AWS would make sense if you don't yet know if your ML algo scales well on a GPU: you can try the hardware without losing much money. Consumer cards are a good

AWS would make sense if you don't yet know if your ML algo scales well on a GPU: you can try the hardware without losing much money. Consumer cards are a good deal if you know your algo works well on GPUs in single (32-bit) floating point and you don't mind dicking around with hardware configurations to save money.