If I'm understanding correctly, each calls to a LLM makes a 3000W GPU run at full speed during the time of inference (let's say from 5 to 60 sec) while Google searches are historically mostly cached so very inexpensive for most searches.

Isn't the environmental impact 10 to 100 times what we were used to?