Looking at the pricing, if you only spin those instances up when you need them, you can go a while before you break even. Otherwise it only takes a few months d

Looking at the pricing, if you only spin those instances up when you need them, you can go a while before you break even. Otherwise it only takes a few months depending on the GPU. I would imagine that someone really serious about training (or any other CUDA workload) uses both.