Starting 2000 Cloud Servers for Benchmarking LLM Inference Speed