Wasted tokens are preferred for local models, I need the GPU mainframe in my bedroom to heat it as I live in a third world country with unreliable heating (Switzerland).