Local llms actually have a larger environmental impact than the data centers in power usage/hardware, since they optimize for LLMs. The models are also based off essentially DDOSing half the internet by constantly scraping sites, and being trained off GPL licensed code without permission. This is a "good" use of llm, but it's more like if the hammer you're using took 2kw of power and 200ml of water every swing, you'd probably want to avoid using that hammer even if it was very helpful.