In some cases, you can patent a major application for that algorithm and then license that patent. From what I know Larry and Sergey patented PageRank for Web Search. My advisor has patent on a Graph Cuts algorithm for image segmentation, which is licensed by Cornell. If you have a fast matrix multiplication algorithm e.g. for use in Fully Connected layers for Neural Nets. You can patent that application. When FB hired people from NYU to work on deep learning, the acquisition involved getting a fast fourier transform optimized version of convolutional layers. Not sure how big of a component it was.