Depends on your use case. The more accurate models these days require lots of data with huge models. Machine Translation, Image Generation, Segmentation, etc. Each experiment takes a couple days/weeks to train, so we were encountering long stretches of constant usage.