It's great when you don't need instant elasticity and traffic is very predictable. I think it's very useful for batch processing, especially owning a GPU clust

It's great when you don't need instant elasticity and traffic is very predictable. I think it's very useful for batch processing, especially owning a GPU cluster could be great for ML startups. Hybrid cloud + bare metal is probably the way to go (though that does incur the complexity of dealing with both, which is also hard).