Nvidia CEO Huang Sees AI Infrastructure Yielding $400 Billion Return

英伟达CEO黄仁勋预计AI基础设施将带来4000亿美元回报 历经60年的预录数据检索计算模式正让位于实时定制的Token生成，这将扩大经济产出并增加整体劳动力需求。 英伟达CEO黄仁勋阐述了计算领域60年来根本性的范式转变，即从静态数据检索走向通过AI工厂进行动态、实时的智能生成。他详细介绍了涵盖能源、芯片、基础设施、模型和应用的五层产业架构，并将AI在全球的铺开与历史上电力和互联网的普及相提并论。在回应有关就业的担忧时，黄仁勋认为AI自动化的是具体任务而非整个岗位，并以放射学和软件工程为例，论证了AI如何提高人类员工的生产力，而非取代整体劳动力需求。 黄仁勋指出计算正迎来60年来最根本的范式转变，传统数据中心的静态文件检索正在全面转向由AI驱动的实时、按需动态生成计算。 (5:00) NVIDIA将AI工厂比作现代工业发电机，输入电能并输出承载智能的Token，预计未来将有上千亿自主智能体联网协同，构成继电力与互联网之后包裹地球的第三层基础设施。 (10:00) AI产业展现出从能源、算力与芯片、物理基建到AI模型与应用的“五层蛋糕”架构，其学习数据规则与结构的能力正向规模达80万亿美元的物理世界与生命科学市场迅速拓展。 (20:00) 黄仁勋驳斥了AI导致人类灭亡等末日叙事，提出“人类不会被AI淘汰，但会被使用AI的人取代”，并以放射医学和软件开发为例，说明自动化将激发更大需求并重塑与升华人类劳动。 (25:00)

NVIDIA CEO Jensen Huang argues that artificial intelligence marks a transition from a 60-year paradigm of data retrieval to on-demand intelligence generation, creating a multi-trillion-dollar industrial layer that will expand economic output rather than displace human labor. The Shift from Retrieval to Dynamic Generation (5:00) Huang traces modern computing back to the 1964 IBM System/360, a foundational model that established six decades of pre-recording and data retrieval. Under that system, human-authored files, media, and code were saved to disks and later served to users via data centers and recommender algorithms. In contrast, generative AI calculates original outputs tailored in real time to specific context, user intent, and constraints. This shift moves AI beyond simple conversational translation, such as summarization or image generation, into agentic systems capable of executing multi-step reasoning, manipulating software tools like AutoCAD and spreadsheets, and operating physical hardware in robotics and autonomous driving. Huang forecasts that the internet will evolve from supporting roughly one billion human users to coordinating tens of billions—or even 100 billion—autonomous AI agents collaborating across enterprise and physical workflows. The Dynamo Analogy and Physical Scale (10:00) Drawing parallels to the rollout of electricity 300 years ago and the internet 35 years ago, Huang describes AI as a third global infrastructural fabric. While early dynamos converted motion and combustion into electrons, Huang likens modern AI factories to dynamos that consume electrons and output mathematical tokens representing text, mathematics, protein biology, and spatial physics. At the hardware layer, Huang highlights the engineering intensity required for these facilities. A standard NVIDIA server rack contains 72 chips, comprises 1.5 million components, weighs two tons, and costs $4 million. Scaling this infrastructure, Huang estimates that constructing a 1-gigawatt AI factory requires roughly $50 billion in capital expenditure, which he projects can yield $300 billion to $400 billion in intelligence token revenue as NVIDIA produces around 8 million chips annually. The Five-Layer Industry Stack (20:00) To frame capital allocation, Huang outlines an AI market architecture organized into five distinct layers: sustainable energy at the base, followed by compute and silicon, scarce physical infrastructure, foundational models, and consumer or enterprise applications. Huang projects that while this ecosystem absorbs approximately $1 trillion in capital deployment in the near term—including $100 billion invested by venture capital into applications over the past year—the stack could eventually address a $20 trillion annual market. Huang emphasizes that large language models are only one part of this landscape. He points toward an estimated $80 trillion physical-world market where models learn the rules and structure of materials, cellular biology, and spatial interactions to solve problems beyond text. Labor Impacts, Task Automation, and Safety (25:00) Huang dismisses existential catastrophe narratives and claims that computer scientists do not understand how AI works, citing consistent yearly software and hardware advancements alongside decreasing model hallucination rates. Addressing workforce displacement, he distinguishes between isolated mechanical tasks and an overarching professional purpose, stating that workers will not necessarily lose jobs to AI directly, but to peers who leverage AI tools. Citing the precedent of computer vision in radiology, Huang notes that predictions of human job elimination failed: automated image analysis raised diagnostic efficiency and department profitability, which in turn increased overall patient volume and hiring. He asserts that automating low-level execution—such as writing syntax or drafting blueprints—elevates trades by turning programmers into higher-level system designers, carpenters into home designers, and opening computing to anyone who speaks a natural language. Huang positions generative AI as an expanding production engine that bridges the digital and physical economies, though realizing his projected returns depends on overcoming real-world infrastructure constraints, securing massive energy capacity, and translating raw token generation into sustainable enterprise value.

英伟达（NVIDIA）首席执行官黄仁勋（Jensen Huang）在红杉资本对话中指出，全球计算正经历自1964年IBM System/360以来六十年未有的根本性范式转移，即从“预先记录与检索数据”全面转向基于上下文的“实时动态智能生成”。他认为，依托能源与算力构建的人工智能工厂将形成继电力和互联网之后的第三层全球基础设施，推动万亿美元级产业重构，并实质性重塑人机协作与劳动力分工。 计算范式重构与智能体演进 (5:00) 黄仁勋指出，现代计算架构由1964年问世的IBM System/360奠定并沿用了六十年。传统数据中心的核心机制是“预先记录与检索”：开发者与内容创作者编写代码、拍摄图像或录制音频后存入磁盘，系统仅负责通过算法索引与推荐预存文件。生成式AI则彻底颠覆了这种静态存储逻辑，转为根据用户的具体上下文和查询，在毫秒级时间内动态计算并生成完全定制化的原创内容。黄仁勋预测，未来所有数字媒介中的文字、像素、广告与视听内容，都将从按需调用转向现场生成。 技术演进路径也已超越早期的文本转换与摘要生成。黄仁勋解释称，AI正从聊天问答工具升级为具备分步推理、工具调用与实际作业能力的自主智能体（Agent）。通过生成操作指令，AI能够驱动网络浏览器、电子表格、AutoCAD与Photoshop等数字工具，甚至控制机器人与自动驾驶汽车。这种从生成娱乐性内容向产生可变现经济价值的跃迁，推动企业和个人开始以每小时20至30美元的单价雇佣AI劳动力，使其成为历史上增长最快的软件商业形态。 智能发电机与全球第三层基础设施 (10:00) 黄仁勋将当前AI工厂的建设与人类历史上的基础设施革命相提并论：300年前西门子（Siemens）发电机将物理燃料与机械运动转化为电能，催生了全球电力网络；35年前源自美国的通信革命则编织了互联人类的互联网网络；而如今由算力驱动的AI工厂则构成包裹地球的第三层网络——智能网络。与传统发电机“输入原子、输出电子”相对照，英伟达的AI工厂则是“输入电力、输出Token”。这些数学Token不仅表达自然语言，还能表征数学公式、蛋白质生物结构、气候物理规律与三维空间物理法则。 针对未来的互联规模，黄仁勋预测互联网的服务对象将从现有的数十亿人类用户，大幅扩展至数百亿乃至1000亿个全天候在线的自主AI智能体。这些智能体将在企业的防护栏与沙盒环境中彼此沟通、协同作业，并在实体世界中驱动车队、工厂机械与智能楼宇系统，形成密度极高的机器间协作网络。 五层产业堆栈与AI工厂投资回报 (15:00) 为了向投资者拆解万亿美元级的产业演进，黄仁勋提出了以能源为底座的“五层蛋糕”架构：第一层为能源层（核能、风能、太阳能等电力供给，带动了GE Vernova、西门子与三菱等工业重工）；第二层为芯片与计算系统（芯片、交换机与硅光子）；第三层为物理基础设施（稀缺的土地、机房与电力配额）；第四层为基础AI模型（如OpenAI和Anthropic）；第五层为终端应用软件（过去一年吸引了约1000亿美元风险投资）。黄仁勋估计，全行业今年正向这五层堆栈注入约1万亿美元资金，长期年度潜在市场规模可达20万亿美元，而涉及物理实体世界的AI应用市场规模预计将达80万亿美元。 在硬件与建造成本方面，黄仁勋详述了英伟达高密度机架的工业指标：单个机架集成72颗专用芯片，整机重达2吨，包含150万个精密零部件，单价为400万美元，英伟达预计本年度将向全球出货约800万颗芯片。针对资本支出的合理性，黄仁勋给出了高回报预测：建设一座规模为1吉瓦（Gigawatt）的AI工厂耗资约500亿美元，但其生命周期内预计可产出3000亿至4000亿美元价值的智能Token，从而为基础设施建造方提供快速的投资回收能力。 驳斥末日论与劳动力价值重构 (25:00) 面对舆论界对于AI生存威胁的担忧，黄仁勋明确驳斥了科幻概念中所谓“终结者模式”或“人类有20%概率灭绝”的末日论断，称其毫无事实根据。他也否认了“科学家不清楚AI底层运作机理”的言论，指出英伟达软硬件系统每年的持续迭代与可控演进本身就印证了工程层面的确定性；同时，行业对护栏机制的大量投入已让模型幻觉在过去两年内显著缩减，使现代系统具备了自我反思和多步骤推理的能力。 关于AI对就业的冲击，黄仁勋提出了明确命题：“你或许不会被AI取代，但你绝对会被使用AI的人取代。”他主张区分工作的“具体任务”（Task）与“核心目的”（Purpose）：十多年前有专家曾预言计算机视觉将彻底淘汰放射科医生，但实际情况是计算机视觉被整合入诊断流程后大幅提升了阅片效率，促使医院承接更多病患并扩充放射科规模，医生雇佣人数反而上升。黄仁勋认为软件工程亦然，敲击代码仅是任务，其核心目的是创意性解决问题；加之自然语言界面的普及使全人类无需学习C++即可操控计算机，AI工具将让木匠升级为房屋设计师、水暖工升级为图纸设计师，从而整体拉升各类工种的生产力水平。 该对话勾勒出计算架构从静态检索向动态生成跃迁的宏大蓝图，强调物理世界模型化与百亿级智能体对全球经济的提振潜能。然而，黄仁勋所预测的巨额Token产出与万亿美元市场规模，仍高度依赖于吉瓦级低成本能源的稳定供应、半导体供应链的极限交付能力，以及基础模型在复杂实体产业中持续变现的确定性，这些关键外部条件仍有待长期的市场实践与工程验证。

## Background 本期内容由红杉资本（Sequoia Capital）于 2026 年 6 月 10 日通过网络播客平台发布，由合伙人 Konstantine Buhler 对谈英伟达（NVIDIA）创始人兼首席执行官黄仁勋。节目围绕计算架构正在经历的六十年一遇之重大变革展开，阐释了计算范式如何从静态“预存检索”跃迁至实时“智能生成”，并深度拆解了被喻为智能时代发电机的 AI 工厂、产业投资的“五层蛋糕”模型，以及技术自动化如何提升劳动力需求而非摧毁就业的核心逻辑。 ## TL;DR 黄仁勋阐述了全球计算范式正经历从 IBM System/360 奠定的“预存检索”向实时定制化“智能生成”的历史性巨变，AI 工厂制造的智能已成为继电力和互联网之后第三层“包裹地球”的基础设施。他提出了涵盖能源、芯片与硬件、基础设施、基础模型以及场景应用的“五层蛋糕”投资模型，并预言 AI 产业将发展成每年 20 万亿美元的巨型生态。面对公众的失业恐慌，黄仁勋严词驳斥了末日叙事，以放射医学在计算机视觉普及后需求暴增的“杰文斯悖论”为例，深刻厘清了日常任务与职业本质目标的区别。他最终强调，自然语言交互正在彻底抚平成百上千万人面临的技术鸿沟，人类绝不会被 AI 淘汰，但会被掌握 AI 这一超级工具的同行淘汰。 ## Atmosphere & style 现场对谈兼具工业文明的宏大历史纵深与极具感染力的极客幽默，黄仁勋在拆解底层芯片架构与信手拈来的生活比喻之间切换自如，频频引发现场投资人与企业领袖的会心大笑与掌声。主持人的提问精准且充满崇敬，黄仁勋在回应末日论与就业焦虑时情绪尤为鲜明激昂，用笃定的工程逻辑与扎实的现实数据有力反击了抽象恐慌。整场交流洋溢着坚定而澎湃的科技乐观主义，带给听众一种亲历新工业革命地基浇筑的震撼感与拨云见日般的清晰笃定。 ## Bulletin - `05:00` 计算范式迎来60年未有之大变局：全球计算架构正从始于1964年IBM大型机的“预存检索”彻底转向“实时生成”，数据中心也正从存储调取静态文件全面演变为生产Token与原始智能的“AI工厂”。 - `16:03` AI基础设施经济模型重塑：一台集成了72颗芯片、重达2吨的NVL72超级机架售价约400万美元且供不应求；投资500亿美元建设的1GW级巨型AI工厂，预计可产出3000亿至4000亿美元的高附加值数字智能产品。 - `20:00` AI投资呈现清晰的“五层蛋糕”结构：自底向上涵盖能源、芯片与硬件、基础设施、大模型与顶层应用，今年全生态注入资本约1万亿美元，长期有望撬动年产值20万亿美元的庞大市场。 - `32:30` 自动化破除就业末日论：AI自动化的是“日常任务（Task）”而非“核心目标（Purpose）”，正如放射科与软件工程在AI深度介入后雇佣需求因杰文斯悖论反而创下新高，人类不会被AI淘汰，但会被善用AI的人淘汰。 ## Key takeaways - `00:00` 从生成到思考与执行：两年前的AI主要局限于跨模态内容生成与翻译，而当下的质变在于内部词元推理与工具调用（Agentic Systems），使AI演进为能交付工作成果的生产力工具。 - `05:00` 计算范式的世纪跃迁：过去60年的计算本质是静态文件的写入与调取，如今全面转向根据用户情境与提示词实时推演生成的智能范式。 - `10:00` 包裹地球的第三层网络：继第一次工业革命的发电机与电网、电信时代的互联网之后，由AI工厂实时驱动的“智能网络”正成为第三种覆盖全球的公共基础设施。 - `10:00` 百亿智能体协同生态：未来的互联网主要流量将不再是服务10亿人类用户，而是由上百亿个在安全护栏下昼夜运转、自主分工协同的Agent所占据。 - `13:14` 智能时代的“发电机”：昔日发电机将物理世界的“原子”动能转化为“电子”，今天的AI工厂则将“电子”在芯片矩阵中重组，输出为表征语言、物理法则与生物机理的“Token”。 - `16:03` 巨型AI工厂的高ROI经济账：虽然单机架造价高达400万美元且1GW级AI工厂基建投入耗资500亿美元，但其可产生高达数千亿美元的数字智能产值，投资回收期极短。 - `20:00` AI投资的“五层蛋糕”架构：产业投资由底部的能源（核能、风光、电网等装备巨头率先爆发）、芯片与系统、基础设施（土地、电力、机房全线紧缺）、模型层及顶层应用构成。 - `22:15` 全要素Token化与80万亿物理世界：AI的本质是理解具结构之物，文本、图像与细胞、基因、车辆在模型眼中皆为Token，实体工业才是AI最广阔的蓝海。 - `25:00` 千亿美元风投资本涌向应用层：海量初创公司正在垂直场景落地AI能力，推动全行业向20万亿美元的智能生产生态演进。 - `27:30` 驳斥“奇点”与“20%毁灭论”等末日恐慌：AI是科学家与工程师透彻理解且每年系统性改进的软硬件产物，安全性伴随技术进步正被坚实构筑。 - `30:00` 职场竞争新法则：人类不会被AI本身取代，但不用AI的人必然会被使用AI的人降维打击，拥抱工具是下一代必备的超级力量。 - `30:00` “工作本质（Purpose）”不等于“具体任务（Task）”：CEO的核心是领导而非打字，将机械性琐碎任务剥离给AI，反而能让人类聚焦于高价值的问题求解与前瞻思考。 - `32:30` 放射科医生悖论与杰文斯效应：十余年前曾被断言将被AI消灭的放射科医生，因读片效率提升大幅拉动了医院检测通量与盈利，行业雇佣人数不降反升，软件工程师亦然。 - `36:30` 职业维度的普遍跃升：AI将重新定义蓝领与白领工种，把水管工、木匠或销售人员从被动执行者升格为具备全套方案输出能力的空间设计师与咨询师。 - `36:30` 自然语言彻底抹平技术鸿沟：40年来编程语言的复杂化导致懂代码的群体始终微小，而大模型让人类语言直接成为编程工具，真正实现了计算红利的全球普惠。 ## Detailed walkthrough ### `00:00` 从生成到思考：AI 如何演变为能够“做工”的 Agent 对谈伊始，红杉资本的 Konstantine Buhler 将当前的 AI 变革定义为人类历史上规模最大、速度最快的技术革命，甚至超越了当年的工业革命。面对这场核心围绕“AI 工厂”（AI Factory）展开的基础设施军备竞赛，英伟达（NVIDIA）创始人兼首席执行官黄仁勋（Jensen Huang）首先从第一性原理拆解了大众对 AI 的理解。大多数人最初接触 AI 是通过浏览器里的聊天机器人（Chatbot），在交互中输入提示词（Prompt），系统给出回应。在短短两三年间，以两年前 ChatGPT 的诞生为分水岭，AI 经历了翻天覆地的演化：彼时的系统核心能力是感知、理解与翻译，能够跨模态地在任意形式间生成内容——将 PDF 摘要为文本（Text-to-Text）、将文本描摹成图像（Text-to-Image），或是理解图像内容生成解说（Image-to-Text），即所谓的“anything to anything else”（跨模态任意转换）。 然而，生成式 AI（Generative AI）真正的质变在于“思考”（Thinking）与“推理”（Reasoning）。黄仁勋指出，如果一个系统无法在内部生成词元（Tokens），它就无法进行逐步推理和问题求解。由此衍生出的另一项关键飞跃，是生成指令去调用工具（Tool Use）的能力——现阶段是驱动浏览器、电子表格、Photoshop、PowerPoint 或 AutoCAD 等数字工具，未来则是操控机械系统（即机器人学，Robotics）或操控带有方向盘的机器（即自动驾驶汽车，Self-driving cars）。两年前，大众可能觉得 ChatGPT 只是个写打油诗、编乡村歌曲或产生胡说八道幻觉的新奇玩意儿，但正是这套底层技术，在两年后催生出了能真正干活的智能体系统（Agentic Systems）。社会从不为单纯“无所不知”的万事通朋友买单，但会毫不犹豫地为解决问题、输出劳动成果付费。正因具备了推理与执行工作（Reason and do work）的能力，AI 一跃成为人类历史上增长最快的软件商业模式。 ```mermaid flowchart LR accTitle: AI 能力演进路径 accDescr: 展示从早期的模态翻译到逐步推理，再到工具调用与最终具身智能的演进过程。 subgraph Phase1["基础生成阶段"] A["感知与模态翻译<br/>(Text / Image / Code)"] end subgraph Phase2["认知推理阶段"] B["内部思考与逐步推理<br/>(Internal Reasoning)"] end subgraph Phase3["智能体执行阶段"] C["数字工具调用<br/>(Browser / Office / CAD)"] D["具身与物理控制<br/>(Robotics / 自动驾驶)"] end A --> B B --> C C --> D ``` - **两年前的生成能力**：主要局限于跨模态内容生成与翻译，被视作新奇有趣的创作玩具。 - **当下的智能体形态**：涵盖 Codex、Claude Code 等编程与工作流 Agent，具备端到端完成复杂任务的实用价值。 - **商业模式转变**：人类开始以“时薪制”购买 AI 服务（例如每小时 20 至 30 美元），推动 AI 成为人类商业史上扩张速度最迅猛的软件业务。 --- ### `05:00` 计算架构的甲子巨变：从“预存检索”到“实时生成” 当多数人将目光聚焦于医疗健康、金融、生命科学、制造、物流与泛娱乐等下游产业（Downstream）的应用前景时，黄仁勋强调更应从上游（Upstream）的工业本质切入思考。现代通用计算架构的建立可以追溯到整整 64 年前——1964 年 IBM 宣布推出 System/360 大型机，奠定了至今沿用 40 多年的现代计算机体系。这套延续了半个多世纪的计算模式本质上是**基于预存的检索系统（Retrieval-based computing）**：无论是文本、手写程序、照片、录音还是流媒体视频，人们都是先将其写入并保存为静态文件；当需要消费时，再从磁盘中提取出来。即便引入了推荐算法，其实质也只是根据用户画像进行“稍微聪明一点的调取”。正因如此，容纳这些设备的建筑被称作**数据中心（Data Center）**，因为它们主要负责存储数据，真正发生的实时计算微乎其微。 而生成式 AI 彻底颠覆了这套底层逻辑。在当前的范式下，用户每一次交互都会提供全新的上下文（Context）与提示词（Query），计算芯片必须依据当前的特殊情境，先理解语义、执行推理，再实时生成独一无二的输出。正如黄仁勋在演讲现场的类比：台下的听众来自 60 多个不同国家、上百个家庭，拥有截然不同的行业背景，他所说出的每一句话都不是事先录制好的现成磁带，而是根据现场听众的反应实时组织、当场生成的智能体现。计算机行业经历了 60 年的检索主导期，如今正不可逆转地转向纯粹的**实时生成（Generation）**。黄仁勋直言，这种理解认知、具备情境感知的生成过程就是“智能”本身；理解这一底层巨变，正是投资者判断未来价值数千亿美元超级产业的核心直觉来源。 | 维度 | 传统计算时代（1964年 IBM System/360 确立） | 生成式 AI 时代（当前范式转变） | |:--- |:--- |:--- | | **底层运作核心** | **预先记录与检索（Retrieval）** | **情境感知与实时生成（Generation）** | | **数据形态** | 静态文件（写入硬盘、定期调取、依靠推荐系统分发） | 动态产物（依据上下文 Context 与查询 Query 现场推理） | | **物理设施定位** | **数据中心（Data Center）**：重在低算力的数据存储与文件提取 | **AI 工厂（AI Factory）**：重在高强度实时计算，输出原始智能 | | **边际产出** | 确定性的录制数据复制（如流媒体、存储文件） | 独一无二、高度定制化的实时词元（Tokens）与逻辑推理 | ### `10:00` 从“检索”走向“生成”：百亿智能体网络与包裹地球的第三层网络 计算机的底层逻辑已经发生了六十年来最彻底的颠覆。过去的计算核心是“检索”（Retrieval），即数据中心存储好现成的内容，用户发起请求时再去调取匹配；而今天的计算已经全面进化为“生成”（Generation）。在不远的未来，人们所看到的每一个像素、每一个单词、每一条新闻、每一则广告甚至每一段视频，都将不再是从某个数据库中提取出来的静态文件，而是根据请求者的身份、背景、动机和提问方式，在毫秒之间实时独立生成的。“康斯坦丁看到的，和我看到的，以及你看到的，将会完全不同。”这种从“检索”到“生成”的范式转移，意味着全球对“生成器”（Generators）的需求将迎来指数级爆发，而制造这些负责生成智能的巨型计算机，正是 NVIDIA 的立身之本。 随着 AI 跨越到智能体（Agentic AI）阶段，计算的规模边界被彻底打破。智能体具备自主执行复杂任务的能力，这意味着智能体之间能够彼此握手、分工协同：“我手头有些工作要做，我们组队把它搞定。”在 NVIDIA 内部，已经有数十万个智能体在沙盒与安全护栏（Sandboxed and Guardrailed）的监管下昼夜不停地自主沟通、协同解决工程难题。目前的互联网主要是为全球约 10 亿人类用户提供服务，但在未来的图景中，互联网的绝大部分流量与交互将由数百亿甚至上千亿个全天候运转的智能体占据——公司之间的业务对接由智能体完成，员工拥有各自的智能体助手，自动驾驶汽车、工业制造系统、具身智能机器人以及每一栋智能建筑本身都是具备自主决策能力的智能体。它们通过网络相互发出生成的指令、理解生成的意图，形成一层时刻在生成智能并紧密“包裹地球”（Cocoon the earth）的庞大计算层。 这种看似不可思议的全球性重构，在人类历史上其实已经发生过两次： - **第一次包裹：能源电网**。几百年前，德国西门子公司打造出了划时代的机器——发电机。只要点燃火焰、驱动蒸汽或借助流水，机器另一端就会涌现出一股当时无人能理解的无形力量，这便是电力。如今，遍布全球的发电机与输电线路交织成电网，如同巨茧般包裹了整个地球。 - **第二次包裹：通信网络**。三十多年前，发源于美国的网络互联矩阵织就了互联网，成为第二层笼罩全球的基础设施。 - **第三次包裹：智能网络**。继能源和通信之后，由 AI 工厂实时驱动的“智能”正在成为第三种包裹地球的全球性大宗商品（Commodity），渗透至人类社会的每一个角落。 ### `13:14` 智能时代的“发电机”：从原子转化为电子，再从电子凝练为 Token 三百多年前人类发明的发电机（Dynamo），其本质是物理运动的转换器：将水流、风力、煤炭燃烧或蒸汽等包含质量与动能的“原子”（Atoms）输入其中，转化并输出为无形的“电子”（Electrons）。而相隔数百年后，NVIDIA 所打造的机器构成了这一进程的下半段接力：将这股由电网输送而来的“电子”送入现代 AI 工厂，在庞大的芯片矩阵中经过运算重组，最终输出为纯粹的“数字”。 这些被输出的数字（即通常所说的 Token）虽然底层看似简单，但通过不同的排列组合与模型架构，能够被重塑为人类所能理解与应用的一切形式的智能。它不仅能转化为人类的自然语言和严谨的数学推导，还能掌握生物世界的底层语言（如蛋白质折叠与基因密码）、物理世界的法则（如气候演变、天气预测与流体力学），以及三维现实空间的物理交互规律（如自动驾驶决策与具身机器人控制）。从物理动能到电流，再从电流到 Token，这便是现代 AI 工厂运行的核心逻辑。 ```mermaid flowchart LR accTitle: 两次工业革命的核心机器对比 accDescr: 展示从物理运动转化为电力，再由 AI 工厂转化为智能 Token 的全流程机制。 subgraph SG1["第一次革命：传统工业"] A["物理世界（原子）<br/>水流 / 风力 / 蒸汽"] -->|"动能输入"| B["发电机 (Dynamo)<br/>西门子工业时代"] B -->|"能量输出"| C["电子 (Electrons)<br/>全球统一电网"] end subgraph SG2["第三次革命：智能时代"] C -->|"电力驱动"| D["AI 工厂 (NVIDIA)<br/>GPU 集群超级计算机"] D -->|"推理生成"| E["数字 / Token<br/>语言、生物学、物理与三维智能"] end ``` 黄仁勋借此幽默地调侃道：“这就讲完了！现在你们明白 AI 是干什么的、它是怎么造出来的、以及未来市场规模有多大了吧？谢谢各位今天到场！”引得全场大笑。主持人康斯坦丁·布勒（Konstantine Buhler）随即接话赞叹，指出这三大基础设施变革在底层逻辑上一脉相承：能源革命的核心是发电机，电信革命的枢纽是网络交换机，而在当前的智能革命中，核心正是以 GPU 和整套 AI 工厂系统为代表的超级计算设备。 ### `16:03` 两吨重的超级机架与 AI 工厂的惊人投资回报率 康斯坦丁提到，在智能革命中，NVIDIA 将所有关键部件高度集成在同一架构之下，从 H100 到最新的 Vera Rubin 架构系统，AI 工厂的工程集成度达到了前所未有的高度。黄仁勋随即披露了当前顶级系统的制造细节： - **单机架集成度**：以最新一代 NVL72 为代表的系统机架，内部集成了 72 颗高性能芯片，包含高达 150 万个精密零部件。 - **物理规格与造价**：单个机架整机重量高达 2 吨，单套售价约为 400 万美元，黄仁勋戏称这是“全世界单体最昂贵的工业设备”，康斯坦丁则打趣说黄仁勋平时肯定就是靠举这些机架来健身的。 - **出货规模**：尽管结构极度复杂，NVIDIA 却像量产智能手机一样开足马力大规模制造这些设备，今年芯片出货量预计达到约 800 万颗，并以整机架的形式源源不断运往全球各地的数据中心。面对如此供不应求的庞然大物，黄仁勋笑着确认：“没有任何批量折扣（No volume discounts）。” 谈及建设这些巨型基础设施的经济账，黄仁勋打破了传统对于重资产投资周期的认知。建设一座功耗达到 1 吉瓦（GW）级别的巨型 AI 工厂，硬件与基建的总投入高达约 500 亿美元，堪称人类工业史上造价最昂贵的单一工厂。然而与传统工厂动辄数十年收回成本不同，这一座 500 亿美元的 AI 工厂能够产生价值 3000 亿到 4000 亿美元的数字智能产品。其极高的生产附加值，带来了极速的投资回报率（ROI）。 ### `19:08` 拆解 AI 工业图景：五层蛋糕架构与能源基底 面对现场来自制造业、能源行业及各大跨国企业的领袖与投资人，康斯坦丁抛出了最关键的问题：身处这场方兴未艾的智能革命正中心，企业与个人究竟该如何看清全貌并参与其中？ 黄仁勋为此提出了一个清晰的“五层蛋糕”（Five-layer cake）产业分析心智模型，用以系统性指导投资决策与商业布局： - 任何智能的诞生都无法脱离物理世界规律，位于蛋糕最底层的第 1 层正是**能源（Energy）**。 - 从历史上最早的发电机，演进至如今各类超大规模的交流发电机与绿色电力网络，充足且廉价的电力供给是支撑一切算力扩张与 Token 生产的不可替代之根基。 黄仁勋强调，理解这个五层架构的层级关系，是把握 AI 时代产业资本流向、基础设施建设节奏以及上层应用价值沉淀的关键起点。 ### `20:00` 人工智能投资的“五层蛋糕”模型 黄仁勋（Jensen Huang）将当前整个人工智能产业的投资架构比作一个“五层蛋糕”（five-layer cake），并指出仅在今年，来自全球市场的资本就将向这套体系注入整整一万亿美元。最底部的第一层是**能源（Energy）**。这是数代人以来能源产业迎来的一场最大规模增长机遇，也是过去近百年来对多国能源电网进行升级的最佳窗口。无论人们看好的是核能、风能、太阳能还是氢能，“只要能够产生能源，就会得到资金支持”。正因如此，西门子（Siemens）、三菱（Mitsubishi）以及通用电气剥离出来的 GE Vernova 等能源装备巨头如今都迎来了爆发式增长。紧随其后的第二层是**芯片与计算系统（Chips and Computers）**，涵盖了芯片、服务器、计算集群、网络交换机以及硅光子技术（Silicon Photonics）等底层算力硬件。 第三层被称为**基础设施层（Infrastructure）**，包括土地、电力供应、数据中心外壳建筑、资金支持以及数据中心的实际运营。目前在这一层级中，几乎每一个环节的资源都处于极端稀缺的状态。第四层则是公众最为熟知、普遍直接等同于 AI 本身的**模型层（Model Layer）**，它构筑在底层硬件与云基础设施之上。黄仁勋感慨，这是近期人类历史上最为波澜壮阔的机遇，如此庞大的市场驱动型资本正在自发汇入这套生态系统之中。 ```mermaid flowchart TD accTitle: AI产业投资五层蛋糕模型 accDescr: 展示人工智能产业投资从底层的能源一直向上到应用层的五层架构。 subgraph HardwareAndInfra["算力与基础设施底座"] L1["第一层：能源 (Energy)<br/>核能 / 风光 / 电网升级"] L2["第二层：芯片与硬件 (Chips & Systems)<br/>GPU / 交换机 / 硅光子"] L3["第三层：基础设施 (Infrastructure)<br/>土地 / 电力接入 / 机房运营"] end subgraph IntelligenceAndApps["智能生成与应用交付"] L4["第四层：基础模型 (Models)<br/>语言 / 物理结构 / 生物智能"] L5["第五层：场景应用 (Applications)<br/>金融 / 法律 / 会计 / 物流"] end L1 --> L2 L2 --> L3 L3 --> L4 L4 --> L5 ``` - **第一层（能源）代表企业**：西门子（Siemens）、三菱（Mitsubishi）、GE Vernova 等。 - **第二层（硬件设备）**：AI 芯片、计算系统、网络设备、网络交换机、硅光子（Silicon Photonics）。 - **第三层（基础设施与要素）**：土地供应、可用电力指标、数据中心机房（Shell）、资本与数据中心现场运营能力，当前处于全线稀缺状态。 - **资本体量**：今年单年注入该生态体系的市场资金约为 1 万亿美元。 ### `22:15` 跨越语言模型：物理世界与结构的标记化 在模型层中，大多数人所熟知的主要是 OpenAI 和 Anthropic 等少数研发语言模型的领军企业。然而黄仁勋强调，绝不能仅将目光局限于语言。正如他此前所阐述的，人工智能的本质在于学习“具有结构之物”的语言与深层含义。他幽默地现场举例：如果他走进演讲厅时看到有人倒挂在天花板上、有人悬浮在半空、还有人的身体器官分散在 17 个不同的地方甚至肉眼可以看穿，那计算机将极难学习，因为那违背结构且不可预测（如同难以捉摸的量子现象）；但现实中人们端正就座、长着眼睛，物理规律一贯成立，他坐上椅子时可以百分之百确信自己不会有 47% 的概率从椅子中穿透摔在地上。 只要事物具备可预测性和规律结构，模型就能从中提炼出意义。计算机并不在乎输入的内容究竟是一段词句、一张图像，还是一个蛋白质、一段基因、一辆汽车或一个活体细胞。黄仁勋以生物学为例：AI 正在理解基因和细胞的深层含义，两张词汇拼在一起会互相激活产生新涵义，两个细胞结合在一起同样如此。从计算科学的角度看，它们全都是“Token（标记）”。计算机科学家的核心任务，就是将物理世界的万物信息转化为计算机能够理解、推理、规划并生成行动的表征形式。相较于公众讨论热烈的纯语言交互，背后对应着高达 80 万亿美元规模的物理世界实体产业，才是 AI 真正最具价值的前沿阵地。 - **结构的本质**：只要现实具备可预测性与物理结构，AI 便能学习其内在规律与含义。 - **全要素 Token 化**：文本、图像、汽车、蛋白质、基因片段、细胞组织，在模型眼中本质上皆可被结构化表征为 Token。 - **物理世界的产业规模**：涵盖物理实体的非纯语言产业体量约为 80 万亿美元。 ### `25:00` 20 万亿未来市场与风投涌向的应用层 在模型层之上，是红杉资本合伙人康斯坦丁·布勒（Konstantine Buhler）每天所接触到的全新浪潮——**第五层：应用层（Applications）**。在这一层，海量初创公司正在将基础模型的智能能力深度植入金融服务、法律科技、财税审计、交通运输以及供应链物流等各行各业，全面提升人类生产生活条件。黄仁勋指出，去年全球风险投资向这一领域注资约 1000 亿美元，创下了人类有史以来单一年度风险投资的历史最高纪录。 黄仁勋坚信这套 AI 产业未来的规模将不可限量。两年前整个产业规模几乎为零，如今已有约 1 万亿美元的市场投入；但他预计，整个 AI 产业最终将迈向每年约 20 万亿美元的巨型生态。因为人类必须思考一个根本问题：智能究竟有多重要？谁需要它？需要多少？无论是用于生成蛋白质分子、驾驶汽车、控制机器人，还是处理人类语言和科研数学，所有智能未来都必须由这些被称为“AI 工厂”的机器不间断地生成。布勒在对话中附和道，仅按每个吉瓦（Gigawatt）规模的电力设施耗资约 500 亿美元、未来数年将有上百个此类项目上线来测算，底层硬件和基础设施投资就将达数万亿美元，再加上应用层带来的衍生价值，将为实体建设者与从业人员创造海量真实的就业岗位。 - **应用层风险投资**：过去一年约 1000 亿美元风险资本涌入应用层，刷新人类风险投资年度历史纪录。 - **应用落地垂直赛道**：金融服务（Financial Services）、法律（Legal）、会计（Accounting）、交通运输与物流（Transportation & Logistics）。 - **生态远景规模**：从两年前的近乎空白，到当前的 1 万亿美元投入，未来有望演变为年产值高达 20 万亿美元的智能生产产业。 - **基础设施建设规模参考**：单个吉瓦（GW）级电力设施对应约 500 亿美元投资，未来数年预计将有 100 多个同等规模单元陆续上线。 ### `27:30` 破除末日叙事：拥抱工具而非被使用者淘汰 面对当前不同国家由于文化差异而对 AI 产生的不同反应，黄仁勋直言不讳地告诫大众：必须警惕那些脱离现实的科幻类比和末日叙事。诸如将 AI 描摹为“终结者（Terminator）”、炒作“技术奇点（Singularity）”或者声称“人类文明有 20% 几率因 AI 灭亡”的论调，完全是“一派胡言”（complete nonsense）。同样荒谬的还有所谓“AI 充满神秘黑盒、开发者自己都完全不知道它为何能运转、明天它可能就会自己长腿跑掉”的煽动性言辞。黄仁勋反驳道，AI 的本质毫无疑问就是计算机硬件与软件，科学家和工程师不仅清楚其运作机理，而且“每年都在实实在在地让它变得更强大——如果你根本不知道它怎么工作，你又怎么可能每年都改进它？” 黄仁勋呼吁人们不要被恐惧吓倒，而要积极主动地拥抱这项技术：“你也许不会被 AI 淘汰，但你绝对会被使用 AI 的人淘汰”（"You may or may not lose a job to an AI, but you will absolutely lose a job to someone who uses AI."）。面对无法确定的末日恐慌，人们应该专注于确凿无疑的事实。黄仁勋甚至以此来告诫身为人父母的听众，绝不该指导下一代在 AI 浪潮前逃跑，而要确保他们掌握这一赋予常人“超能力”的工具。与此同时，工程界也倾注了巨大的心血来保障该技术的安全性。黄仁勋以产品的迭代速度为例：两年前的 ChatGPT 与今天的最新系统相比，模型产生幻觉的概率已经断崖式下降；现在的 AI 不仅能够输出切合当下语境的精准知识，在面对未知问题时还会自主进行研究检索，甚至在输出答案之前进行多维度的自我质疑与校验。 - **反驳 AI 末日论**：明确驳斥“20% 人类毁灭概率”、“终结者觉醒”、“工程师不知道系统如何运转”等恐慌叙事。 - **核心职场法则**：决定胜负的关键不是人类与机器对抗，而是使用 AI 的人对不用 AI 的人产生降维打击。 - **安全性与模型表现进化**：对比两年前的 ChatGPT（原音频口误提及其代际），最新模型在事实幻觉率骤降、上下文连贯度、自主调研能力及自我反思纠错机制上实现了质的跃迁。 ### `30:00` AI 安全性、拥抱超级力量与“工作岗位”辩论 黄仁勋（Jensen Huang）深入探讨了现代 AI 的思考机制与安全性演进。当前的 AI 模型在给出最终回答前，会经历自我反思（reflection），生成两到三个不同的候选答案并进行权衡校准。行业在安全护栏（guard railing）和事实锚定（grounding of truth）上的技术迭代极其迅猛。黄仁勋打了个生动的比方：比起 100 年前的老式汽车，他绝对更偏爱今天的现代汽车——尽管现代汽车内部集成了成千上万项复杂的技术发明，但它却无比安全。这说明安全本身需要依托大量先进技术来实现。因此，科技产业、科学家和工程师的首要天职就是构建安全的 AI。 黄仁勋语气严肃地向台下观众发出呼吁：每个人都肩负着责任，必须去告诫身边的亲人、孩子、孙辈、所在的公司乃至整个国家——无论如何，一定要主动拥抱 AI。如果 AI 是一种“超级力量”（superpower），就必须全力以赴去掌握它。黄仁勋直言： > “我们不会因为 AI 而失去生命（或饭碗），但我们可能会输给那些善用 AI 的人。” 当主持人提到“工作岗位”（job）这个词时，黄仁勋的情绪立刻被点燃。他直言自己对外界针对 AI 扼杀就业的种种无端臆测深感反感（trigger）。他反问台下：“今年我们向全球生态系统注入了整整一万亿美元，这笔钱究竟在做什么？在创造就业！”从能源行业、芯片制造，到包括土地、电力、数据中心外壳及金融服务在内的基础设施层，再到 AI 模型层以及获得千亿美元注资的应用层，全产业链各环节所催生的就业岗位比以往任何时期都要多。 ```mermaid flowchart TD accTitle: AI 五层生态投资与就业催生逻辑 accDescr: 展示从能源与底层硬件到顶层应用如何全链路创造更多就业岗位 subgraph InfraLayer ["底座支撑层"] A["能源与电力板块"] --> C["数据中心与基建 (土地/电力/金融)"] B["芯片与半导体制造"] --> C end subgraph ModelAppLayer ["模型与应用层"] C --> D["AI 核心大模型研发"] D --> E["顶层行业应用生态 ($100B 注入)"] end E --> F["创造历史新高的全产业链就业机会"] ``` 对于传统岗位是否会被取代的担忧，黄仁勋提出了一个核心区分：**工作（Job/Purpose）与日常任务（Task）紧密相关，但绝不是同一回事**。他拿自己作为首席执行官打趣：作为 CEO，他的工作职责是领导英伟达，但他大部分时间具体执行的“任务”无非是“打字”和“说话”。如果按照字面任务来看，“CEO = 打字 + 说话”，而这两项任务 AI 如今都能以超人类的水平完成，然而他自己却比以往任何时候都更加忙碌。 --- ### `32:30` 放射科医生悖论与“任务（Task）与本质目标（Purpose）”之辨 为了透彻阐明工作与任务的区别，黄仁勋分享了一个极具启发性的科技历史教训。大约在十多年前，全球顶尖的计算机科学家之一（暗指杰弗里·辛顿 Geoffrey Hinton）曾公开警告世人 AI 的恐怖威力。当时那位科学家断言：第一个将被 AI 彻底消灭和摧毁的职业就是**放射科医生（radiologist）**，甚至公开劝诫所有年轻人千万不要再报考或涉足这个领域，因为该行业必将覆灭。早在 12 年前，计算机视觉（computer vision）在检测医学影像异常上就已具备超人类能力——不知疲倦、明察秋毫。 预测的前半部分确实应验了：计算机视觉以极快的速度彻底渗透进放射医学的每一个环节，如今每一位放射科医生都在使用计算机视觉辅助系统。然而，现实的走向却给所有人上了一课：**全球对放射科的需求不降反升，放射科医生的实际雇佣人数也创下新高**。 黄仁勋现场邀请观众思考其中的深层机制并给出了剖析： 1. **放射科医生的本质目标**：看片扫描只是占用时间的“任务”，其核心工作目标（purpose）是与各科临床医生紧密协作以诊断疾病。 2. **生产力释放与科室盈利**：当看片任务被自动化工具提速后，医生效率大增。医院能接收更多排队候诊的痛苦患者，放射科的检测通量和盈利能力随之暴涨。 3. **扩大招聘规模**：为了接纳更多患者、帮助更多病患并创造更多效益，各大医院反而开始大举招聘更多的放射科医生。 这一现象完美诠释了“杰文斯悖论”（Jevons paradox）：自动化降低了单次操作成本，反而激发了远大于以往的总体需求。 ```mermaid flowchart LR accTitle: 放射医学自动化的正向经济循环 accDescr: 展示计算机视觉提高诊断效率后如何反向扩大全行业用人需求 A["AI 自动化读片任务"] --> B["医生工作效率大幅提升"] B --> C["医院检测通量倍增<br/>收治更多等待病患"] C --> D["科室盈利能力显著提升"] D --> E["大举扩招放射科医生"] ``` 黄仁勋严肃指出，计算机科学家不负责任的末日预言险些酿成严重灾难——当时医学界报考放射科的人数确实一度因恐慌而下滑，但临床前线其实极其缺乏人手。同理，当下有人宣称“90% 的软件代码将由 AI 完成，因此不再需要软件工程师”，但英伟达却在以创纪录的速度招募软件工程师。黄仁勋直言： > “我招聘软件工程师时，绝不会递给他一个键盘说：‘来，展示一下你一秒钟能打多少个单词。’打字和敲代码不是他们的工作，**解决问题和提出前瞻性问题才是他们的工作**。” 在工业发展史上，工作站出现前就有放射科医生，AI 时代后放射科医生依然会存在；在软件编程出现前就有工程师，在代码被高度自动化后，工程师也同样会屹立不倒。 --- ### `36:30` 职业维度的跃升与彻底消除“技术鸿沟” 黄仁勋进一步总结了对工作形态演变的看法：AI 不会消灭工作，而是会**重塑并拔高（elevate）各行各业的职业维度**。 - **水管工（Plumber）**：今天的水管工拿到的是具体的施工任务单或管道示意图；明天借助 AI，水管工可以一键生成令人惊叹的现代化厨房设计图，一跃兼任空间设计师。 - **家具销售员（Furniture Salesperson）**：不再只是被动卖桌椅的人，而是能即时根据客户房屋户型提供高阶全屋软装方案的室内设计师。 - **木匠（Carpenter）**：不再只是机械地把木头拼凑在一起，而是直接升级为全套家居空间定制的设计师。 对于社会舆论中蔓延的恐慌情绪，黄仁勋尖锐地指出，许多唱衰 AI 的叙事本质上是通过制造恐慌来牟利，阻碍公众走向技术红利。 更重要的是，在黄仁勋长达 40 余年的计算机设计生涯中，AI 是他所见过的**彻底消除技术鸿沟（technology divide）的最伟大力量**。回顾过去 40 年，底层计算系统被设计得越来越庞大、精密且极其晦涩，导致全社会中真正懂得编写程序操纵计算机的人口比例持续缩减。黄仁勋现场半开玩笑地询问台下：“在座有谁真正懂 C++？”台下只有零星两排初创团队成员举手。黄仁勋打趣道：“在这群极其特殊的精英听众里也就 2% 左右懂 C++，而在普通大众里这个比例甚至远低于 2%。但是，有百分之多少的人懂人类语言（Human Language）？所有人！” 由于大语言模型可以直接听懂人类自然语言，如今全人类都获得了给计算机“编程”的能力。以往横亘在技术精英与普通公众之间的技术深渊被彻底抹平，科技红利第一次真正走向了普罗大众。 --- ### `39:50` 总结陈词：从检索到生成的智能纪元 红杉资本主持人 Konstantine Buhler 登台进行总结，高度评价了黄仁勋的乐观主义视野。Buhler 指出，黄仁勋是当今全球最接近底层核心计算技术脉动的人，他的观点为整个技术产业锚定了方向： - **计算范式的世纪跃迁**：全球计算范式正在彻底摆脱陪伴了几代人的“信息检索”（retrieval）模式，迈向实时、定制化的“智能生成”（generation）时代。在这个新纪元里，无论是文字、知识还是合成视频，都将根据使用者的需求即时生成。 - **智能作为驱动文明的新力量**：正如昔日发电机推动了电气化革命、光纤推动了电信与互联网革命，当下的生成式 AI 工厂正在让智能成为环绕全球的第三股澎湃力量。 - **计算语言的无限延展**：计算机如今不仅能理解英语、德语等自然语言，甚至掌握了生命底层的“蛋白质语言”，正在重塑生物学与医疗。 - **技术赋能人类使命**：正如黄仁勋所展现的图景，AI 将把人类从枯燥重复的打字、读片或机械劳作等任务碎片中彻底解放出来，赋能每个人去思考真正的宏大命题，让更多匠人蜕变成为文明的“架构师”。 在全场热烈的掌声与欢呼声中，主持人向一手打造了智能时代基础设施的英伟达掌舵人 Jensen Huang 致以由衷的敬意，全场对谈圆满落幕。 ## Notable quotes > `00:00` anything to anything else — Jensen Huang > `10:00` 康斯坦丁看到的，和我看到的，以及你看到的，将会完全不同。 — Jensen Huang > `13:14` 这就讲完了！现在你们明白 AI 是干什么的、它是怎么造出来的、以及未来市场规模有多大了吧？谢谢各位今天到场！ — Jensen Huang > `16:03` 没有任何批量折扣（No volume discounts）。 — Jensen Huang > `20:00` 只要能够产生能源，就会得到资金支持 — Jensen Huang > `27:30` You may or may not lose a job to an AI, but you will absolutely lose a job to someone who uses AI. — Jensen Huang > `30:00` 我们不会因为 AI 而失去生命（或饭碗），但我们可能会输给那些善用 AI 的人。 — Jensen Huang > `30:00` 今年我们向全球生态系统注入了整整一万亿美元，这笔钱究竟在做什么？在创造就业！ — Jensen Huang > `32:30` 我招聘软件工程师时，绝不会递给他一个键盘说：‘来，展示一下你一秒钟能打多少个单词。’打字和敲代码不是他们的工作，解决问题和提出前瞻性问题才是他们的工作。 — Jensen Huang > `36:30` 在这群极其特殊的精英听众里也就 2% 左右懂 C++，而在普通大众里这个比例甚至远低于 2%。但是，有百分之多少的人懂人类语言（Human Language）？所有人！ — Jensen Huang ## Mentioned Jensen Huang (黄仁勋), Konstantine Buhler, Sequoia Capital (红杉资本), NVIDIA (英伟达), ChatGPT, OpenAI, Anthropic, Codex, Claude Code, IBM System/360, Siemens (西门子), Mitsubishi (三菱), GE Vernova, Geoffrey Hinton (杰弗里·辛顿), NVL72, H100, Vera Rubin ## Video description Jensen Huang, founder and CEO of NVIDIA, makes the case that computing is undergoing its biggest shift in 60 years: from retrieval, where data centers store files we look up, to generation, where every word, image, and video is produced in real time and customized for whoever is asking. He explains why NVIDIA's AI factories are the dynamos of this era: machines that take in electrons and send out tokens of intelligence, just as Siemens' dynamo once turned motion into electricity. Jensen frames intelligence as the third force to "cocoon" the planet after electricity and the internet. He describes the five-layer cake of AI investment—energy, chips, infrastructure, models, applications—and dismantles the fear that AI will erase jobs, using radiology and software engineering to show how automation raised labor demand instead of killing it. His bottom line: you won't lose your job to AI, but you might lose it to someone who uses AI. Hosted by Konstantine Buhler, Sequoia Capital [Source](https://pscrb.fm/rss/p/traffic.megaphone.fm/CPUAI2705410687.mp3)

[0:00] Jensen Huang: anything to anything else [10:00] Jensen Huang: 康斯坦丁看到的，和我看到的，以及你看到的，将会完全不同。 [13:14] Jensen Huang: 这就讲完了！现在你们明白 AI 是干什么的、它是怎么造出来的、以及未来市场规模有多大了吧？谢谢各位今天到场！ [16:03] Jensen Huang: 没有任何批量折扣（No volume discounts）。 [20:00] Jensen Huang: 只要能够产生能源，就会得到资金支持 [27:30] Jensen Huang: You may or may not lose a job to an AI, but you will absolutely lose a job to someone who uses AI. [30:00] Jensen Huang: 我们不会因为 AI 而失去生命（或饭碗），但我们可能会输给那些善用 AI 的人。 [30:00] Jensen Huang: 今年我们向全球生态系统注入了整整一万亿美元，这笔钱究竟在做什么？在创造就业！ [32:30] Jensen Huang: 我招聘软件工程师时，绝不会递给他一个键盘说：‘来，展示一下你一秒钟能打多少个单词。’打字和敲代码不是他们的工作，解决问题和提出前瞻性问题才是他们的工作。 [36:30] Jensen Huang: 在这群极其特殊的精英听众里也就 2% 左右懂 C++，而在普通大众里这个比例甚至远低于 2%。但是，有百分之多少的人懂人类语言（Human Language）？所有人！

[25:00] Jensen Huang: You may or may not lose a job to an AI, but you will absolutely lose a job to someone who uses AI.

[0:00] Thank you so much, Jensen. So we are in the middle of a massive AI revolution. It is probably bigger and faster than even the Industrial Revolution. And you have called out what's happening right now as the largest infrastructure build out in human history. At the center of that build out is the AI factory. And the company enabling all of that is NVIDIA. Can you tell us what is an AI factory and why is it the best investment for any enterprise in the next decade? Okay, so you could understand AI in a particular number of ways. The way that you understand AI probably most is through a chatbot, through a web browser. You're interacting with it. You give it a prompt. It says something back to you. And even those of you who have been using AI for some time, you've seen in the last couple, two, three years, a very significant evolution improvement in the capabilities of AI Two years ago you heard about ChatGPT And ChatGPT basically is a computer software that understands the input you give it It can perceive understand information, and it can translate and generate the information into something else. Okay, so you can give it a prompt and you can say, here's this PDF I gave you. I would like you now to summarize it. It went from text to text. You could also tell it, here's a PDF I gave you. I would like you to now generate an image of that story. It goes text to image. You could use go from image to text, meaning you could give it a picture and you can, what's happening inside this picture. It goes text, image to text. Does that make sense? Anything to anything else. And AI in the last, in two years time two years ago was largely able to do this translation we call it generation generative models okay generative ai but the the thing that is very big deal inside that word generative ai is in order to do something even more valuable than generation understanding and generating is thinking Well you can think if you don generate words And so the foundation of generative AI gave us the ability to generate internal thoughts thinking, reasoning, step-by-step reasoning, problem solving. It also allowed us to do another thing that is now very important, which is generate intelligence to control something else, to generate control to use a tool. Does it make sense? To use a browser, use a spreadsheet, use Photoshop, use PowerPoint, use something, use AutoCAD, use another tool. Now, that tool today is digital, but someday that tool will be mechanical. So if I generate a command to a mechanical system, that would be called robotics. If I generate commands for a machine with steering wheel that would be called self-driving cars. Does that make sense? Okay. And so two years ago, two years ago, in fact, you saw the foundations, we call it ChatGPT, and everybody said, ah, you know, it's fun, it's silly, or it produced a whole bunch of crazy hallucinated text. That's all true, but it was the foundational technology that led to all of this. two years later we now have agentic systems now that one view of AI I just described the view which is what can AI do Right? And so now all of you realize you see it from ChatGPT, you see it from Codex, you see it from Cloud Code, you see that it's now able to not just understand, but it's able to do work. Reason and do work. Now, two years ago, when AI was able to understand you and generate information that was interesting, novel, a little cute, whenever you need a poem written, great way to do it, right? Who doesn't want to write a country song? And so that was two years ago. But now, because it's able to do work, AI is valuable. Valuable meaning it can generate information, it can generate useful work, and it could be paid for. Because we're interested in having friends that are smart, we love people who are know-it-alls, but we don't pay them for it. We pay for people who do work. Does that make sense? All right, which is what happened in the last two years. AI went from having this capability to now, Agentech went from not very valuable to [5:00] now producing useful work. So much useful work that you and I are doing this every day. We're paying AI by the hour, right? And so we might pay them $30 an hour to do the work, $20 an hour to do the work. We're basically paying AI a lot of money today, the fastest growing software business in the history of mankind, because now it's doing useful work and we can pay them to do it. Now that's one view of AI, which is what it can do. But one other view of AI that's really important to help reason through what Constantine's saying. So for example, the reason why some companies, some people are able to build great businesses and could maneuver themselves into the center of very large industries is because when they see this capability, this is very interesting. One interesting thought is, if we're able to do this, what is the implication to this downstream industries? That's an interesting conversation we should have. Okay. So now that AI can do this, what happens to all the industries like healthcare and financial services and life sciences, manufacturing, logistics, transportation on and on and on E advertising future entertainment the list of conversations you can have about now that AI can do this what can it do as a result subsequently That an interesting conversation But you should go upstream meaning industrially what does that mean And so the first thing you realize is this. Go back to first principles. I had told you just now that AI is software and it's being produced by a computer. Now, what happened to the computer that has made it possible to do this? Well, the big idea is about, if you think about the computer as we know it today, really emerged about 64 years ago. IBM System 360 was the biggest announcement of computing. And 64 years ago, IBM was the most valuable company in the world. And they created the modern understanding of computers. Everything that we can describe about a computer was really described in 1964. For 40 years, largely has remained the same. And what happened was that form of computing is called pre-recording. You write down your story, you save it to a file. You write a program by hand, you save it to a file. You take a picture, you save it to a file, you record music, you save it to a file, you make a video right now. We streaming right now but somebody going to record it you going to save it to a file And when you want to use it later you retrieve it from the disk drive Does that make sense And the retrieval process is done intelligently So that's why everybody's retrieval of a news story is a little bit different. It's called a recommender system. But basically, computers as we know it today is a retrieval-based system, which is the reason why these buildings are called data centers. They store data. notice they don't call them computer centers because not doing much computing they just store data that you retrieve based on what you touch on your phone well what happened now if you look at what i just described in order for this ai to work as i described it every time i say something to it i have to give it new information we call it context i give it a new prompt that's called a query between the context and the query, it will understand it first, reason about it, and it will produce an output based on that context in that query, based on the circumstance. Does that make sense so far? OK, give me one nod. And so if that is the case every time you use the AI the content is produced originally every single time Everything I saying to you right now is being produced in real time And it's because my explanation is based on the fact that I realize all of you come from 60 different countries, 128 different families. You all have many different backgrounds. Some of you probably came from the computer industry. Most of you probably did not. And so I'm explaining the information to you in a way that is sufficiently deep. But ultimately, my goal is this so that you know how to make your next investment. That's what I'm leading to. And so I'm going to give you enough sufficient information that you can reason about it for yourself so that when you see something in the next time you go, that's worth investing in. That's going to be a big industry. That's one hundred billion dollars right now. And it looks really big, but that's nothing compared to how big it's going to be. I'm going to give you the intuition to solve that problem. OK, so so here we are. Went from a computer industry that was largely based on retrieval for 60 years. And all of a sudden, one day, it's completely generated in real time. We call it intelligence. This this is what I'm doing right now for you. I'm demonstrating intelligence, contextual awareness. He gave me a promise. [10:00] And here comes my answer. Does it make sense? Extremely artificial intelligence. Okay. And so, and so here I, so here, here, what's going on now? I just gave you one word earlier. It's called generative AI. The computer of today has been completely reinvented. It's now generative. and every single toe every single letter you see every single word you see in your in the future every video every image every ad every tv commercial every single time you read a story a news story every one of them will be different what constantine sees what i see what you see will be completely different because it'll be generated for you because your your interest your context, who you are, for what reason you asked, how you asked is completely different. Does it make sense? And so therefore, therefore, every single pixel that you see, every single sound that you hear in the future, every video you see in the future will be originally generated, not retrieved which means in the future we need a lot more generators And these generators is what we build That what we build for a living these are large computers and they generating intelligence now the next question is this well how big could it be how big can it be and so it turns out it turns out the amount of information the amount of generation of intelligence is we do it for about a billion people in the world today now that I told you that AI has become agentic, meaning that it can actually do work by itself. Well, if it can do work by itself, then one agent can communicate with another agent and say, I have some work to do. Let's team up together and let's do some work. And now you have all these different agents and they're all working together to solve problems inside your company, say. So inside our company, Constantine knows that we're huge users of agentic AI. We have hundreds of thousands of agents probably running around right now that are doing work and they're talking to each other and they're solving problems. They're all guardrailed, all sandboxed, all guardrailed and sandboxed, but they're all working with each other, which means in the future it is very likely that the internet that we use today for a billion people will likely mostly be several billion call it 100 billion agents working around the clock and they using the internet and talking to each other And what are they saying? So for example, there'll be companies working with companies, employees, agents working with other employees, agents. There'll be self-driving cars, which are agentic. There'll be robots, which are agentic. All the manufacturing systems, every building will be agentic. There'll be agents all over the place and they'll be using the internet and all of those commands that they generate to each other will be generated. All of the thoughts that they have to understand all generated. Does that make sense? And so basically the world is going to be this layer of computing that's going to cocoon the earth and it's going to be generating intelligence all the time. Now, I just said something that sounds ridiculous, except, in fact, it's already happened twice. So 300 years ago, a company in Germany called Siemens produced a machine. And this machine is a really interesting machine. You go up to this machine, you light it on fire. And then this incredible invisible force comes out the other end Nobody understood what it was We understand it now as electricity How many power generators are there in the world? We call it a grid. Power generation cocoons the planet. We call it the grid. And then, of course, 20-some-odd years ago, earlier than that, 35 years ago, this networking scheme, networking matrix fabric was created here in the United States, eventually became the internet. And where is it? It cocoons the world. And so now you have energy, communications, intelligence, and it will cocoon the world. And we'll use it for, you know, it'll just be a commodity and we'll use it all over the place. And so what NVIDIA does for a living is this new machine. The machine that was invented 300 years ago is called the Dynamo. that dynamo anything that moves comes in it could be waterfalls, it could be wind, fire steam transfer it from motion, atoms to electrons atoms to electrons we then take the electrons into our machine [15:00] of NVIDIA. Electrons now comes into our machine, comes in this factory, and what comes out are numbers. These numbers, depending on how you combine them, turns into language, math. It can also turn into a new language we've learned, proteins. We learned the language of human biology. We learned the language of the physical world, physics, climate, weather. We learned the language of the 3D world, robotics, self-driving cars. We learned the language of all kinds of different forms of intelligence. But the point being, now, these two machines, 300 years apart, atoms in, electrons out, electrons in, numbers out, and those numbers could be rejiggered, reformulated into all kinds of different intelligence. That's what we do for a living, and that's why I call it a factory, because it's producing, we call them tokens, but they're just numbers, tokens, and these tokens are intelligence. That it That what we do It not that hard Brilliant Now you know what AI is for, and now you know how AI is built, and how big it's going to be. Thank you all for joining today. Thank you. Good job. Excellent question. I really felt that carried it, you know, the prompt. You laid it up. Okay, so this is a massive revolution. Yeah. And you laid out three transformations. Energy transformation, which touches everyone today. And a lot of the people in the audience are part of these manufacturing and energy producers globally. Telecommunications, which connects all of us. And now intelligence. and in energy, you talked about the generator, telecommunications, I guess the comparable would be the switch or something along those lines for routing communications globally. And now in the intelligence revolution, it's the GPU at the core and the AI factory like the H100 or any of the new systems that bring everything you need under the same hood Vera Rubin what happens these factories so you know these generators it each one of our each one of our units we call a rack there 72 chips inside we we we um we manufacture uh call it eight million of them this year but but uh 72 of them go into a rack that rack weighs two tons It is $4 million, has one and a half million parts, and it's the most expensive piece of equipment in the world. And we manufacture them like, I guess it would be like phones. I mean, we crank them out, and they go into data centers all over the world. We build these machines in volume. They are big devices. This is how you do your weightlifting. Yes. I understand. No volume discounts. Okay. So you laid out this picture of a very exciting world that we are in. We're in the middle of this revolution. We you could say decades in you could say years in certainly now in the mainstream of the intelligence revolution How do we participate I sure everybody here wants to participate in this revolution Yeah excellent question Let start with big enterprises and then let's get to individuals as well. How do people join this movement? And so now I gave you two mental models. I'm going to give you one more mental model. So one mental model, they're really, you know, we could talk about this story and hopefully we cover all four phases. I just talked about what AI can do. I talked about how AI is made. And these things are in these factories. These factories are, you know, each gigawatt is about $50 billion. So have you ever seen it's the most expensive factory in the world? But it also that one $50 billion factory also generates $300, $400 billion in intelligence. And so the production value is incredible. Okay, the return on investment is extremely fast. And so these, so that's the factory part. The part that I'm going to tell you now, and this is very important when you think about investment, is what does the industrial layout look like for AI? And the way to think about the industrial view is think of it as a five-layer cake. Now, I told you on the bottom is energy. The bottom is, remember I said the dynamo, okay? Now we have, of course, different AC generators and things like that. [20:00] power generation. And so on the lowest layer is energy. This is the single greatest opportunity in several generations for the energy industries to grow. The very first time in probably, I don't know, 100 years since the energy grid in many countries could be invested in, this is the best opportunity to invest in sustainable energy. If you care about sustainable energy, nuclear, air, you know or wind solar you name it whatever form hydrogen whatever form is so long as it produces energy it's going to get funded and so that tells you something about how great of a time this is because we have a trillion dollars just think this one year this year alone we're going to put a trillion dollars from the market into this entire five layer cake i'm about to describe you so the first layer is energy that's the reason why siemens is doing so well that's mitsubishi is doing fantastically. GE, Vernova. I mean, everybody. The first layer of the cake is energy. The second layer of the cake is chips and computers and networking and switches and silicon photonics. Does that make sense? It's all the computers. The third layer of the cake, we call it infrastructure land power shell money data center operations every one of them is scarce applied today And so that's the next layer, it's the infrastructure layer. And then the layer that everybody sees, that everybody thinks is AI, is the model layer. Does it make sense? That's the next layer. It sits on top of the computers, the cloud infrastructure. And this is the greatest opportunity in recent human history that I know, that so much market-driven investment is naturally coming into the ecosystem. This is a great time to build. And so now that's the model layer. The model layer is OpenAI. It's anthropic. But this is the part that you can't overlook. This is very important. So you have two companies that you know of, that you hear about. However, don't forget, AI, as I was explaining earlier, has learned the language. It learns the language and the meaning. the mean the language the meaning of anything that is structural so that layer that what's really important so we hear we talk about all the language but don't forget we learn any you can learn anything with structure and so let me give you an example something with structure today when I walked in turn I was expecting a lot of people and it wasn unexpected the way you appeared Now if some of you were hanging off the ceiling and floating in midair and some of you you know the one human but body parts in 17 different places and I could see through some of you then it hard to learn that Because every time it's different, okay? Because it's hard to learn quantum things. However, things with structure we can learn, right? people have eyes and so you can learn these things okay and so 3d i learned the laws of physics uh i i sat down and notice i sat down with confidence i wasn't i wasn't 50 percent 53 percent of the time i landed safely on the chair the other 47 percent of the time i went right through it and so i can't trust it but 100 time you guys understand and so if if you if things are predictable and are predictable, then there's structure you can learn from it and you can learn the meaning of it. Okay. And so we learned the meaning of protein. We learned the meaning, we're learning the meaning of genes, not just sequencing it, not just CRISPR editing it, but what is the meaning of that gene? What is the meaning of a cell? Why does a cell do what the cell does What happens in two cells coming together And so this is no different than imagine I learned the meaning of a cell the way I learned the meaning of a word And what happens when I take two words put them together These two words activate each other, turn into something else of another meaning. Okay? And so from a computer's perspective, it doesn't care if it's a cell, a protein, a word, an image, a car. Does that make sense? It's just tokens. And so we have to figure out as computer scientists, we have to figure out how to represent the world's information in all these different ways so that the computer can understand it understand it understand it reason about it come up with a plan generate an action the intelligence loop proteins the same way cells the same way the human anatomy is the same way it must be predictable it's predictable because tomorrow morning i'm largely the same it must be predictable okay and so uh we're learning all these different my point is There are two language models that you guys know about, but AI is a giant industry. The industry of everything else physical is about $80 trillion. It is actually the most important frontier, the parts that we're not talking about. And then on top of that, this model, this technology then feeds into all the stuff that Constantine gets to see. [25:00] These days, which is all of these startups that are coming up with revolutionary ideas in financial services, in legal, in accounting, in transportation, logistics. Isn't that right? And so what that layer above this last year, a hundred billion dollars of venture capital investment, the single largest year of VC investment in the history of humanity. all of that money is going into that fifth layer the top layer which is the the layer that apply applications to enhance human condition and so there are five layers when you think about ai and you want to invest in this future and i promise you this future is going to be gigantic because two years ago zero it's approximately we're about to put one trillion dollars in but that's 1 trillion out of the You know, we're gonna be putting in probably the AI industry. I'm gonna guess for a second Probably something along the lines of 20 trillion dollars a year. We're 1 trillion dollars in of a 20 trillion dollars a year ecosystem Because the production of intelligence you just got to ask yourself how important is intelligence and who needs it and how much of it do you want And so those are kind of the basic questions And all of that intelligence, whether it's for proteins or cars or robots or language or math, science, whatever it is, has to be generated by these machines. And so this five-layer cake is the industrial version. And I think that's a good way to think about where to invest. Hugely so. So you've described what is a multi-trillion dollar opportunity to become part of this revolution. And that includes the hardware and the facilities. If it's $50 billion for a gigawatt and there's 100 plus coming online in the next several years, that is trillions plus the application layer where that is many, many more trillions plus, plus, plus. And that means real jobs for people doing the hands in building. Right now, exactly. We have to really emphasize this, and this is very important to you. Every country has a different attitude about AI today. Would you guys agree with that? Every country, because everybody's culture is a little different, okay? And here my recommendation Be careful with the analogies and the science fiction stories that this is Terminator and words like singularity and ideas that somebody say in 20% chance, this will be the end of humanity as we know it. Those kind of articulations of AI is just nonsense. It is complete nonsense. Oh, we have no idea how it works. This is so mysterious. we don't even know how it works. It might just get up out of its seat and walk out tomorrow morning. There's no question in my mind it's computer and software. And there's no question in my mind they know how it works. And do you know how they know how it works? Because every single year, apparently, it's getting better. If you don't know how something works, how do you make it better? I have no idea how it works, but I know how to make it better. That's nonsense. So why are they saying these things? That's an interesting question. However, don't let it scare you. You must engage it. You may or may not lose a job to an AI, but you will absolutely lose a job to someone who uses AI. Would you agree with that? Okay, so let's not worry about the things you not sure about and focus on the things that you are sure about I am absolutely certain I will lose my job to someone who uses AI So before I worry about AI let just go make sure I use AI And so the part that now why is that common sense so important for me to tell you? Because some of you have children. What are you advising them? Run away? Or make sure whatever this technology is that gives people superpowers, that you go make sure you use it. So I'm hoping that we do two things. One, we are and we're doing everything we can to build this technology safely for the world. I promise you, so much computer science, so much investment, so much passion dedicated to making this technology safe for everybody to use. And I can prove it. Use ChatGP2 two years ago and use it again. the amount of hallucination completely reduced to almost nothing to the point where it's producing knowledge not only accurately contextually relevant relevant to the moment and if it doesn't know the answer it does research and when it comes up with an answer it even questions itself before it [30:00] tells you an answer, it reflects on it. And it comes up with two or three different answers, and it reflects on those before it produces the answer for you. The amount of safety and guard railing, grounding of truth, the technology advanced so fast to make it safe. I am certain I can tell you this completely with fact. I prefer my car today than the car that was 100 years ago. the technology is a lot better but it's a lot safer and it takes a lot of technology to be invented in order for it to be safe and so i can tell you that it is our job it is the responsibility of the technology industry things the responsibility scientists and engineers for us to build ai safely two it is your responsibility to make sure that you tell the people that you love whether it's your family, your kids, your grandkids, or the company you work for, or the country that you're in. Whatever we do, engage AI. If we think it's a superpower, engage it. Because if we don engage it somebody else will We not going to lose our lives to AI We going to lose our lives to somebody who uses AI And so that my Well that too serious That's too serious. That was because he said the word job. So I've got a trigger, and the trigger is a bunch of people making stuff up about jobs. We put a trillion dollars into the world's ecosystem this year, did we not? What's it doing? Making jobs. right now, the energy sector, more jobs than ever. The chip sector, more jobs than ever. Infrastructure layer, more jobs than ever. Everything from land, power, shell, finances. AI model layer, more jobs than ever. And we just said $100 billion last year went into the upper layer, more jobs than ever. We're creating so many more jobs. Now, somebody might say, well, what about the traditional jobs? So let me give you the example. You know that everybody's job and their task is related not the same A job and the task you do in the job is related not the same So for example, my job is to be the CEO to lead the company Most of most of the time and I spent a lot of it today most of my time my task is Typing and talking and so you could say ceo equals typing and talking both of them ai does in a superhuman way and i busier than ever and i giving you that of course that's a really cute example but let me give you the a more a deep example and you can now apply it. So 10 years ago, slightly more than that, one of the world's leading computer scientists wanted to warn everybody about the power of AI. And so he said, and Constantine probably knows who it is, he said the first job that AI will destroy and eliminate, and I advise nobody goes into this field because this field will be wiped out, is radiology. Computer vision is superhuman, already 12 years ago. Computer vision. A computer can recognize images, detect anomalies with superhuman capability, never gets tired, never misses detail. 12 years ago, it was able to do that and he predicted as a result of that, radiology is going to be wiped out. Well, he was absolutely right. Radiology was completely penetrated by computer vision Computer vision proliferated through every single form of radiology and every radiology stack and every radiologist today is augmented by computer vision. However, the interesting thing is this. Radiology demand went up. The number of radiologists in the world went up. Why? Audience participation, please. Why? I heard some of the things. It's all true. It turns out radiology spends a lot of time studying scans. But the purpose of the radiology, the purpose of the radiologist is to work with doctors to diagnose disease. To work with doctors to diagnose disease. And because it's now automated, they are more productive. So two things happen. more patients are admitted into the hospital they do more scans the radiology department became more profitable when they realized they were more profitable and they were admitting more patients they hired more radiologists so that they could admit more [35:00] Patients so that they could make more money take care of more people because as it turns out there are a lot of people Who are suffering and they're waiting to get into the hospital? so now Let's pretend for a second Do you appreciate the computer scientists tell you it's going to be the end of the world for radiologists? my point is we have to be responsible about what we make up because we could have done harm and it turns out the number of people who want to be radiologists after his speech because it permeated through everything the number of radiologists started to decline but we need more radiologists now somebody recently said 90 of software coding will be gone and therefore we don't need software engineers. Meanwhile, we're hiring more software engineers than ever. And the reason for that is because a software engineer's job is to solve problems and dream up problems to solve. Innovate. I never hired somebody and said, hey, guess what? You're a software engineer. Listen, here's a keyboard. Show me how many words a second you can type. Typing is not the job of a software engineer Coding is not their job Solving problems is their job And so I just gave you two examples Task versus purpose. Does that make sense? It turns out this example happens all over the place. But because we have such a contrived, such a naive understanding that computer scientists could say things like 50% of the jobs will be gone, software coding is completely irrelevant radiology is going to be wiped out because we think about it from the task perspective we forgot the purpose of the job there were radiologists before there were workstations they're going to be radiologists after ai there were soft there were engineers before software coding i promise you that there will be engineers after does that make sense And so that's the way to think about jobs. And I've now covered two things. One, if your country is not investing in AI, there's a massive boom of jobs that you're missing out on. If your country or your company is not investing in AI there a level of elevation of your people that you missing out on AI is not going to eliminate jobs AI is going to elevate your job if I were a plumber today Largely I get a job task sheet or a schematic However, if I'm a plumber tomorrow, it is very likely I'm a designer as well Does that make sense because you and I both know yeah that we could just use AI to generate decent incredible designs of a kitchen if i'm a carpenter does i can turn if i were a sales person of furniture i'm going to be an interior interior designer for sure and so i've elevated my craft went from somebody who sells furniture to somebody who could advise you on how beautiful your home could be i went from somebody who's a carpenter you expected me to come and just you know put some wood together and now i'm your home designer you i've elevated my craft i've given you so many examples but that's my point i think the narrative about ai is absolutely wrong and the goal is to scare everybody out of it so that some people could benefit from it but ai as you know is the greatest force for eliminating the technology divide in my entire career I spent 40 years My entire life has been in computer design. I spent 40-some-odd years, and this entire time, the technology we created became more and more and more and more complex. And the number of people who could program these computers as a percentage of population declined. Who in this room knows C++? come on cut it out you weirdos this row is just is a startup company okay and so so so this row okay so we're looking at two percent okay two percent and this is a very strange room this is a very strange room and so two percent of society knows c plus plus how many people know human okay more than two percent and so and so everybody now know everybody now can program a computer and yet in the past only 2% can. We have closed the technology divide. We got to bring everybody with us. Does that make sense? Okay. Anyways, that's it. On a Friday night, that's too serious. That is extremely optimistic and I agree. And it's great to hear from someone who is closer to... [40:00] building of the actual technology that powers everything than anyone else in the world. So Jensen, you talked about a future where we move from retrieval, this paradigm that we've had our entire lives in, to generation, where everything is customized, knowledge is customized for the individual, a world where we have the generation of intelligence paralleling from the energy to the telecommunications revolution to now the intelligence revolution. You talked about these languages that the computer can speak, not just English or German, but even protein. You talked about five layers of participation, an abundant opportunity to participate in this revolution for everyone in this room and everyone listening. And you talked about how this transformation is going to be something that has real consequences, real consequences that allow people to move from just doing the task to dreaming the problems and solutions, maybe even a life of purpose and a life where we move from carpenters to architects. Thank you, Jensen. Please join me in thanking Jensen Huang, the man who made this all happen. Thank you. Thank you so much. You're welcome. Thank you. Thank you. Thank you. Thank you. Thank you. Thank you. Thank you.