I've seen this several times by now. It's pretty ridiculous. One of my clients was an AI startup. They got $100K in AWS credit. So yeah, of course they aren't going to move off AWS until that credit runs out, even though GCP's offering for GPU is much better: for one thing on GCP you can vary the number of cores allocated to a machine with a GPU (or several). Inexplicably on AWS you can't do that. One GPU - you get 8 cores. Why? Because fuck you, that's why. As a result, you can't use that V100 fully if you do any significant data augmentation - CPU is often a bottleneck. IO and networking are faster on GCP as well. And if your GPU is not "free", you can cut costs by running on preemptible instances (few of their training runs are longer than 8 hours). On top of that if scale is needed (which it was), you can wrap it all into k8s, too.

Disclosure: ex-googler (left 4 years ago), own no GOOG, use GCP to run my own business. Clients use whatever they want.