Because this stuff is very compute-intensive and that is simply expensive.

Also, I don't think there is as much free training data publicly available as Google or Amazon have gathered privately.