Theoretical Bottlenecks for Scaling LLM Inference to Get Higher Token per Second · HackerLangs