Horrific comparison point. LLM inference is way more expensive locally for single users than running batch inference at scale in a datacenter on actual GPUs/TPUs.