One benefit of using cloud GPUs is the ability to train multiple models simultaneously. This is difficult to do with your own hardware at a reasonable cost, especially since a large portion of your time will be spent on hyperparameter tuning.
While you're getting started with your own hardware I'd love it if you could share any feedback on something we've been building[0] as far as cloud training goes. It's currently in free public beta and works directly with TensorFlow models so there's no lock-in.
[0] https://fomoro.com (jim at fomoro.com)