why not shutdown the GPU clusters training large models at night or when its cloudly, and powering them back on when its shiny and windy?