I think they mean the assumed parameter size of GPT-4 is so large you couldn't run it on commodity hardware even if you could get hold of the model.