Hope we see affordable hardware for doing inference in reasonable time on a 400B parameter model soon.