The whole premise for this article is just wrong. The only way demand for deep learning chips will drop is if the current transformer model hits a theoretical c

The whole premise for this article is just wrong. The only way demand for deep learning chips will drop is if the current transformer model hits a theoretical ceiling. Otherwise you can always improve the models "for free" just by scaling compute, without any improvements in architecture or training algorithms. And since there is nothing in current research that would point to such a theoretical ceiling, everyone will be craving GPUs in the hope of beating the next guy's product in the AI race. If a paper comes out that says something like "yeah, we're no longer seeing improvements after X trillion parameters" then - and only then - can you begin to doubt the shovelmakers.