A PhD arguably gives you the opportunity to run very fast into one particular direction. Much longer than you would be allowed to run in any commercial setting.
It seems like most cutting edge research (and the ability to run fast) in machine learning is done in commercial settings these days, not academia.