As a relative laymen in this hardware inference space, I am curious what exactly was Groq useful for vs. the typical hardware architecture? Or was this a “step in before they become more generally useful” situation for Nvidia/Groq?