Tbf the energy issues are getting better, or at least there are some more efficient models being created. Back in April there was a version of Llama that only needed 8gb to almost match GPT4