A 501B model that only activates 23B parameters per token sounds like a bargain, until you look at what it takes to host.