Inference, development cycles, any of the application domains of PyTorch that don't involve training frontier models... all of those are complicated by excessive container layers.

But mostly dev really sucks with writing out an extra 10GB for a small code change.