Fully on-prem can be done if you've got the LLM compute power in place.