils throughout the contiguous United States using mid-infrared (MIR) spectroscopy. Excellent performance was achieved over an extensive geographic and chemical diversity of soils. A single model for all soil types performed very well with a root mean square error of prediction (RMSEP) of 12.6 g kg-1 and was further improved if Histosols were excluded (RMSEP 11.1 g kg-1). Exclusion of Histosols was particularly beneficial for accurate prediction of CO3 values when the national model was applied to an independent regional dataset. Little advantage was found in further narrowing the taxonomic breadth of the calibration dataset, but higher precision could be obtained by running models for a restricted range of CO3. Ten absorbance peaks enabling CO3 prediction by mid-infrared (MIR) models were identified and evaluated for individual predictive power and the directness with which peak shape was translated into the loading vectors and cumulative loading function of a partial least squares (PLS) model. An absorbance peak centered at 1796 cm-1 was found to be the most informative with an RMSEP of 13.5 g kg-1 for carbonate prediction. This predictive power is attributed to the strength and sharpness of the peak, and an apparent minimal overlap with confounding co-occurring spectral features of other soil components. Soil CO3 is an excellent example of a soil parameter than can be predicted with great effectiveness and generality, and MIR models could replace direct laboratory measurement in many contexts. This eCommons item holds the datasets from this work.