Not really an option right now. The current generation of cameras out there are basically a Raspberry-Pi level computer with a better camera and a hardware h264 encoder. The moment you start doing something fancy, like running any non-trivial motion-detection algorithm, you are bound to run into performance or thermal (read: overheating) issues. Let's not even talk about machine learning.
Just think how much money Nest would save in server time with such a setup :)
Just think how much money Nest would save in server time with such a setup :)