If you want to do "budget ML" for training or long-running GPU jobs, you should check out Lambda GPU cloud.
Our on-demand instances are 50% the cost of GCP/AWS. Our 8x V100 instances are $12/hr and our 1x Quadro RTX 6000 instances are $1.25/hr. They come set up with TensorFlow/PyTorch/Jupyter notebook so you don't need to do much DevOps work to get started.