在谈主义之前，先把工业搞上去。
GTC 柏林开幕前的两周，半导体板块走得并不顺，费城半导体指数出现了 2008 年以来最长的一轮连跌。市场在反复算同一笔账，巨额的 GPU 投入，什么时候能变成实打实的回报。
10 月 21 日，在柏林 Tempodrom 的舞台上，黄仁勋没有发布新芯片。早已发布的 Vera Rubin 平台，这次只是确认「已全面量产」。他把更多时间留给了三个更落地的问题：算力中心怎么建得更快，越来越能干的智能体怎么管住，物理 AI，怎么走进欧洲的工厂和马路。
用他自己的话说，「我们正处于人类历史上最大规模的基础设施建设之中」，需求不缺，缺的是产能和落地速度。
在英伟达 GTC 柏林大会之前，黄仁勋和德国总理默茨的对谈走红社交平台，当下 AI 基建绝对王者告诉后者，面对汹涌而至的全新生产力革命，「紧箍咒」可能并不是最优解。
把 AI 工厂装进集装箱
过去两年，建一座 AI 数据中心更像一场土木工程。服务器、网络、供电和液冷要在现场逐一安装、布线、调试，一个大型集群从设备进场到真正点亮，拖上大半年是常态。而模型公司的迭代是按周算的，晚一个月拿到算力，可能就晚了一个版本。
英伟达在柏林给出的答案是 DSX POD。大屏上的标语很直白，「集装箱就是 AI 工厂」。
它是一套完整 Vera Rubin AI 工厂的模块化参考设计。一套 DSX POD 由 7 个液冷计算机架组成，内部约 17 公里线缆互连，采用 800V 直流供电，计算、网络、供电和液冷都在工厂里集成、测试完毕，再装进集装箱运到现场。戴尔、HPE、联想、施耐德电气、维谛，以及鸿海、广达、纬创等厂商参与制造，预计 2027 年初开始交付。
按照官方的说法，在现场电力、冷却和网络已经就绪的前提下，部署时间可以从约 6 个月缩短到最快 14 天。
只看建设速度，其实低估了这件事。黄仁勋把 AI 经济比作五层蛋糕，最底下一层是能源。当算力规模越做越大，真正卡脖子的已经不只是芯片，而是电。
所以和 DSX POD 一起发布的，还有两套用电工具。DSX MaxLPS 在 GPU 之间动态分配功率，官方称在产出同样多 Token 的情况下，功耗最多能降低 30%，换个角度说，同样的电力额度里能塞进更多 GPU。DSX Flex 则让数据中心和电网联动，按电价和供电状况调整负载，在模拟电网事件的早期测试里，一个 Blackwell Ultra 集群的用电负荷最多降了 40%。冷却改用闭环液冷加室外干式冷却器，减少冷却塔的蒸发耗水。
英伟达投资的欧洲数据中心开发商 Nscale 给了一组早期数据，在 GB300 NVL72 上，同样的电力预算下多部署了 37% 的 GPU，总推理吞吐提升近 50%。
这套逻辑的落点很清楚，衡量 AI 工厂的指标，正在从「有多少卡」变成「每度电产出多少 Token」。
当然也有官方没说的部分。14 天的前提是电力和场地已经到位，而从公开信息来看，欧洲不少地区的电网接入本身就要排很久。DSX POD 的价格、总拥有成本也还没有公布，集装箱方案能帮客户省下多少钱，要等首批项目交付后才见分晓。
给智能体装上硬件刹车
如果说 DSX POD 解决的是「建得快」，OpenShell 企业参考架构要解决的是「用得放心」。
今年的智能体已经不只待在对话框里，它们开始读写文件、调用工具、访问企业系统。权限越大，风险越大，近期 AI 圈就出现过智能体擅自关闭安全监控的事件。
黄仁勋在台上说得很直接，「最重要的是确保智能体是安全的」。智能体的权限要降到最低，并持续监控，而且「每次启动智能体，权限都不一样」，所以管控也得跟着动态调整。他把 OpenShell 称为英伟达做过最重要的事情之一。
OpenShell 本身在 6 月的 GTC 就已推出，是一个开源、按策略控制的运行时，用内核级沙箱把智能体和它能影响的系统隔开。这次发布的，是在英伟达全栈硬件上怎么部署它的企业级方案。
分工是这样的，Vera Rubin GPU 负责模型推理，Vera CPU 承载智能体的运行框架和工具调用，OpenShell 划定隔离和权限边界，BlueField-4 DPU 上则跑着英伟达的 Sentry 安全框架。黄仁勋的概括是，智能体在 Rubin 上思考，在 Vera 上行动，在 BlueField 上被管住。
这套设计的意义在于，安全边界不再只依赖模型「自觉」或软件层的护栏。模型遇到提示词注入时，软件防线可能失守，而跑在 DPU 上的管控独立于智能体本身。英伟达称，智能体一旦越界，Sentry 能在毫秒级把它隔离并叫停，这一点还有待第三方验证。
合作名单相当长，Anthropic、OpenAI、微软、谷歌云、AWS、SAP、ServiceNow、CrowdStrike、Palo Alto、摩根大通都在大屏上。ServiceNow 已经可以用自家治理平台制定权限、由 OpenShell 执行，并能快速撤销智能体的访问权。不过多数安全厂商仍处在集成阶段，一些银行还在评估。
另一面也要看到，完整的参考架构建立在英伟达自家的 GPU、CPU 和 DPU 之上。对企业来说，获得硬件级安全的同时，也意味着和英伟达全栈绑得更紧。
物理 AI 在欧洲
把 GTC 放在柏林，很大程度上是冲着欧洲的工业底子来的。黄仁勋这次说，「Robotaxi 是机器人和物理 AI 第一个大众市场应用。」
最具体的一单来自欧洲出行平台 Bolt。它选用英伟达的 Alpamayo 技术栈开发 Robotaxi，计划获批后 2027 年先在马德里采集道路数据，目标是在欧洲部署至少 2.5 万辆基于 Hyperion 平台、具备 L4 能力的 Lucid 车辆，德国是重点市场之一。沃尔沃也宣布采用 Hyperion 和 Alpamayo，黄仁勋还提到，奔驰、大众商用车和捷豹路虎的下一代车型将兼容 Hyperion。
车端的新模型是 Alpamayo 2 Edge，20 亿参数，针对 DRIVE AGX Thor 芯片优化，每秒运行 10 次，负责理解路况和规划运动。配套的 Halos SAFE 把仿真、软硬件测试、道路测试和车队监控串起来，帮车企积累可追溯的安全证据，这恰恰是自动驾驶在欧洲拿牌照绕不开的一环。
人形机器人这边，德国思灵机器人推出了 Agile ONE SN 参考平台。它高 1.4 米、重 40 千克，本体 31 个自由度，两只五指灵巧手各有 20 个自由度，由 Jetson Thor 驱动，跑英伟达的 Isaac GR00T 模型，2027 年初推出。Agility Robotics、Figure、库卡等公司在同一套平台上开发。
英伟达对这个平台的定位相当克制。它承认人形机器人的开发门槛依然很高，目前基本是资金雄厚的公司才玩得转。Agile ONE SN 加上新的 Isaac OS、Isaac Lab 3.0 和仿真资产库，目标是把研究人员搭建和配置的时间从几周缩到几天，但这并不等于机器人已经能在产线上稳定干活。
工业现场的演示也是同样的思路。西门子埃尔朗根工厂让智能体先在数字孪生里验证新品投产方案，达索系统展示了智能体把大众高尔夫的车门设计改到途观上，仿真发现振动问题后继续迭代，最后交给工程师拍板。物理 AI 在欧洲的切入点很务实，先进入最耗工程师时间的仿真和测试环节，而不是急着替换产线。
演讲临近结束，四只 Microduck 小机器鸭走上了舞台。它们听指令坐下、站起、趴下「睡午觉」，有一只没跟上节奏，绊了一下又自己爬起来，黄仁勋连声招呼它「集中精神，过来」，全场都笑了。他对这几只小鸭子说，它们在 RTX Spark 上诞生，在云端的英伟达超级计算机上训练，又在数字孪生里学会了技能。
这几只摇摇晃晃的机器鸭，恰好是这场演讲的缩影。AI 从芯片走到机房，从对话框走到工厂和马路，每往前一步，都需要有人把电、安全和工程这些不那么性感的问题，一件件解决掉。
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。
Let's start with industry before we talk about it.
Two weeks before the opening of the GTC Berlin, the semiconductor plate was not going well, and the Philadelphia semiconductor index showed the longest consecutive drop since 2008. The market is double-counting, with huge GPU inputs, and when will it turn into real returns.
On 21 October, on the stage of Berlin Tempodrom, Hoang In-hoon did not have a new publish chip. The Vera Rubin platform, already in publish, was only confirmed as "full production". He left more time for three more landing questions: how to build a computer center faster, how to manage the increasingly dry AI agent, physics AI, how to walk into factories and roads in Europe.
In his own words, "We are in the process of building the largest infrastructure in human history" and there is no shortage of demand, but of capacity and speed of landing.
Before the Yin Weida GTC Berlin Congress, Huang In-hoon and German Chancellor Metz spoke to the Red Social Platform, where the absolute king of AI capital told the latter that the "tight spell" might not be the best solution in the face of a new productivity revolution.
Load the AI factory in a container.
In the past two years, an AI data centre has become more like a civil engineering project. Servers, networks, electricity and liquid cooling are to be installed, wired, debugged on site, and a large cluster is normal for much of the year, from equipment to real light. The model company's anecdotal is calculated on a week-by-week basis, with one month's credit, and one version may be late.
The answer given by Yin Weidar in Berlin is DSX POD. The slogan on the screen is plain, "The container is the AI factory."
It is a complete modular reference design for the Vera Rubin AI plant. A set of DSX PODs consisting of seven liquid cold computer racks, with approximately 17 km of internal cable interconnection, with 800 V current power supply, and computing, network, power supply and liquid cooling are integrated, tested and then loaded into the plant for transport to the site. Dell, HPE, Union, Schneider Electric, Wittung, and Honghai, Guangda, Latitude, etc. were involved in manufacturing and delivery is expected to begin in early 2027.
According to official accounts, the deployment time could be reduced from about 6 months to 14 days at maximum, provided that electricity, cooling and the network were in place on the ground.
It's an underestimation of the speed of construction. Wong In-hoon compares the AI economy to five-storey cakes, the bottom one being energy. When the calculus is bigger and bigger, it's not just chips, it's electricity.
So with DSX Pod, publish, there are two power sets. DSX MaxLPS dynamically allocates power between GPUs, officially claiming a 30% reduction in power consumption in the same output as Token, in other words, more GPUs can be inserted into the same power quota. DSX Flex allows data centres and grid connections to adjust loads to the price and state of power supply. In the early testing of simulated grid events, the load of a Blackwell Ultra cluster was reduced by up to 40 per cent. Refrigeration has been replaced with closed-cycling liquid cooler plus outdoor dry cooler to reduce the evaporation of cooling towers.
Nscale, a European data centre developer, invested in Weida, gave an early set of data, and on GB300 NVL72, 37% more GPUs were deployed under the same power budget, and the total reasoning was up by almost 50%.
This logic is clear, and the measure of the AI plant is moving from "how many cards" to "how many tokens per hour."
There are, of course, parts that are not officially mentioned. 14 Days are premised on the availability of electricity and sites, while, in terms of public information, grid access in many parts of Europe itself will be long. The price and total cost of ownership of DSX POD has not yet been disclosed, and the container programme will save the client how much it will save until the first project is delivered.
Put the hardware brake on AI agent.
If the DSX POD solves "quick", the OpenShell business reference architecture addresses "no need to worry."
This year ' s AI agent is not just in the dialogue box, they start reading and writing files, calling tools, accessing enterprise systems. The greater the permission, the greater the risk, the more recent incidents in the AI Circle of AI agent unauthorized security surveillance.
Huang In-hoon was very direct on stage, "first and foremost to make sure that AI agent is safe". As AI agent's permission has to be kept to a minimum and kept under constant surveillance, and "it's different every time AI agent, permission starts," controls have to be adjusted to the dynamics. He called OpenShell one of the most important things that I've ever done.
OpenShell itself was launched in June as an open-source, tactically controlled operation to separate AI agent from the system that it can influence by using a kernel. This time, publish is an enterprise-level programme on how to deploy its entire hardware in Weida.
The division of labour is as follows: Vera Rubin GPU is responsible for model reasoning, Vera CPU carries the operating framework of AI agent and tool call, OpenShell delineates the boundary of isolation and permission, and Bluefield-4 DPU runs the Suntry security framework of Britain. Huang In-hoon summed up AI agent as thinking on Rubin, acting on Vera, being held on Bluefield.
The significance of this design is that a secure boundary no longer relies solely on a model of " self-consciousness " or software layers. When the model is injected into prompt, the software line may fail, while running controls on the DPU are independent of AI agent itself. According to Ying Weida, once AI agent has crossed the border, Centry can isolate it in milliseconds and stop it, which is yet to be verified by third parties.
The list of cooperation is quite long, and the screens are on Anthropic, Open AI, Microsoft, Google Clouds, AWS, SAP, Servicenow, CrowdStrike, Palo Alto and Morgan Chase. Servicenow has been able to develop permission from its own governance platform, implemented by OpenShell, and quickly revokes AI agent’s access rights. However, most security manufacturers are still in the integration phase and some banks are still assessing them.
On the other hand, the complete frame of reference is based on the GPU, CPU and DPU of Weidar. For enterprises, obtaining hardware-grade security also means being tied more closely to the whole of Britain.
Physical AI in Europe
Putting GTC in Berlin, it's coming to Europe's industrial base. Wong In-hoon said this time, "Robotaxi is the first mass market application for robots and physics AI."
The most specific one comes from Europe's travel platform Bolt. It selected the Alpamayo technology warehouse in Weida for the development of Robotaxi, which was planned to collect road data in Madrid in 2027, with the goal of deploying at least 25,000 Lucid vehicles with L4 capabilities in Europe, one of the priority markets. Volvo also announced the adoption of Hyperion and Alpamayo, and Huang In-hoon also mentioned that the next generation of Mercedes, Mass Commercial Car and Jaguar Rover models would be compatible with Hyperion.
The new model for the end is Alpamayo 2 Edge, a 2 billion parameter that optimizes the DRIVE AGX Thor chip 10 times a second and is responsible for understanding the road and planning the movement. The accompanying Halos SAFE, which combines simulations, software and hardware tests, road tests and convoy surveillance to build up traceable security evidence for cars, is exactly the link where automatic driving can't get a license in Europe.
On the side of the human robot, the German Sing robot launched the Agile ONE SN reference platform. It is 1.4 metres tall, 40 kg heavy, 31 degrees of freedom in the body, 20 degrees of freedom in each of the five fingers, driven by Jason Thor, and launched in early 2027 by the Isaac GR00T model running in Weida. Companies such as Agility Robotics, Figure and Cooka developed on the same platform.
Young Weidar has exercised considerable restraint in positioning this platform. It recognizes that the threshold for the development of human robots remains high, and that it is largely financially funded companies that are now turning around. Agile ONE SN, with new Isaac OS, Isaac Lab 3.0 and a virtual repository, aims to reduce the duration of the build-up and configuration of researchers from a few weeks to a few days, but this does not mean that robots are already working steadily on the production line.
The same is true of industry field demonstrations. The Siemens El Langen factory allowed AI agent to validate the new product's commissioning programme in a digital twin, and the Dasso system showed AI agent changing the popular Golf door design to a road view, i.e. to discover the vibration problem and to continue to evolve and then hand it over to the engineer. Physical AI ' s entry point in Europe is practical, starting with the most time-consuming simulation and testing of engineers, rather than rushing to replace production lines.
The speech is nearing its end, and four Microduck little machine ducks are on stage. They sat down on instructions, stood up, lay down for their naps, one of them missed the rhythm, tripped and climbed up on their own, and Wong said, "Concentrate, come." And he said to these little ducks, they were born on RTX Spark, trained on the clouded British supercomputer, and learned their skills in digital twins.
These rocking machine ducks are just a microcosm of the speech. AI goes from chip to machine room, from dialogue boxes to factories and roads, and every step forward requires someone to solve less sexy problems of electricity, safety and engineering in one piece.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.