Frontier AI labs require chips that can run models orders of magnitude faster to maintain an advantage over open-source models