← 回總覽

「推理之王」黄仁勋:你们都误会了,1 万亿美元其实很保守

📅 2026-03-18 10:07 极客公园 人工智能 12 分鐘 14753 字 評分: 78
英伟达 黄仁勋 AI Agent 智能体 GTC 2026
📌 一句话摘要 黄仁勋在 GTC 2026 上阐述了英伟达从硬件向 Agentic AI 基础设施转型的战略,指出 1 万亿美元市场预测仅是保守起点,并发布了企业级 Agent 平台 NemoClaw。 📝 详细摘要 本文深度报道了英伟达创始人黄仁勋在 GTC 2026 期间的战略思考。黄仁勋指出,此前预测的 2027 年 1 万亿美元 AI 基础设施市场规模实际上非常保守,因为该数字仅涵盖了 Blackwell 和 Vera Rubin 核心方案,未计入 CPU、存储系统及后续架构的贡献。他提出了 AI 演进的三个关键拐点:ChatGPT(生成)、o1(推理)以及 OpenClaw(智能
Skip to main content ![Image 7: LogoBestBlogs](https://www.bestblogs.dev/ "BestBlogs.dev")Toggle navigation menu Toggle navigation menuArticlesPodcastsVideosTweetsSourcesNewsletters

⌘K

Change language Switch ThemeSign In

Narrow Mode

「推理之王」黄仁勋:你们都误会了,1 万亿美元其实很保守 ============================

!Image 8: 极客公园 极客公园 @极客公园

One Sentence Summary

At GTC 2026, Jensen Huang outlined NVIDIA's strategy of transitioning from hardware to Agentic AI infrastructure, stating that the $1 trillion market forecast is merely a conservative starting point, and unveiled the enterprise-grade Agent platform, NemoClaw.

Summary

This article provides an in-depth report on NVIDIA founder Jensen Huang's strategic thinking during GTC 2026. Huang pointed out that the previously predicted $1 trillion AI infrastructure market size for 2027 is actually very conservative, as this figure only covers the Blackwell and Vera Rubin core solutions, without accounting for the contributions of CPUs, AI storage systems, and subsequent architectures. He proposed three critical inflection points in AI evolution: ChatGPT (generation), o1 (inference), and OpenClaw (Agentic systems). By launching the NemoClaw platform, NVIDIA aims to transform the AI Agent strategy from a privilege of tech giants into universally accessible infrastructure. Furthermore, Huang shared his 'three no's' principle for running a company and envisioned an organizational future where 75,000 human employees collaborate with 7.5 million AI Agents.

Main Points

* 1. The $1 trillion market forecast is severely underestimated, representing merely the tip of the iceberg for NVIDIA's business.This forecast is based solely on sales expectations for Blackwell and Vera Rubin architectures, and does not include the significant incremental growth potential from CPU business, AI storage systems, network interconnects, and subsequent Feynman architectures. * 2. AI development is entering its third inflection point: the engineering popularization of Agentic Systems.Following generative AI and inference systems, OpenClaw marks AI's entry into a new phase of autonomous learning, self-accumulation, and capability transfer, making enterprise-grade Agent strategies possible. * 3. Future enterprise organizational structures will undergo fundamental changes, achieving an exponential increase in the human-to-machine ratio.Jensen Huang envisions 75,000 human employees driving 7.5 million AI Agents to work around the clock, completely transforming sci-fi scenarios like robotics and autonomous driving into engineering realities.

Metadata

AI Score

78

Website mp.weixin.qq.com

Published At Today

Length 2343 words (about 10 min)

Sign in to use highlight and note-taking features for a better reading experience. Sign in now

!Image 9!Image 10

黄仁勋在 GTC 上说的那个数字,其实只是入门价。 作者|靖宇

当一个人被冠以「推理之王」、世界首富、AI 行业(甚至是全球)最有影响力的人,他会怎么说话,怎么给别人建议?

别被炒鱿鱼、别弄得太无聊、别死了。」

在 GTC2026 年第二天的媒体会上,英伟达创始人黄仁勋给出了自己的「黄氏三不」原则——即便在 GTC 这样高强度的工作之中,黄仁勋依然能用玩笑的口吻给出自己认真的思考。 当地时间 3 月 16 日的开场演讲中,黄仁勋给出了「1 万亿美元」这个 2027 年的目标,在 17 日的记者会上,黄仁勋再次详细解释,这个数字不是英伟达公司的营收,而仅仅是 Vera Rubin 和 Blackwell 方案的预期销售额。

英伟达站在了 Agentic AI 时代的起点,黄仁勋看到了未来,现在他要让人们再次相信他看到的、让人难以置信的画面。

_01_ **「一万亿」,被低估的起点**

就在昨天的 GTC Keynote 上,黄仁勋发布了 Vera Rubin 完整平台——Vera CPU 加 Rubin GPU 构成的新一代计算架构,配套 NVLink 72 机架级互联,在推理吞吐量上比上一代 Blackwell 提升 35 倍,成本压到三十五分之一。这是 NVIDIA 目前最强的推理系统,也是他口中「明天的推理王者」。

但在当天下午的媒体问答上,黄仁勋想聊的,是一件更大的事:那个被媒体反复引用的「1 万亿美元预测」,他说算少了。

去年 GTC,他预测到 2027 年 AI 基础设施市场规模将从 5000 亿美元增长到 1 万亿美元。这个数字立刻成了整个科技圈的讨论焦点。这次站在台上,黄仁勋亲自拆解了那 1 万亿究竟算了什么:只有 Blackwell 和 Vera Rubin,只到 2027 年。他刻意保持和去年同样的口径做对比,以便让外界看清增速。

但被排除在外的部分,才是真正的「冰山水下」。

!Image 11!Image 12 1 万亿美元只是起点|图片来源:极客公园

单独销售的 CPU 业务——Vera 和 Grace 架构——将会是「数十亿美元」量级的独立业务,跟 GPU 收入分开算。然后是 Groq,他给出了一个极为具体的数字:给一个 1 吉瓦的数据中心部署 Groq,可以额外提升 25% 的经济效益,对应的就是再加 2500 亿美元进来。还有 STX、QDF 这些 AI 存储系统,还有 BlueField,还有 Vera Rubin Ultra,还有下一代架构 Feynman,还有 Feynman Next。 这一串东西,一个都没进那 1 万亿

黄仁勋说,从他站着的地方往前算,还有 21 个月。言下之意是,这只是已知的、最保守的可见范围。

这让人想起一个对比:三年前,谁能预测 NVIDIA 会从一家游戏显卡公司变成全球最重要的基础设施供应商?而现在,连「最保守估计」都已经是 1 万亿美元起步。

_02_ **AI 的第三拐点**

同一场 Keynote 上,NVIDIA 还正式发布了 NemoClaw——这是 NVIDIA 面向企业的 AI Agent 平台,叠加了企业级安全、合规治理和高性能运行时,被视为 NVIDIA 在「后推理时代」的新布局。理解 NemoClaw 是什么,得先听黄仁勋怎么讲它的母体:OpenClaw。

黄仁勋把 AI 这几年的演进,归纳成三个拐点。

第一个是 ChatGPT,它把生成式 AI 带给了普通人。第二个是 o1,它把推理系统带给了世界,让 AI 从「生成」进化到「思考」。而第三个拐点,黄仁勋认为是 OpenClaw。

他用的词是「推理之后的 AI 时代的拐点」。

OpenClaw 是开源的、工程化完善的 Agentic System——一套让任何人、任何企业都能拥有自己的 Agent 战略的基础设施。在这之前,Claude Code 这类 AI 编程和 Agent 工具,主要活在大企业内部和顶级科技公司里,大多数人虽然听说过「AI Agent」这个词,但从来没真正用上过。OpenClaw 出来之后,几周内就成了有史以来增长最快的开源项目——Linux 花了 30 年积累的声望,OpenClaw 在几周内就达到了类似的关注度。 而 NemoClaw,就是「OpenClaw + NVIDIA 技术栈」的企业版本:在 OpenClaw 基础上,叠加企业级安全、合规治理、隐私保护,以及 NVIDIA 的高性能运行时和参考设计。一行代码安装,任何公司都能跑起来一个高性能的开放模型。

他在台上描述了一个让人印象深刻的场景:

你打开 OpenClaw,告诉它:「帮我重新设计厨房,预算上限是 X,自己去学需要的工具,订阅必要的服务,找家具,然后把渲染图给我看。」 然后它真的就去做了。

它会自己摸索,学会它不会的工具,生成第一稿渲染图,然后跟你给它的参考风格对比,找出差距,再迭代。完成之后,它会把整个过程总结成一份「技能文档」。下次你跟它说「帮我设计客厅」,它已经是这个领域的专家了。 这不是演示里的假 demo,这是 Agentic System 的底层逻辑:自主学习、自我积累、能力可迁移。

黄仁勋说,OpenClaw 让「拥有自己的 AI Agent 战略」这件事,从科技公司的专属特权,变成了所有人触手可及的选项。这个门槛的消失,才是真正的拐点意义。

!Image 13!Image 14 黄仁勋:嫉妒活在未来的人|图片来源:极客公园

_03_ **「三个不要」,和一份嫉妒未来的宣言**

记者问黄仁勋,NVIDIA 最大的风险是什么?

他没有给出华尔街分析师会写的那种答案。他说了三个不要:不要被客户炒鱿鱿鱼,不要让公司陷入无聊,不要倒闭。

这三条听起来像废话,但他是认真的。

被客户炒鱿鱼是真实风险——竞争对手在追,AMD 在追,自研芯片的大客户们在追。让公司陷入无聊则是更隐性的危险:一旦组织开始享受惯性、开始管理成功而不是创造成功,那比竞争对手更危险。而倒闭——在这个速度下,一次重大判断失误就可以是致命的。

他说这三条每年都是高风险,去年是,今年是,永远是。

然后他聊到了未来十年。

NVIDIA 现在有 7.5 万名员工。他的设想是,这 7.5 万人,配上 750 万个 AI Agent,全天候不间断工作。人类员工不需要跟 Agent 同步作息,但 Agent 不会停歇。

这不是比喻,这是他在认真描述组织形态的变化。

机器人、自动驾驶、药物发现,正在从「科幻」变成「工程问题」。黄仁勋说,十年前没有人敢相信这些是可以被解决的,但现在,完全可信。

他说他嫉妒 40 年后还活着的人。因为那时候摆在人类面前的问题,是我们今天连想象都想象不到的。

这句话有点意思。通常站在时代浪尖的人,会享受当下的位置。但黄仁勋选择了嫉妒未来——不是担忧,而是嫉妒。 说白了,他自己也不知道终点在哪里。但他有把握的是,方向没有错,加速还在继续。

1 万亿,只是个开始。

!Image 15: 极客公园 极客公园 @极客公园

One Sentence Summary

At GTC 2026, Jensen Huang outlined NVIDIA's strategy of transitioning from hardware to Agentic AI infrastructure, stating that the $1 trillion market forecast is merely a conservative starting point, and unveiled the enterprise-grade Agent platform, NemoClaw.

Summary

This article provides an in-depth report on NVIDIA founder Jensen Huang's strategic thinking during GTC 2026. Huang pointed out that the previously predicted $1 trillion AI infrastructure market size for 2027 is actually very conservative, as this figure only covers the Blackwell and Vera Rubin core solutions, without accounting for the contributions of CPUs, AI storage systems, and subsequent architectures. He proposed three critical inflection points in AI evolution: ChatGPT (generation), o1 (inference), and OpenClaw (Agentic systems). By launching the NemoClaw platform, NVIDIA aims to transform the AI Agent strategy from a privilege of tech giants into universally accessible infrastructure. Furthermore, Huang shared his 'three no's' principle for running a company and envisioned an organizational future where 75,000 human employees collaborate with 7.5 million AI Agents.

Main Points

* 1. The $1 trillion market forecast is severely underestimated, representing merely the tip of the iceberg for NVIDIA's business.

This forecast is based solely on sales expectations for Blackwell and Vera Rubin architectures, and does not include the significant incremental growth potential from CPU business, AI storage systems, network interconnects, and subsequent Feynman architectures.

* 2. AI development is entering its third inflection point: the engineering popularization of Agentic Systems.

Following generative AI and inference systems, OpenClaw marks AI's entry into a new phase of autonomous learning, self-accumulation, and capability transfer, making enterprise-grade Agent strategies possible.

* 3. Future enterprise organizational structures will undergo fundamental changes, achieving an exponential increase in the human-to-machine ratio.

Jensen Huang envisions 75,000 human employees driving 7.5 million AI Agents to work around the clock, completely transforming sci-fi scenarios like robotics and autonomous driving into engineering realities.

Key Quotes

* Don't get fired, don't get bored, don't die. * $1 trillion is actually very conservative. * OpenClaw makes 'having your own AI Agent strategy' transform from an exclusive privilege of tech companies into an accessible option for everyone. * I envy those who will be alive 40 years from now. Because the problems facing humanity then are ones we can't even imagine today.

AI Score

78

Website mp.weixin.qq.com

Published At Today

Length 2343 words (about 10 min)

Tags

NVIDIA

Jensen Huang

AI Agent

Agent

GTC 2026

Related Articles

* GPT-5.4 Released: OpenAI's First Unified Model, Truly Native * After Reading the New 23,000-Word "AI Constitution," I Understand Anthropic's Pain * 2026: The Explosion of 'One-Person Companies'—No Employment, No Layoffs * This Smart Ring Achieved 4 Million Shipments: An Interview with the Founder * How to Design an AI Agent System * What You Don't Know About Claude Code: Architecture, Governance, and Engineering Practices - Tw93 * 128. Manus Co-founder's Final Interview Before Sale: The Fantastical Drift to 2025... * Nvidia: Unrivaled in Next-Gen AI Computing * Deconstructing Clawdbot: Local Architecture, Memory Management, Agent Orchestration, and Context Assembly Principles * Four Chinese LLM 'Trainers' Discuss China's AI Landscape in 2026 HomeArticlesPodcastsVideosTweets

Jensen Huang, the 'King of Inference': You've Got It All ... ===============

查看原文 → 發佈: 2026-03-18 10:07:00 收錄: 2026-03-18 14:00:41

🤖 問 AI

針對這篇文章提問,AI 會根據文章內容回答。按 Ctrl+Enter 送出。