The short answer is that there are more things that matter than parameter count, and we are probably nowhere near the most efficient way to make these models. Also: the big AI labs have shown a few times that internally they have way more capable models