The article is subtitled "Its deep pool of data may let it lead in artificial intelligence" for a reason, you can't just pull a "well trained" model out of thin air, and proprietary data is absolutely defensible.
Instead, what is needed is user labeled images, videos, text, speech and of course, simulations (games, VR) for virtual robotics and reinforcement learning agents. These kinds of data and simulations are open sourced and growing. Once there is a sufficient collection of training data, it's cheap to train your neural nets.
Even today, there is much more data than scientists care to use. For example, ImageNet is over 1TB, but a smaller portion of it is used routinely (it has 21841 topics but usually just 1000 of them are used in research papers). So it's not the lack of images that's keeping them. China would have no advantage here. Everyone has approximately the same level, the best paper of last year is the average of this year. That's about AI in general - but regarding self driving cars and ads Google is cooking in secret so it might be a little more ahead.
umm:
google image search, all html embedded/related metadata, adwords itself serving as a refined metadata/tagging and ontology database building engine, recaptcha 'how many buildings are in this picture' puzzles, android debug settings where things are uploaded, google drive cloud storage, gmail, the google speaker gizmo and voice activated phone agent, google voice voip, google chat, maps annotations to pictures/gps coordinates, etc etc etc and yadda yadda yadda...
not sure what else you need as far as those sources..