Nvidia 再創季度紀錄，代幣需求爆炸性增長

🔑 增強重點摘要

• •OpenRouter processed 13 trillion AI tokens in the week ending February 9, 2026, doubling from 6.4 trillion in early January, signaling a sharp surge in third-party AI inference activity[1].
• •Launch of OpenClaw, an open-source AI agent system in November 2025, has driven much of the recent AI agent activity and coincided with rebounding Nvidia H100 GPU rental prices[1].
• •Nvidia unveiled the Rubin platform with six new chips promising up to 10x reduction in inference token cost compared to Blackwell, with early deployments by AWS, Google Cloud, Microsoft Azure, and Oracle[3].
• •Nvidia guided for $78 billion revenue in fiscal Q1 2027, exceeding analyst consensus of $72.78 billion, alleviating concerns over AI demand sustainability[5].

🛠️ 技術深入

• •Grace Blackwell with NVLink delivers an order-of-magnitude lower cost per token for inference, positioning it as the leading platform[3].
• •NVIDIA Run:ai GPU fractioning enables 77% of full GPU token throughput (152,694 tokens/sec at 64 GPUs) using 0.5 GPU allocations, with linear scaling[4].
• •For Phi-4-Mini model, 0.25 GPU fractions via Run:ai support 72% more concurrent users than full GPUs, achieving ~450K tokens/sec at 32 GPUs with P95 TTFT under 300ms[4].
• •Vera Rubin platform chips aim for up to 10x inference token cost reduction versus Blackwell[3].

🔮 前景展望AI analysis grounded in cited sources

⏳ 時間線