Hmm... GPT-4 may not be 100 T param. Yet, Cerebras' new chip is, critically, *efficient* enough to run 120 T parameters well. So, once the chip is available, I expect researchers at the major institutions will end-up using it... Because the $cost/model is capping current growth (alongside the understanding that "this is *already* good enough to start making products", which directs attention toward applications instead of insights). More efficient steam engines => more and bigger steam engines. Cerebras is likely to do the same, due to *efficiency* at scale.