This study tested strategies for the prediction by NIRS of the chemical composition and nutritive value of large poaceae (sorghum, millet, sugarcane, miscanthus etc.). The main objective was to evaluate the possibility of prediction of samples when no specific calibration is available because the plant part of the species is not common enough. The first strategy was to build a global calibration gathering several species. With a database of 750 samples from 8 species or plant parts, the precision of the models was close to the values usually recorded in single-species calibrations. The second strategy was to test extrapolation capacity by removing one species from the previous database and validate on this species. The results depended on the species and on the parameters: while good results were obtained for protein content in most cases, results were less precise for more aggregative criteria (fibre, digestibility) particularly when removing species with extreme (high of low) values.