英伟达的护城河守得住吗?

Will Nvidia’s moat persist?

黄仁勋 Jensen Huang · Dwarkesh 播客 · 2026-04-15 · 约 103 分钟 · 原视频 ↗

打开互动全文版（中英对照 + 朗读 + 问答）→

本期速览 · Overview

算力之王谈 GPU、AI 基建，以及物理 AI。

The compute kingpin on GPUs, the AI buildout, and physical AI.

要点 · TL;DR

• 英伟达的护城河是其生态系统和可编程性，而不仅仅是供应链。
  Nvidia's moat is its ecosystem and programmability, not just supply chain.
• 能源政策而非芯片供应是人工智能的长期瓶颈。
  Energy policy, not chip supply, is the long-term bottleneck for AI.
• 对中国的出口管制可能催生独立的技术栈。
  Export controls on China risk creating a separate tech stack.

核心观点 · Key points

• 从电子到 token 的转化极其难以商品化。
  The transformation from electrons to tokens is incredibly hard to commoditize.
• 英伟达的护城河是其生态系统、安装基础和可编程性，而不仅仅是供应链。
  Nvidia's moat is its ecosystem, install base, and programmability, not just supply chain.
• 芯片供应的瓶颈是 2-3 年的问题；能源政策才是真正的长期制约。
  Bottlenecks in chip supply are 2-3 year problems; energy policy is the real long-term constraint.
• CUDA 的价值在于其丰富的生态系统、庞大的安装基础和推动算法创新的能力。
  CUDA's value lies in its rich ecosystem, vast install base, and ability to enable algorithmic innovation.
• 英伟达的总拥有成本全球最佳；ASIC 并未提供有意义的成本优势。
  Nvidia's TCO is best in the world; ASICs don't offer a meaningful cost advantage.
• 中国拥有充足的算力和人才；孤立他们可能导致形成独立的技术栈。
  China has abundant compute and talent; isolating them risks creating a separate tech stack.

反共识 · Contrarian takes

• 英伟达 70%的利润率可持续；ASIC 利润率也很高。
  Nvidia's 70% margins are sustainable; ASIC margins are also high.
• 中国拥有充足的算力和能源；出口管制无法阻止他们。
  China has abundant compute and energy; export controls won't stop them.
• 向中国销售芯片有利于美国技术栈在全球的采用。
  Selling chips to China benefits US tech stack adoption globally.
• AI 不像核武器；将两者比较是「荒谬的'。
  AI is not like nuclear weapons; comparing them is 'lunacy'.
• 英伟达不哄抬价格；采用固定定价和先到先得分配。
  Nvidia doesn't price gouge; it uses fixed pricing and FIFO allocation.
• 英伟达的增长由架构和软件驱动，而不仅仅是光刻技术。
  Nvidia's growth is driven by architecture and software, not just lithography.

本期章节 · Chapters（共 31）

• 软件商品化与英伟达地位 Commoditization of Software and Nvidia's Position
• 对企业软件和工具制造商的影响 Impact on Enterprise Software and Tool Makers
• 英伟达供应链策略与采购承诺 Nvidia's Supply Chain Strategy and Purchase Commitments
• 供应链准备与扩展挑战 Supply chain preparation and scaling challenges
• 瓶颈与扩展 Bottlenecks and Scaling
• 竞争对手与 TPU Competitors and TPU
• 算法创新 vs 摩尔定律 Algorithmic Innovation vs. Moore's Law
• Crusoe Cloud 与推理优化 Crusoe Cloud and Inference Optimization
• CUDA 在超大规模云中的作用 CUDA's Role with Hyperscalers
• CUDA 的价值与客户优势 CUDA's value and customer advantages
• 公司选择其他加速器的原因 Why companies choose other accelerators
• 英伟达为何未早期投资 Why Nvidia didn't invest earlier
• 英伟达作为云提供商角色 Nvidia's role as cloud provider
• 不挑选赢家 Not picking winners
• 支持新云服务的兼容性 Compatibility of supporting neo clouds
• 构建 AI 协同研究工具 Building an AI co-researcher tool
• GPU 短缺与分配 GPU shortage and allocation
• 英伟达在 AI 行业的地位 Nvidia's position in AI industry
• 关于中国与芯片销售的问题 Question about China and chip sales
• 对中国与 AI 能力的回应 Response on China and AI capabilities
• 对算力差距与黑客攻击的担忧 Concern about flop difference and hacking
• 计算瓶颈与中美 AI 竞争 Compute bottleneck and US-China AI competition
• 关于 AI 芯片出口到中国的辩论 Debate on AI chip exports to China
• 美国必须在 AI 全层保持领先 US must stay ahead in AI across all layers
• 关于计算出口到中国的辩论 Debate on compute export to China
• 出口管制与芯片竞争 Export controls and chip competition
• 向后节点迁移的可行性 Backward node migration feasibility
• 多种芯片架构 Multiple chip architectures
• 架构与推理市场 Architecture and Inference Market
• 若深度学习革命未发生 If Deep Learning Revolution Didn't Happen
• 结束语 Closing

阅读全文双语转录 →