Using GPU is essential for performance of some/many ML/AI algorithms...
Linear Algebra Shootout: NumPy vs. Theano vs. TensorFlow
(less time is better)
Sydney Region with Andrew Walker and Graham Polley | Google Cloud Platform Podcast
in the same podcasts there is indication that using GCP "serverless / services" is much less costly than AWS IaaS. While this is comparing "apples and oranges", they are both fruits :)
Azure push for PaaS suggest that enterprises prefer using services than managing IaaS.
Google Cloud Platform Blog: How we built a brand new bank on GCP and Cloud Spanner: Shine