SemiAnalysis has published an article challenging the market's focus on "time-to-first-token" (TTFT) latency in large model inference. The report argues that the primary concern should not be speed alone, but the balance between single-user interaction speed and overall throughput efficiency. According to SemiAnalysis, TTFT has a limited impact on user experience, with token generation speed during the generation phase being more critical. The analysis suggests that only 10%–20% of inference tasks are truly latency-sensitive, while most scenarios benefit more from optimizing throughput and cost-efficiency.