I could believe that 90% of code will be generated by LLMs, because it takes almost no effort to generate some common boilerplate using an LLM. It's kind of like saying that 90% of code is BSD/MIT/GPL licensed because many people compulsively fork repos on github.

This just means that a smaller percentage of code will be useful/must-have/must-read code. Kind of like the market for films and books and games -- many more get made every year, which means that many more will get ignored every year.