英伟达 CEO 黄仁勋探讨三大 AI 扩展定律、推理需求增长十亿倍，以及与 OpenAI Stargate 项目的战略合作。
Nvidia CEO Jensen Huang discusses three AI scaling laws, the billion-fold increase in inference, and the strategic partnership with OpenAI's Stargate project.
要点 · TL;DR
推理计算需求将因思维链推理而激增十亿倍。 Inference compute demand will surge a billion times due to chain-of-thought reasoning.
英伟达的年度发布周期和极致协同设计创造了不可逾越的护城河。 Nvidia's annual release cycle and extreme co-design create an insurmountable moat.
每个国家都需要主权 AI 基础设施，就像能源和互联网一样。 Every country needs sovereign AI infrastructure, just like energy and internet.
核心观点 · Key points
现在有三条缩放定律：预训练、后训练和推理时思考。 Three scaling laws now: pre-training, post-training, and inference-time thinking.
由于思维链推理，推理算力需求将增加十亿倍。 Inference compute demand will increase by a billion times due to chain-of-thought reasoning.
OpenAI 很可能成为下一个价值数万亿美元的超大规模公司。 OpenAI is likely to become the next multi-trillion-dollar hyperscale company.
通用计算已经终结，未来是加速计算和 AI。 General-purpose computing is over; the future is accelerated computing and AI.
英伟达的年度发布周期和极致协同设计创造了不可逾越的护城河。 Nvidia's annual release cycle and extreme co-design create an insurmountable moat.
每个国家都需要主权 AI 基础设施，就像能源和互联网一样。 Every country needs sovereign AI infrastructure, just like energy and internet.
反共识 · Contrarian takes
竞争对手的 ASIC 可以免费，但客户仍会选择英伟达，因为智能总成本更低。 Competitors' ASICs could be free, but customers would still choose Nvidia due to total cost of intelligence.
中国在 AI 上并非落后数年，而是落后纳秒。 China is not years behind; they are nanoseconds behind the US in AI.
对华芯片出口管制适得其反，让华为获得垄断利润。 Export controls on chips to China are counterproductive, giving Huawei monopoly profits.
AI 将创造更多就业，而非摧毁，因为它拓展了创意的边界。 AI will create more jobs, not destroy them, because it expands the realm of ideas.
在所有通用计算被取代之前，算力过剩的风险极低。 The risk of a compute glut is extremely low until all general-purpose computing is replaced.
与中国脱钩是错误的理念；两国经济深度交织。 Decoupling from China is the wrong concept; the two economies are deeply intertwined.
本期章节 · Chapters（共 33）
三大缩放定律Three scaling laws
OpenAI与Stargate合作OpenAI Stargate partnership
驱动算力需求的双指数Two exponentials driving compute demand
OpenAI与英伟达直连OpenAI building direct relationship with Nvidia
华尔街质疑vs行业现实Wall Street skepticism vs. industry reality
超大规模向AI转型Hyperscale transition to AI
AI收入与投资展望AI Revenue and Investment Outlook
推理作为第二指数Reasoning as the second exponential
超大规模建设不足Hyperscalers underbuilt
加速数据处理计划Accelerated data processing initiative
循环收入担忧Circular revenue concerns
Token生成指数增长Exponential Growth in Token Generation
英伟达性能飞跃：从Kepler到BlackwellNvidia's Performance Leap: Kepler to Blackwell
极致协同设计解析Extreme Co-Design Explained
竞争护城河与供应链Competitive Moat and Supply Chain
规模与协同设计Scale and Co-Design
GPU vs ASIC与市场时机GPUs vs ASICs and Market Timing
芯片类别与ASIC局限Chip categories and ASIC limitations
机会成本与年度迭代Opportunity Cost and Annual Cadence
马斯克与ColossusElon Musk and Colossus
主权AI与地缘政治Sovereign AI and Geopolitics
主权AI与国家安全Sovereign AI and National Security
技术作为国家财富与贸易Technology as National Treasure and Trade
中国竞争与市场份额China Competition and Market Share
在中国竞争Competing in China
H1B签证费H1B Visa Fee
移民政策与H-1B滥用Immigration Policy and H-1B Abuse
中国人才管道萎缩Declining Chinese Talent Pipeline
爱国主义与美国梦Patriotism and the American Dream
美国再工业化与AI作为平衡器Reindustrializing America and AI as an equalizer