十年前要两个人抬上楼的机器，今天被老黄单手举在了台上！刚刚，GTC柏林的舞台上，黄仁勋举起一台轻薄的笔记本 RTX Spark，内存128GB。2016年8月，他把刚发布的DGX-1送到OpenAI的旧金山办公室。那台机器的内存同样是128GB，却得两个人才抬得上楼。十年过去，AI性能涨了100万倍，同期摩尔定律只涨了约100倍。差出来的部分，靠的是「极致协同设计」，每一层软件、每一台计算机、每一种新算法，连同架构一起改。
- AI工厂装箱发货：DSX POD在工厂里装好、测完，场地水电就绪后，部署从约六个月缩到最短14天。
- Agent上了锁：OpenShell参考架构把规矩写在Agent之外，越界操作由BlueField-4毫秒级隔离。
- 机器鸭子上台：MicroDuck在RTX Spark上造出来，云端训练，自己走上了舞台。
- CUDA能打电话了：英伟达和诺基亚在T-Mobile网络上，打通了全球首个实网AI-RAN通话。
- 人形机器人有了参考机：英伟达和Agile Robots推出Agile ONE SN，预计2027年初上市。
RTX Spark的FP4 AI算力最高1 Petaflops，统一内存最高128GB。它把Blackwell RTX GPU和Grace CPU做成了一颗超级芯片，老黄在台上叫它N1X。这台电脑10月16日已经开售。过去几十年，电脑一直是工具，你敲下指令，它照着执行。现在，这个定义被彻底推翻了。在黄仁勋眼里，RTX Spark是你的「个人Agent」「个人助手」，甚至「数字伴侣」。你和它相处，更像和一个帮你干活的伙伴相处。要当好个人Agent，它得有模型、Harness、安全运行时OpenShell，还得有知识和工具。其中模型既可以在本地跑，也可以放到云上跑。工具这一项，英伟达有现成的家底。这台电脑能加速OpenGL、DirectX和全部CUDA，所以英伟达做过的所有软件，都能在上面跑。为了这台电脑，英伟达和微软秘密合作了四年，前后投入了数千个工程师年。整个PC行业也倾巢出动，台式机、笔记本、工作站，全线拥抱RTX Spark。未来的工作场景，是你把任务交给RTX Spark，它去调用软件，帮你把活干完。你不再是「使用」软件，你是在「雇佣」电脑。这几年，大家都在谈大模型。但在黄仁勋看来，单靠模型还不够，模型只是大脑，真正能干活的是Agent。Agent会思考，会推理，会迭代，会使用工具。它想得越多，消耗的token就越多，GPU的需求也跟着冲上了天。
- Harness： 编排任务，调动活跃记忆、工作记忆和长期记忆，激活工具。
- 沙箱： 把Harness隔离起来，和系统的其他部分分开。
这五样东西，各自对应不同的硬件。模型在Vera Rubin NVL72上思考，Harness在Vera MGX计算节点上跑，知识由Vera CMX存储节点提供，工具交给专跑CUDA-X库的Vera RTX计算节点。所以Vera Rubin驱动的，是整个Agent。Vera Rubin的硬件本身也变了。黄仁勋在台上拉出一个计算托盘，里面没有线缆，也没有电线，100%液冷。上一代Grace Blackwell组装一台要两个小时，线缆到处都是，Vera Rubin组装一台不到5分钟。至于Agent怎么搭，黄仁勋的答案很简单：「让另一个Agent帮你建。」后面那场机器鸭子的演示，就是由一台个人电脑上的Agent驱动的。Agent能自己动文件、上网、调工具，所以权限得压到最小，还要一直盯着。英伟达介绍OpenShell的片子，开场先自嘲了一把，说眼下能拦住Agent的，只有提示词里一句客客气气的「请别删东西」。它的思路，是把规矩写在Agent之外。每个Agent关进自己的沙箱，能读写哪些文件、能连哪些网络，细到每一次系统调用，都由沙箱外的策略说了算，密钥也不交到Agent手里。硬件上再加一道锁。Sentry跑在BlueField-4上，像一个不睡觉的哨兵，守在主机之外，死死盯住Agent的一举一动，一越界，毫秒级就会被隔离。旧时代的工厂，把煤炭变成电力。新时代的AI工厂，把电力变成智能。黄仁勋整场都在反复讲一个判断，AI是一种制造业，不只是软件。「想成为这个AI经济的一部分，你不能只做AI的消费者，还要做AI的生产者。」黄仁勋说。一座1吉瓦的AI工厂，把土地、电力、厂房、计算、网络、存储和散热都算上，大约要500亿美元。按一年400亿美元的收入算，晚投产三个月，就有约100亿美元的收入没了。为了让企业更快建厂，英伟达推出了 NVIDIA DSX POD。一个POD由7个液冷计算机架组成，机架之间用17公里长的Spectrum-X线缆连起来。整套设备在工厂里就完成集成、布线和测试，再装进可运输的集装箱，拉到数据中心。到了现场，用干冷器散热，800V直流电直接供到机架。只要场地的水电就绪，部署时间就能从约六个月缩到最短14天。背后的逻辑，是把装机搬进供应链。三年前，Grace Blackwell NVLink 72先把整机架挪进了工厂，可机架运到现场后，仍要再组装好几个月。DSX POD把搭建、点亮、验证和老化测试也全搬进工厂，出厂就能运行。在老黄看来，这是对全球供应链的一次彻底改造。电和水也省了。同样的token产出，DSX MaxLPS最多能少用30%的电；闭环液冷加干冷器不蒸发水，放大到2吉瓦规模，每天约能省下5400万升水，相当于18万户家庭一天的用水量。如果说大模型是AI的灵魂，那么Agent是AI的双手。压轴演示的主角，是一只机器鸭子MicroDuck。它是Hugging Face的机器人，由Pollen Robotics造出来。这场演示要看的，是怎么用Agent教会它走路。从搭训练场到训练、再到上真机，整套流程都跑在一台RTX Spark上，只有遇到大负载，才交给云端。第一步，开发者用文字描述训练场地和要学的技能，Hermes加Nemotron就在RTX Spark上规划好场地，列出需要的物体。接着，OpenAI的Astra调用FreeCAD建3D物体，用Blender贴材质，用Omniverse SimReady库补上物理属性，在Isaac Sim里把场景搭起来。然后，Astra再给MicroDuck配上关节、执行器和连杆，用Newton物理引擎模拟它怎么动。训练这一步在云端。Isaac Lab里数千只虚拟MicroDuck并行训练，摔一跤就是一课，策略一遍遍变好。训练好的策略再下载回RTX Spark，评估通过后，部署到真机上。随后，四只真实的MicroDuck走上了舞台。黄仁勋像个老父亲一样，先让它们走两步，叮嘱了一句「别绊倒」，又问它们「你们能坐下吗？」，鸭子们应声坐了下来。他又让它们站起来、睡个觉，最后喊它们翻个跟头，自己连说了好几遍「太棒了」。「你们在RTX Spark上诞生，在云端一堆英伟达超级计算机上训练，又在数字孪生里学会了本事。」黄仁勋对着鸭子说。这三步，对应的就是物理AI的三台计算机。第一台训练AI模型，第二台在数字孪生Omniverse里让AI学技能，不能刚编好就丢进现实世界乱试，第三台装在机器人身上。鸭子之外，老黄更看好人形机器人，在他看来，这会是下一个大产业。为了让开发者不用从零开始，英伟达和Agile Robots发布了参考人形机器人 Agile ONE SN。以前做机器人研究，软件、计算系统、传感器和机械系统都得自己拼。现在这些全装在一台机器上，搭一个测试平台，从几个星期缩到了几天。Agile ONE SN身高1.4米，体重40公斤，有31个自由度，预置Isaac GR00T，由Jetson Thor驱动，预计2027年初上市。在黄仁勋看来，机器人任务太分散，如果每个任务各配一套，行业就会被切碎，技术飞轮也转不起来。手机、PC和NVLink 72，都是靠标准化配置把量做起来，量再反过来推着技术加速。人形不一定在每个场景都最优，但能覆盖足够多的场景，所以很可能成为出货量最大的机器人形态。加上英伟达开放的数据集、Isaac仿真平台和GR00T模型，Agile ONE SN想当的，是全球机器人开发者的「标准教材」。前面说的Agent五要素里，模型和知识这两样，英伟达这次也都给了现成的。先说模型。大模型不再垄断在少数人手里，英伟达把各领域的开放模型全摆了出来。老黄也承认，这些模型，尤其是语言模型，不一定处在最前沿。可英伟达把权重、训练数据和复现脚本整套开放了，企业拿自己的数据做后训练，就能把它练到前沿水平。最终，你将得到一个绝对忠诚、绝对保密、只懂你业务的「超级智能」。再说知识。为了让Agent读懂企业数据，英伟达推出了 Agent Knowledge Foundation。不管是杂乱的文档，还是结构严谨的数据库，Nemotron都能解析、嵌入、重排序，新的Kumo模型还能直接从数据里做预测。「让Agent变成超级Agent的，是你给了它访问公司知识的权限。」老黄说。懂你公司知识的超级Agent，这才是企业真正需要的AI。
演讲最后，黄仁勋留下一句宣言：「每一家欧洲公司，都将成为AI公司。毫无疑问。」
一首「欧洲歌唱大赛」风格的歌，把本场的发布全编成了调侃，从OpenShell给Agent划的权限，到BlueField上的Sentry，再到机器人在仿真里练本事。「OpenShell守大门，没有权限赶紧滚。BlueField像施瓦辛格一样护卫沙箱。」「你无法拼写Fungible（可互换）如果没有Fun（乐趣）！」伴随着萨克斯的独奏，全场合唱《We are Europa》。
参考资料：
https://blogs.nvidia.com/gtc-berlin-2026-news/
编辑：大卫 摩西
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Ten years ago, the machine that had to be carried upstairs was put on stage today by the yellow one! Just now, on the GTC Berlin stage, Wong In-hoon raised a light notebook RTX Spark, RAM 128GB. In August 2016, he sent the DGX-1 of just publish to the San Francisco office in Openai. The memory of the machine was also 128 GB, but it had to be carried upstairs by two people. Over the past decade, AI performance has increased 1 million times, while Moore's law has increased only about 100 times over the same period. The missing part is based on "extremely synergistic design", with each layer of software, every computer, every new algorithm, together with structures.
- Delivery by container at the AI plant: DSX POD is installed in the plant and measured, the site is ready for hydropower and deployment has been reduced from about six months to a minimum of 14 days.
- Agent is locked: the OpenShell reference structure puts the rules outside Agent and cross-border operations are isolated from Bluefield-4 milliseconds.
- Machine ducks on stage: MicroDuck was built on RTX Spark, cloud-trained, himself on stage.
- CUDA was able to make a telephone call: In Weida and Nokia, on the T-Mobile network, reached the world's first live-networked AI-RAN.
- Human robots have a reference machine: Young Weida and Agile Robots launched Agile ONE SN, which is expected to be available early in 2027.
RTX Spark has the highest FP4 AI calculator of 1 Petaflops and a single memory of 128GB. It turned Blackwell RTX GPU and Grace CPU into a super chip called N1X on stage. The computer was sold on October 16th. The computer has been a tool for the past few decades, and you knock down the command, and it's working. This definition has now been completely overturned. RTX Spark is your personal "Agent" "personal assistant" and even "digital partner" in Wong In-hoon's eyes. You're with it, more like a partner who works for you. To be personal, Agent must have models, Harness, OpenShell when running safely, knowledge and tools. They can run locally or on clouds. The tool, we have a home. This computer accelerates OpenGL, DirectX and all CUDA, so all of the software that Yvette did can run up there. For this computer, Yin Weida and Microsoft worked secretly for four years and spent thousands of years in engineering. The entire PC industry is also penetrating, desktops, notebooks, workstations, full-line embraces RTX Spark. The future work scene is that you give the mission to RTX Spark, which calls the software to help you finish the job. You're no longer using software. You're hiring computers. All these years, everybody's talking about LLM. But in Huang In-hoon's view, models are not enough, models are just brains, and it's Agent who can really work. Agent will think, reason, overlap, use tools. The more it thinks, the more token it consumes, the more the GPU needs.
- Harness: Organize tasks, mobilize active memory, work memory and long-term memory, activate tools.
- Sandbox: Segregate Harness from the rest of the system.
These five things correspond to different hardware. The model reflects on Vera Rubin NVL72, Harness runs on the Vera MGX computing node, knowledge is provided by the Vera CMX storage node, and tools are given to the Vera RTX calculation node of the CUDA-X library. So Vera Rubin drives the whole Agent. Vera Rubin's hardware itself has changed. Wong In-hoon pulled out a computing tray with no cable or wire, 100 per cent liquid cold. The last generation of Grace Blackwell assembled one for two hours, cable was everywhere, and Vera Rubin assembled one for less than five minutes. As for how Agent works, Huang In-hoon's answer is simple: "Let another Agent build it for you." The demonstration behind the machine duck was driven by Agent on a personal computer. Agent was able to move his own files, access the Internet, adjust tools, so permission had to be pressed to the minimum and kept staring. Young Weidar introduced OpenShell's film, and started by mocking himself, saying that the only one who can stop Agent right now is the polite phrase "Please don't delete anything" in prompt. The idea is to write the rules outside Agent. Every one of Agent's sandboxes, which documents to read and write, which networks to access, and which systems to call, is determined by the strategy outside the sandbox, and the key is not delivered to Agent. Add another lock to the hardware. Suntry runs on Bluefield-4, like a non-sleeping sentry, outside the mainframe, staring at Agent's every move and crossing the border, the millisecond level is quarantined. Old-age factories turned coal into electricity. New age AI factory, turning power into intelligence. Wong In-hoon has been saying a lot of things about it: AI is a manufacturing industry, not just software. "To be part of the AI economy, you can't just be an AI consumer, you can be an AI producer. Wong In-hoon said that an AI factory, one giva, counted land, electricity, plant, computing, network, storage and heat dispersion, at about $50 billion. At $40 billion a year, three months later, about $10 billion was lost. In order to speed up the construction of the plant, Inverda launched the NVIDIA DSX POD. A POD consists of seven liquid cold computer racks, which are linked by a 17-km Spectrum-X cable. The complete set of equipment is integrated, wired and tested in the plant, then loaded into transportable containers and pulled to the data centre. Upon arrival, dry chillers were used to disperse the heat and 800V direct currents were supplied directly to the rack. The deployment time could be reduced from approximately six months to a minimum of 14 days as long as the site ' s utilities were ready. The logic behind this is to move the rigs into the supply chain. Three years ago, Grace Blackwell NVLink 72 moved the whole rack into the factory, which was to be assembled for several months after it arrived at the site. The DSX POD has also moved the set-up, light-up, validation and ageing tests to the plant, where they can be operated. In the old yellow perspective, this is a complete overhaul of the global supply chain. Electricity and water were saved. The same token output, DSX MaxLPS, has up to 30 per cent power reduction; closed-cycling liquid chillers and dry chillers do not evaporate, zoom in to 2 GW, saving approximately 54 million litres of water a day, equivalent to 180,000 households. If LLM is AI's soul, Agent is AI's hands. The main character of the axle presentation is a machine duck, MicroDuck. It's H. Ginger Face's robot, built by Pollen Robotics. This demonstration is about how Agent teaches it to walk. From the training ground to the training and then to the real machine, the entire process is running on a RTX Spark, which is handed over to the cloud only when it meets a heavy load. As a first step, the developers describe the training site and the skills to be learned in writing, Hermes plus Nemotron plans the site on RTX Spark, listing the required objects. The Astra of Open AI then called FreeCAD to build 3D objects, used Blender to post the material, added physical properties to the Omniverse SimReady library, and set up the scene in Isaac Sim. Then Astra gives MicroDuck a joint, an implementer and a pole to simulate how it moves with a Newton physical engine. Train this step in the clouds. There are thousands of virtual MicroDucks in Isaac Labs who are trained in parallel, falling is a lesson, and the strategy has changed over and over. A well-trained strategy was downloaded back to RTX Spark, after which it was deployed to the real machine. Then four real MicroDucks went to the stage. Wong In-hoon, like an old father, took them two steps, said, "Don't trip" and asked them, "Will you sit down?" And the ducks sat down. And he put them up and slept, and he called them back and forth, and he said, "It's great." "You were born on RTX Spark, you trained on a bunch of supercomputers in the clouds, and you learned your skills in digital twins. Wong In-hoon says to the ducks. Three steps, three computers for physics AI. The first training AI model, the second in the digital twin, Omniverse, allowed AI to learn skills that could not just be programmed into the real world and the third in robots. Aside from ducks, yellow looks more like a good-looking robot, which, in his opinion, is the next big industry. In order to avoid the need for developers to start from scratch, Iveda and Agile RobotsZ94Z have a reference human robot, Agile One SN. Previously, robotic research, software, computing systems, sensors and mechanical systems had to be crafted. Now these are all mounted on a machine, with a test platform, which has shrunk from weeks to days. Agile ONE SN, which is 1.4 metres tall and weighs 40 kg, has 31 degrees of freedom, pre-positions Isaac GR00T, driven by Jason Thor, and is expected to be listed early in 2027. From Wong In-hoon ' s point of view, robotic missions are too dispersed, and if each task is combined, the industry will be shredded and the technology wheel will not be able to move. Cell phones, PCs and NVLink 72 are done with standardized configurations, which in turn push technology acceleration. The human form is not necessarily the best in every scene, but it can cover enough of it, so it is likely to be the robotic form with the largest volume of output. Together with the Open Data Set, the Isaac Simulation Platform and the GR00T model, Agile ONE SN is a "standard textbook" for global robot developers. Of the five above-mentioned Agent elements, both models and knowledge are now available at the same time. Let's start with the model. LLM is no longer monopolized by a few, and Ying Weidar has put out all the open models in all fields. It is also acknowledged that these models, especially language models, are not necessarily at the forefront. But Yin Weidar opened up weight, training data and replicas. I can get it to the front level. In the end, you'll get a "super smart" with absolute loyalty, absolute secrecy, only knowledge of your business. Say knowledge. In an effort to get Agent to read business data, Young Waida launched Agent Knowledge Foundation. Whether it is a messy document or a well-structured database, Nemotron can decipher, embed, reorder, and the new Kumo model can make projections directly from the data. "Agent became superAgent, and you gave it access to the company's permission. Yellow says. The super Agent who knows your company's knowledge, that's what companies really need.
At the end of the speech, Hoang In-hoon left a message: "Every European company will be AI." No doubt about it.
A song like the "European Singing Contest" made up the whole of publish from OpenShell to Agent's permission to Suntry on Bluefield, where robots practice in simulations. "Open Shell guards the gate without permission. Bluefield guards sandboxes like Schwarzenegger. "You can't spell Fungible without Fun!" along with Sax's solo, "We are Europa."
References:
https://blogs.nvidia.com/gtc-berlin-2026-news/
Editor: David Moses
Auto-fetched and translated by AI Atlas. All rights belong to the original author.