Analysts cite NVIDIA–Groq integration cutting latency as Samsung vs TSMC yields drive risk. NVIDIA Groq inference chip, prefill vs decode, Samsung 3nm yield