Collab is awesome. It's not really for any serious training, it's for evaluating and sharing ML models that (may) need a GPU, and learning. But for many situations, it replaces the need to have a laptop with a GPU - it's good enough for getting set up before you engage the real compute to do serious training. If collab was not free, I would consider paying a subscription for it, which is maybe the end game.
They have a paid version now that is supposed to be faster and more lenient about disconnecting you, but I recently read a comparison that said it wasnt worth it.
Yes, given Google's reputation, relying on anything run by Google that doesn't make (a lot!) of money is a high risk venture. It's one thing to do so when you have options to fall back to, but if you are buying yourself a machine learning laptop as a long term (2 year+) investment on the basis that you don't need GPU hardware because of Colab ...
Google collab doesn't need to make money. It needs to suffocate space so much, that no potential competitor with commercial offering ever has a chance. The money they make here is money they save by not having to buy competitor for nine figures later.
Considering the computationally intensive nature of ML, does it make more sense to train on specialized cloud-based processors?