It was not a "prove it" challenge.

I'm pointing out that there's no known information theoretic constraint about the impossibility of frontier AI models being improved to fit/run on a small GPU.

Please do not make up plausible sounding science facts.