MoonshotAI unveils Kimi's large-scale LLM serving architecture

I have been wondering the reason why online generative AI can serving so many requests. This really gives me an explanation.