Yellow cake is a commonly used name for powdered uranium concentrate, produced with the uranium ore. It is the first step in the fabrication of nuclear fuel. As it contains fissile material its circulation needs to be controlled in order to avoid proliferation. In particular there is an interest in onsite determination of the geographical origin of a sample. The yellow cake elemental composition depends on its production site and can therefore be used to identify its origin. In this work laser-induced breakdown spectroscopy (LIBS) associated with chemometrics techniques is used to discriminate yellow cake samples of different geographical origin. 11 samples, one per origin, are analyzed by a commercial equipment in laboratory experimental conditions. Spectra are then processed by multivariate techniques like Principal Components Analysis (PCA) and Soft Independent Modeling of Class Analogy (SIMCA). Successive global PCAs are first performed on the whole spectra and enable one to discriminate all samples. The method is then refined by selecting several emission lines in the spectra and by using them as input data of the chemometric treatments. With a SIMCA model applied to these data a rate of correct identification of 100% is obtained for all classes. Then to define the specifications of a future onsite LIBS system, the use of a more compact spectrometer is simulated by a numerical treatment of experimental spectra. Simultaneously the reduction of spectral data used by the model is also investigated to decrease the spectral bandwidth of the measurement. The rate of correct identification remains very high. This work shows the very good ability of SIMCA associated with LIBS to discriminate yellow cake samples with a very high rate of success, in controlled laboratory conditions.