According to @_avichawla, moving LLMs to faster GPUs barely helps TTFB because network, cold starts, and RAG hops dominate latency.