“The speaker predicts that within one to three years, LLM inference will be viewed as a core capability that determines a model's peak intelligence, rather than just a cost or convenience factor.”