Non-negative matrix factorization(NMF) is a widely used tool because it constructs interpretable part-based representations of high-dimensional count data. The interpretability comes from the fact that the non-negative data matrix is factorized in to two matrices that is also constrained to be non-negative. The factorization is usually found with an EM-algorithm or stochastic gradient decent. The method is especially popular with-in cancer genomics to identify both mutational signatures and gene programs. In applications such as spatial transcriptomic, NMF does not incorporate the known location of the transcripts. Incorporating the location in the analysis creates spatially crisp landscapes that identifies multicellular hubs with similar gene expression. I will present a spatial NMF method that incorporates the spatial information into the model in a very simple way, which makes it possible to scale to several million observations at once. The new method includes the spatial information by including Gaussian smoothing into the EM-algorithm. I will show applications of the method to three colorectal cancer samples with around two million observations and compare the results to regular NMF.