黄仁勋整场演讲，其实就讲了一条主线：从美国过去的科技发明 → 到传统芯片变慢、摩尔定律失效 → 到用新型 GPU 和"AI 工厂"继续提速 → 再到让 AI 走出屏幕，变成懂物理世界的机器人。
他的核心意思是：
下一轮工业革命的"电厂和工厂"，会是 AI 工厂 + 物理 AI（机器人），而英伟达要用 Blackwell / Rubin 这些新平台和全套生态来当发动机，同时把生产基地放在美国，配合"美国重新搞制造业"的大方向。
1. 从美国创新史，到"AI 工厂时代"
开场用一段很燃的视频回顾美国的发明：晶体管、微处理器、超级计算机、个人电脑、互联网、智能手机、登月……
重点是一个：每一代大突破，背后都有更强的"算力"和工程能力。
现在，他说我们又到了一个类似"登月计划"的时刻：AI 会像电力、互联网一样，变成所有国家、所有企业都离不开的基础设施。
2. 摩尔定律不灵了，要换玩法：加速计算
以前有个"摩尔定律"：大概每两年，芯片上的晶体管数量翻倍，电脑就越来越快又省电。
现在问题来了：晶体管数量还在涨，但性能和省电的提升跟不上了，传统 CPU 那种一条一条顺序算的方式，遇到物理瓶颈。
英伟达这 30 年押的牌，是另一条路：加速计算（GPU + CUDA）
可以理解为：CPU 是一个人算题，GPU 是一大群人一起算。真正难的是：怎么让这一大群人配合好，所以英伟达花大力气做了 CUDA 和 350+ 个软件库。
这些库已经渗透到很多行业（芯片制造、金融、医学、量子、图形……），形成一个软件 + 硬件互相推动的良性循环：用的人越多 → 生态越好 → 更多人愿意用 → 芯片卖得越多 → 再投钱把生态做得更好。
3. 新概念："AI 工厂"取代传统数据中心
传统数据中心是：存文件、跑各种网站和应用（像一个大型综合商场）。
AI 工厂不一样：它专门干一件事——生产"智能答案"。
这些"答案"在技术上叫 Token，可以是文字、图片的一部分、化学分子的一块、3D 结构的一片，甚至是机器人一个动作的单位。
为什么要"工厂"级别？因为有两个"指数级"在叠加：模型越来越大、越来越聪明，所以训练和"思考"本身就越来越耗计算。模型越好用，用的人越多，用的场景越多，调用次数（也就是推理量）也涨得飞快。
结论：全世界的算力被这个需求往死里拉，唯一活路是——做出更极端优化的系统，让"每个答案（每个 Token）的成本持续降下来，否则这套循环会被自己撑爆。
4. Blackwell & Rubin：从"做芯片"变成"做整套机器"
传统做法：设计一块芯片，其他系统交给别人拼。
英伟达现在的做法：把 芯片 + 封装 + 芯片之间的高速连接（NVLink）+ 机柜内部的网络（Spectrum-X）+ 整个机架 + 整个机房 当作一个整体的超级电脑来设计，这就是所谓"极限协同设计"。
结果：新一代 Blackwell（GB200/GB300 + NVLink 72 机架） 相比上一代 H200：➜ 性能大约提升 10 倍 ➜ 生成同样多 Token 的成本，大约只有原来的 1/10。
生意上：未来几年已经能看到的订单规模，大约 5000 亿美元级别。Blackwell GPU 出货量预期远远超过上一代 Hopper。
Rubin 是下一代：机架做到几乎没有线缆、全部液冷，散热和维护更高效。目标是一年一代，每年再把"每个答案"的成本往下砸。
5. "美国制造" + 再工业化
演讲里很详细地展示了 Blackwell 的生产路线：亚利桑那做晶圆 → 印第安纳做高带宽内存 → 德州做整机 → 最后组装成完整 AI 机架。
这不只是炫技，背后还有两个意思：回应美国"制造业回流"的政治诉求。为供应链安全和核心技术"握在自己手里"做布局。
他特别提到：从被要求"把制造带回美国"，到在亚利桑那实现 Blackwell 量产，中间只用了 9 个月，并多次感谢美国政府的能源和制造政策支持。
6. 6G + NVIDIA ARC：把基站变成"AI 小机房"
英伟达和诺基亚合作，推出一个新系统：NVIDIA ARC。
简单理解：以前的基站，只负责"信号塔"这件事。现在的基站，要变成一个小型"算力节点"，既能负责通讯，也能顺便跑 AI。
它能做两件关键事：用 AI 优化通讯本身——让信号分配更聪明、频谱利用更高、用更少电传更多数据。在基站上跑 AI 应用——比如附近有机器人、无人机、自动驾驶车，都可以直接连到最近的基站上用 AI 算力。
大目标是：下一代 6G 的技术标准和核心设备，美国要重新成为主导之一。
7. 量子计算：让"未来计算机"和现在的超算一起干活
量子计算的关键进展：业界终于做出了可以纠错、相对稳定的"逻辑量子比特"。
但现实问题是：量子比特非常脆弱，需要大量纠错运算，这反而非常吃传统算力。
英伟达给的方案是：用 NVQ Link 把量子芯片（QPU）和 GPU 超算高速连起来。让 GPU 帮忙做控制、纠错、模拟，QPU 专心干自己擅长的那部分。通过 CUDA-Q 这个平台，把两者当一个整体来编程。
他们还联合多家量子硬件公司和美国能源部 8 大实验室，建设 7 台新的 AI 超级计算机，让：传统计算 + AI + 量子 一起为国家级科研服务。
8. AI 重写了整个"计算堆栈"：从工具变"数字员工"
过去的模式：程序员手写代码 → 在 CPU + 操作系统（如 Windows）上跑 → 给你一个工具（Excel、Word、浏览器）。这些都是"工具"，需要人来用。
现在的模式：把各种数据切成"Token"，在 GPU 上训练出来的模型，本身就能思考、推理、决策。它不仅会用工具，还能帮你把很多工作直接干了。
黄仁勋的角度是：AI 不再只是工具，而是可以自己用工具的"数字劳动力"。这意味着：AI 不会只影响几万亿美元的 IT 行业，而是会跑到约 100 万亿美元的实体经济里去——制造、医疗、物流、金融都要被重塑。
而且，真正最耗算力的不是训练，而是：预训练——学基础知识。后训练——学解题方法、推理方式。日常"思考和推理"——不停地理解新问题、查资料、规划步骤，再给出答案。这一整套加起来，远比很多人想象中更吃算力。
9. Omniverse DSX：先在"虚拟工厂"里造好，再去现实世界装
NVIDIA Omniverse DSX 被当作一种"设计和运营超大规模 AI 工厂的说明书 + 模拟器"。
用大白话说，就是：先在一个非常逼真的 3D 数字世界里，把整个 AI 工厂——楼怎么建、机房怎么布局、电怎么走、冷却怎么做、机柜怎么排——都模拟一遍。和西门子、施耐德、Vertiv、Bechtel 等传统工程巨头一起，用仿真软件反复试。
这样做的好处：工厂还没建就先"试运营"，找出问题再开工，建造周期大幅缩短。建成之后，这个数字孪生就变成工厂的"控制面板"：AI 代理可以在虚拟空间里模拟不同的运营方式，看怎么省电、省钱、产量高，再指导真实工厂。
对一个 1 吉瓦（非常巨大的）AI 工厂来说，这种优化每年可能带来数十亿美元级别的经济收益。
10. 开源模型和企业生态：不只自己赚钱，还要"带生态一起飞"
黄仁勋特别强调：美国要在开源 AI 模型上保持领先，这对研究、开发者、初创公司都非常关键。英伟达自己做了很多模型并开源出来（语音、推理、物理世界理解、生物学等），下载量非常大。
同时，他们把自己的软件和模型：深度整合进各大云（AWS、Google Cloud、Azure、Oracle）。嵌到各种企业软件里：ServiceNow、SAP、Synopsys、Cadence 等。和 CrowdStrike 做安全 AI，和 Palantir 做"数据 + 决策"的智能平台。
意思很简单：不管你在云上、在大企业、在安全或数据平台，最后背后跑的那一层算力，很可能都是英伟达的。
11. 物理 AI：让 AI 走出屏幕，进工厂、进手术室、进乐园
"物理 AI"的定义：不仅会算和聊天，还要理解真实世界的物理规律（重力、摩擦、碰撞、因果）并能在现实中行动的 AI。
英伟达觉得要做好这件事，需要"三台电脑"配合：训练用超算——大模型在哪里被"教会本事"。数字孪生仿真机——用 Omniverse 做虚拟世界，给机器人一个"安全的练习场"。机器人身上的小电脑——Jetson Thor 放在机器狗、人形机器人或自动驾驶车上，负责实时控制。
实际合作场景包括：跟富士康、西门子搭建数字化工厂——先在虚拟工厂里训练机器人，再让它们进真实产线干活。跟 Figure、Agility Robotics 做人形机器人、仓储机器人。跟强生做手术机器人，在数字孪生手术室里反复练习复杂手术。跟迪士尼研究院做超级可爱的机器人 Blue，用新一代物理引擎 Newton 让它在模拟世界里学会走路、摔倒、再爬起来。
12. 轮子上的机器人：无人车 + Uber 网络
在各种机器人里，真正最接近大规模落地的，是"轮子上的机器人"——也就是无人驾驶出租车（RoboTaxi）。
英伟达搞了一个标准方案：Drive Hyperion——帮车厂定好：车上要装哪些摄像头、雷达、激光雷达、用什么计算单元。车厂照着这个标准造车，自动驾驶公司就可以把自己的软件往上面一装，直接测试和商用。
他们还和 Uber 合作：以后打车软件里，除了人类司机的车，还可能有"无人驾驶车车队"可以叫。这些车背后，用的就是类似 Hyperion 这种统一的平台。
简单说：英伟达想把"无人车"变成全球统一的"机器人计算平台"，再接入到 Uber 这样的超级网络里。
13. 收尾：两个"世纪级转型"叠在一起
黄仁勋最后的归纳是：我们正处在两次大转型叠加的时代——
从 通用计算 → 加速计算（CPU 为主 → GPU 为主）
从 手写软件 → 人工智能（模型成为主角）
这两股浪叠在一起，再加上 6G、量子、AI 工厂、物理 AI、机器人和无人驾驶，带来的就是：计算机行业自身被彻底重塑；美国有机会通过"AI 工厂 + 制造业回流"，在新一轮科技和工业革命中重新占据制高点。
Jensen Huang's entire speech essentially followed one main thread: From America's past technological inventions → to traditional chips slowing down and Moore's Law reaching its limits → to using new GPUs and "AI factories" to continue accelerating → to enabling AI to move beyond the screen and become robots that understand the physical world.
His core message is:
The "power plants and factories" of the next industrial revolution will be AI factories + Physical AI (robots), and NVIDIA aims to be the engine driving this with its new platforms like Blackwell / Rubin and its full ecosystem, while locating production bases in the U.S., aligning with the broader direction of "America's reindustrialization."
1. From American Innovation History to the "AI Factory Era"
Opening with an inspiring video montage of American inventions: Transistors, microprocessors, supercomputers, personal computers, the internet, smartphones, moon landings...
The key point: Every major breakthrough was backed by more powerful "computing power" and engineering capability.
Now, he states we are at another "moonshot" moment: AI will become infrastructure as indispensable as electricity or the internet for every nation and every enterprise.
2. Moore's Law is Fading; Time for a New Game Plan: Accelerated Computing
Moore's Law used to state: Roughly every two years, the number of transistors on a chip doubles, making computers faster and more power-efficient.
The problem now: Transistor counts are still increasing, but gains in performance and power efficiency are lagging. The traditional sequential computing approach of CPUs has hit physical limits.
NVIDIA's bet over the past 30 years has been on a different path: Accelerated Computing (GPU + CUDA)
Think of it as: A CPU is one person solving a problem, while a GPU is a massive group of people solving it together. The real challenge: How to get this massive group to coordinate effectively. That's why NVIDIA invested heavily in CUDA and 350+ software libraries.
These libraries have permeated numerous industries (chip manufacturing, finance, medicine, quantum, graphics...), creating a virtuous cycle where software and hardware drive each other: More users → better ecosystem → more users adopt it → more chips sold → more investment to improve the ecosystem.
3. New Concept: "AI Factories" Replace Traditional Data Centers
Traditional data centers are for: Storing files, running various websites and applications (like a large shopping mall).
AI factories are different: They specialize in one thing—producing "intelligent answers."
Technically, these "answers" are called Tokens, which can be parts of text, images, chemical molecules, 3D structures, or even units of motion for a robot.
Why does it need "factory" scale? Because two exponential trends are converging: Models are getting larger and smarter, making the training and "thinking" process itself increasingly computationally demanding. The better the models, the more users and use cases emerge, leading to a rapid surge in the number of calls (inference volume).
Conclusion: Global computing demand is being stretched to the limit by this need. The only viable path forward is—creating extremely optimized systems that continuously lower the cost per answer (per Token); otherwise, this cycle will collapse under its own weight.
4. Blackwell & Rubin: From "Making Chips" to "Building Complete Machines"
Traditional approach: Design a chip, let others assemble the rest of the system.
NVIDIA's current approach: Treating chip + packaging + high-speed chip-to-chip interconnects (NVLink) + intra-rack networking (Spectrum-X) + the entire rack + the entire data center as one integrated supercomputer to design. This is "extreme co-design."
The result: The new-generation Blackwell (GB200/GB300 + NVLink 72 rack) compared to the previous H200: → Performance improves by ~10x → The cost to generate the same amount of Tokens is only ~1/10th.
On the business side: The visible order pipeline for the coming years is on the order of $500 billion. Blackwell GPU shipments are expected to far exceed the previous Hopper generation.
Rubin is the next generation: Racks are designed to be nearly cable-free, fully liquid-cooled for more efficient cooling and maintenance. The goal is a one-year generation cycle, relentlessly driving down the "cost per answer" each year.
5. "Made in America" + Reindustrialization
The speech detailed Blackwell's production route: Arizona (wafers) → Indiana (HBM memory) → Texas (full systems) → final assembly into complete AI racks.
This isn't just for show; it carries two underlying messages: Responding to the political push for "reshoring manufacturing" in the U.S. Securing supply chains and keeping core technology "in our own hands."
He specifically noted: It took only 9 months from being asked to "bring manufacturing back to the U.S." to achieving Blackwell production in Arizona, and he repeatedly thanked U.S. government support on energy and manufacturing policy.
6. 6G + NVIDIA ARC: Transforming Base Stations into "AI Micro Data Centers"
NVIDIA collaborated with Nokia to launch a new system: NVIDIA ARC.
Simply put: Traditional base stations only handled being "signal towers." Now, base stations will become small "computing nodes," handling both communication and running AI workloads.
It enables two key things: Optimizing communication itself with AI—Smarter signal allocation, higher spectrum efficiency, transmitting more data with less power. Running AI applications on the base station—Nearby robots, drones, or autonomous vehicles could directly connect to the nearest base station to access AI computing power.
The grand goal: For the next-gen 6G technology standards and core equipment, the U.S. aims to re-establish itself as a leading force.
7. Quantum Computing: Making the "Future Computer" Work with Today's Supercomputers
Key progress in quantum computing: The industry has finally produced relatively stable, error-corrected "logical qubits."
But the practical problem: Qubits are extremely fragile, requiring massive error-correction computations, which ironically demands significant traditional computing power.
NVIDIA's proposed solution: Use NVQ Link to connect quantum chips (QPUs) to GPU supercomputers at high speed. Let GPUs handle control, error correction, and simulation, allowing QPUs to focus on their specialized tasks. Use the CUDA-Q platform to program both as a unified system.
They are also collaborating with multiple quantum hardware companies and eight U.S. Department of Energy national labs to build 7 new AI supercomputers, enabling: Traditional computing + AI + Quantum to jointly serve national-level scientific research.
8. AI is Rewriting the Entire "Computing Stack": From Tool to "Digital Employee"
Past model: Programmers write code → runs on CPU + OS (e.g., Windows) → delivers a tool (Excel, Word, browser). These are "tools," requiring human operation.
Current model: Models trained on GPUs from data chopped into Tokens can themselves think, reason, and make decisions. They can not only use tools but can also perform many tasks directly for you.
From Jensen Huang's perspective: AI is no longer just a tool but a "digital labor force" capable of using tools itself. This means AI won't just impact the multi-trillion-dollar IT industry; it will enter the ~$100 trillion physical economy—reshaping manufacturing, healthcare, logistics, and finance.
Moreover, the most compute-intensive part isn't just training, but: Pre-training—Learning foundational knowledge. Post-training (fine-tuning/RLHF)—Learning problem-solving methods, reasoning approaches. Daily "thinking and inference"—Continuously understanding new problems, retrieving information, planning steps, and delivering answers. This entire pipeline combined is far more computationally demanding than many realize.
9. Omniverse DSX: Build and Perfect in the "Virtual Factory" Before Deploying in the Real World
NVIDIA Omniverse DSX is positioned as a "blueprint and simulator for designing and operating hyper-scale AI factories."
In simple terms: First, simulate the entire AI factory—building construction, data center layout, power distribution, cooling, rack arrangement—in a highly realistic 3D digital world. Collaborate with traditional engineering giants like Siemens, Schneider, Vertiv, and Bechtel to iterate repeatedly in simulation.
The benefits: "Test operation" before construction begins, identifying issues upfront, significantly shortening build cycles. After completion, this digital twin becomes the factory's "control panel": AI agents can simulate different operational strategies in the virtual space to see how to save energy, reduce costs, and increase output, then guide the real factory.
For a 1 Gigawatt (massive scale) AI factory, this optimization could yield tens of billions of dollars in annual economic benefits.
10. Open Source Models & Enterprise Ecosystem: Not Just Profiting Alone, but "Elevating the Entire Ecosystem"
Jensen Huang strongly emphasized: The U.S. must maintain leadership in open-source AI models, crucial for research, developers, and startups. NVIDIA itself has created and open-sourced many models (speech, reasoning, physical world understanding, biology, etc.) with massive download counts.
Simultaneously, they have deeply integrated their software and models: Into major clouds (AWS, Google Cloud, Azure, Oracle). Into various enterprise software suites: ServiceNow, SAP, Synopsys, Cadence, etc. Collaborated with CrowdStrike on security AI and Palantir on intelligent "data + decision" platforms.
The message is simple: Whether you're on the cloud, in a large enterprise, or using security/data platforms, the underlying compute layer powering it will likely be NVIDIA's.
11. Physical AI: Enabling AI to Move Beyond the Screen, into Factories, Operating Rooms, and Theme Parks
Definition of "Physical AI": AI that not only computes and converses but also understands real-world physical laws (gravity, friction, collision, causality) and can take action in reality.
NVIDIA believes achieving this requires coordination among "three computers": Training Supercomputer—Where the large model is "taught its capabilities." Digital Twin Simulator—Uses Omniverse to create a virtual world, providing a "safe training ground" for robots. Onboard Computer for Robots—Jetson Thor placed in robot dogs, humanoids, or autonomous vehicles for real-time control.
Practical collaboration scenarios include: With Foxconn, Siemens building digital factories—Train robots in virtual factories first, then deploy them on real production lines. With Figure, Agility Robotics on humanoid robots, warehouse robots. With Johnson & Johnson on surgical robots, practicing complex procedures repeatedly in a digital twin operating room. With Disney Research on the incredibly cute robot Blue, using the new physics engine Newton to teach it to walk, fall, and get back up in simulation.
12. Robots on Wheels: Autonomous Vehicles + the Uber Network
Among all robots, the ones closest to mass deployment are "robots on wheels"—RoboTaxis (autonomous ride-hailing vehicles).
NVIDIA offers a standardized solution: Drive Hyperion—Helps automakers define: Which cameras, radars, LiDARs, and computing units to install in the vehicle. Automakers build to this standard, allowing autonomous driving software companies to install their software directly for testing and commercialization.
They also partner with Uber: In the future, ride-hailing apps might offer not just human drivers but also a fleet of "autonomous vehicles" to hail. These vehicles would be built on unified platforms like Hyperion.
Simply put: NVIDIA aims to turn "autonomous vehicles" into a globally unified "robotic computing platform," integrated into super-networks like Uber.
13. Conclusion: Two "Century-Level Transformations" Converging
Jensen Huang's final summary: We are in an era where two massive transformations are converging—
From General-Purpose Computing → Accelerated Computing (CPU-centric → GPU-centric)
From Handwritten Software → Artificial Intelligence (models take center stage)
These two waves, combined with 6G, quantum, AI factories, Physical AI, robots, and autonomous driving, will lead to: The complete reshaping of the computer industry itself; The opportunity for the U.S., through "AI factories + manufacturing reshoring," to reclaim a leading position in the new wave of technological and industrial revolution.