I wonder to what degree the rise of CUDA played into this. A lot of CV tasks are better performed on the gpu and today nvidia is the goto platform for that.
Yeah, they have a parallel strategy (for now) with their Xeon Phi hardware, but there's still a performance penalty vs Nvidia. A big problem is that all the software is written for CUDA and optimized for particular Nvidia cards - hence, software / hardware synergy.
Wouldn't be surprised to see them acquire more deep learning algo startups as well.