LLM-powered NPCs will quickly fall out of fashion as people realize they're literally just talking to chatGPT.
The either forced always-online requirement with privacy violating telemetry for server-side LLMs, or immensely high GPU memory requirements for local LLMs will also cripple their games.
How small can you make an LLM before it starts having issues with grammar and coherency? I would argue that the bare minimum still would be rather large, and in videogames we're already using vram for other resources. In a 3D game especially I imagine very little vram is left to utilize.