GPU Inference: Understanding the Real Cost and Speed of Running LLMs

What actually drives LLM inference cost and latency — utilization, batching, cold starts, TTFT vs throughput — and how to get frontier speed without owning