Dataproc has per minute billing + very fast startup times (sub 90s for decent-sized clusters). So with Dataproc, you can rethink how you run Hadoop. Rather than starting with a cluster, which you fill up with jobs, you can have "cluster setup" as part of the job path. This is possible because you don't pay a large penalty in startup, nor do you pay a large penalty in per-hour billing granularity. Oh and Google is much more inexpensive (http://fortune.com/2016/01/08/google-amazon-cloud-price-war/).
Something like this is impossible on AWS without pre-warming, pre-setup, or maybe even calling someone - https://www.youtube.com/watch?v=xT6tQAIywFQ&feature=youtu.be
Here's another example: EMR versus Dataproc.
Dataproc has per minute billing + very fast startup times (sub 90s for decent-sized clusters). So with Dataproc, you can rethink how you run Hadoop. Rather than starting with a cluster, which you fill up with jobs, you can have "cluster setup" as part of the job path. This is possible because you don't pay a large penalty in startup, nor do you pay a large penalty in per-hour billing granularity. Oh and Google is much more inexpensive (http://fortune.com/2016/01/08/google-amazon-cloud-price-war/).