Jensen Huang explains why extreme co-design across chips, systems, software, and infrastructure is essential for scaling AI performance beyond Moore's Law.
要点 · TL;DR
英伟达的主导地位源于跨硬件、软件和系统的极致协同设计。 Nvidia's dominance stems from extreme co-design across hardware, software, and systems.
CUDA 的安装基础是英伟达的主要护城河，建立在数十年的生态系统和信任之上。 CUDA's installed base is Nvidia's primary moat, built on decades of ecosystem and trust.
AI 扩展现在有四个定律，都受限于计算而非数据。 AI scaling now has four laws, all limited by compute, not data.
核心观点 · Key points
极端协同设计在整个技术栈中是必要的，因为大规模分布式计算需要共同优化每个组件。 Extreme co-design across the entire stack is necessary because distributed computing at scale requires optimizing every component together.
英伟达最大的护城河是 CUDA 的安装基础，而不仅仅是技术；这是数十年建立的生态系统和开发者信任。 Nvidia's biggest moat is the CUDA installed base, not just technology; it's the ecosystem and developer trust built over decades.
AI 已从基于检索的计算转向生成式计算，使算力成为主要价值驱动力而非存储。 AI has shifted from retrieval-based computing to generative computing, making compute the primary value driver instead of storage.
存在四种缩放定律：预训练、后训练、测试时和智能体式缩放，最终都受算力限制。 There are four scaling laws: pre-training, post-training, test-time, and agentic scaling, all ultimately limited by compute.
工作的目的和任务是相关但不相同的；AI 将增强角色而非消除它们，如放射科医生所示。 The purpose of a job and its tasks are related but not the same; AI will augment roles, not eliminate them, as seen with radiologists.
反共识 · Contrarian takes
推理比预训练更难，因为思考和推理需要的算力比记忆更多。 Inference is harder than pre-training because thinking and reasoning require more compute than memorization.
数据不是瓶颈；合成数据将占主导，训练现在受算力限制而非数据。 Data is not a bottleneck; synthetic data will dominate, and training is now limited by compute, not data.
电网浪费可以通过利用过剩容量和动态数据中心负载转移来解决，而不仅仅是建设更多电厂。 Power grid waste can be solved by using excess capacity with dynamic data center load shifting, not just building more plants.
开源 AI 模型对于广泛行业采用至关重要，英伟达从理解模型演进中受益。 Open-source AI models are essential for broad industry adoption, and Nvidia benefits from understanding model evolution.
如果 AGI 定义为创建十亿美元公司，那么它可能已经实现，因为 AI 智能体现在可以构建病毒式网络服务。 AGI may already be achieved if defined as creating a billion-dollar company, as AI agents can now build viral web services.
本期章节 · Chapters（共 41）
极限协同设计Extreme Co-Design
从加速器到计算公司From Accelerator to Computing Company
CUDA 早期困境与 GeForce 角色CUDA's early struggles and GeForce's role
垂直整合与平台开放Vertical integration and platform openness
对扩展定律的信念Belief in scaling laws
倾听行业与灵活架构Listening to the industry and flexible architecture
智能体 AI 扩展的障碍Blockers for Agentic AI Scaling
供应链与电力瓶颈Supply Chain and Power Bottlenecks
数据中心动态电力分配Dynamic Power Allocation in Data Centers
电网机遇Utility Grid Opportunity
马斯克建造巨像的方法Elon Musk's Approach to Building Colossus
与英伟达系统工程相似Parallels with NVIDIA's Systems Engineering
第一性原理 vs 持续改进First Principles Thinking vs Continuous Improvement
系统设计中的复杂 vs 简单Complexity vs Simplicity in System Design
中国技术生态与创新China's Technology Ecosystem and Innovation
开源愿景Open Source Vision
拒绝 CEO 职位Declining the CEO offer
CUDA 安装基础为主要护城河CUDA installed base as the primary moat
向 AI 工厂演进Evolution to AI factories
太空边缘 AIAI at the edge in space
英伟达增长与计算未来Nvidia's growth and the future of computing
英伟达潜力与市场份额Nvidia's potential and market share
代币作为产品与代币 iPhoneTokens as product and the iPhone of tokens
应对压力与责任Dealing with pressure and responsibility
分解问题与管理负担Decomposing Problems and Managing Burden
谦逊与向他人学习Humility and Learning from Others
AI 垃圾与游戏图形工具AI slop and game graphics tools
最伟大游戏：毁灭战士与 VR 战士Greatest games: Doom and Virtual Fighter
AGI 时间线与定义AGI timeline and definition
放射科医生短缺与 AI 影响Radiologist shortage and AI impact
编程定义与未来程序员Coding definition and future programmers
传统编程的价值Value of traditional programming
工作焦虑与学生建议Job anxiety and advice for students
AI 作为人人工具AI as a tool for everyone
人类意识 vs AIHuman consciousness vs AI
智能 vs 人性Intelligence vs. Humanity
死亡与继承Mortality and Succession
对人类的希望Hope for Humanity
太空人形机器人与数字意识Humanoid in space and digital consciousness