Carbon nanotubes come in different species having different properties. So, it is useful to develop automated ways to quantify species purities and trace impurity content. Spatially scanned Raman spectra make hyperspectral data sets that can be used to discriminate between species and determine purities, and their analysis can be automated. A promising analytical machine learning approach is non-negative matrix factorization, which is a multivariate algorithm well adapted to hyperspectral Raman scattering data sets, and, significantly, available in open source. We prepare samples from different concentrations of pure nanotubes and acquire spatially scanned Raman scattering hyperspectra. We compare the known concentrations of the source dispersions to those determined from hyperspectra acquired from deposited materials on substrates. Here, we demonstrate and deal with several metrological issues: We show that the stability of the focusing conditions is critical. We show that if there are strong peaks that are not significant, normalization is helpful. We show that this approach compares favorably to the “best case” situation where the spectra are factored into a priori known library spectra. Scans of around 100 data points provided good bounds on concentrations down to about the parts per thousand level. Using more than one laser wavelength, so that different species are brought into resonance with each laser should enable higher relative purity measurements. However, there is an important consideration if the difference in laser energies is less than or comparable to the phonon energy. Overall, this approach is promising for the determination of chemical purity of carbon nanotubes and could be generalized to other chemical mixtures.