blog/post/avoid-costly-llm-apis-scalable-cpu-gpu-inference/index.html