Jensen Huang Unveils DSX POD AI Factory Design and OpenShell Agent Architecture at GTC Berlin

On the evening of October 21 Beijing time, Nvidia's GTC conference opened in Berlin, Germany. In his keynote, Nvidia founder and CEO Jensen Huang showcased a new generation of AI infrastructure, enterprise agents and robotics platforms, and devoted a large portion of his talk to industrial production and Physical AI.

Huang said Vera Rubin has entered full production, and Nvidia is expanding its computing platform from handling large language models to running complete agents. In his view, models handle perception, reasoning and planning, but for AI to actually complete tasks, it also needs an agent framework responsible for orchestration, enterprise knowledge, tools, and a secure runtime environment that limits its permissions.

Before the Berlin GTC, Huang visited France, the UK and Sweden, among other places. He revealed that over the past year Nvidia has invested nearly $10 billion in Europe's technology ecosystem and now works with more than 12,000 European partners.

On the delivery speed of AI data centers, Huang said a 1-gigawatt AI factory can cost about $50 billion to build, from land and power to computing, networking and cooling systems. For operators, how quickly equipment can be put into operation after construction directly determines when the investment starts generating revenue.

Nvidia that day unveiled the DSX POD AI factory design, in which Vera Rubin computer racks, network cabling and liquid cooling systems are integrated and tested at the factory in advance, then delivered as containers. Nvidia estimates that, provided site power and other supporting facilities are already in place, on-site deployment time can be cut from about 6 months to as fast as 14 days. Partners including Dell, HPE, Lenovo and Schneider Electric plan to launch related products in early 2027.

Shortening installation time in the server room does not mean all construction challenges are solved; power connections, site construction and upstream equipment supply remain independent constraints.

At a media briefing the previous day, Ian Buck, Nvidia's vice president of hyperscale and high-performance computing, also said in an interview with Jiemian News and other global media that hyperscale data centers still have economies of scale, and that the near-term opportunity for DSX POD is mainly to help hundreds-of-megawatts-scale projects deploy faster, not to fully replace traditional server rooms.

Image source: Nvidia

Huang also presented an energy-adjustment solution for AI factories. According to Nvidia, DSX MaxLPS can dynamically allocate power across different workloads, cutting electricity use by up to about 30% while maintaining the same token output. Closed-loop liquid cooling, combined with dry cooling in suitable climates, can also reduce water consumption from evaporative cooling.

Agents were the other main thread at this year's GTC Berlin. Huang believes traditional models mainly answer questions, while agents can call tools, read materials and repeatedly execute tasks. That means enterprises care not only about model capabilities, but also about which files it can access, whether it can call internal systems, and whether it can be stopped immediately if something goes wrong.

To that end, Nvidia that day released the OpenShell reference architecture, which places agents in an isolated environment constrained by permission policies and uses software running on BlueField-4 processors to independently monitor agent behavior and enforce security policies. Huang stressed that an agent may need different permissions each time it starts, so security controls cannot rely on the model alone.

Such concerns are not unique to Nvidia. Gartner predicts that by 2027, 40% of enterprises may reduce the permissions of autonomous agents or stop using them after real operational problems arise due to governance deficiencies.

Nvidia went further with the Agent Knowledge Foundation, which helps agents retrieve enterprise documents, link order and supplier data, and even predict the impact of possible supply disruptions.

The industrial robot market itself is also expanding. A report released by the International Federation of Robotics (IFR) in September shows that in 2025, more than 600,000 new industrial robots were installed in factories worldwide, up 11% year on year, with about 5 million in operation.

However, industrial robot installations do not directly reflect the progress of general-purpose robots. Huang that day also introduced Physical AI applications such as autonomous driving and adaptive radiotherapy. Compared with traditional automation equipment that executes tasks according to preset programs, such systems need to make judgments and take actions based on changes in the actual environment, placing higher demands on reliability and safety.

Deepu Talla, Nvidia's vice president of robotics and edge AI, also discussed deployment boundaries at the media briefing. He said tasks such as mobile navigation, inspection, and grasping, placing and inspecting rigid objects are beginning to show clearer application opportunities, but general-purpose robots involving soft materials, dexterous hands and complex operations still have a lot of work stuck in the research stage. Model accuracy, reliability and safety are prerequisites for large-scale deployment.

Regarding the gap between robot training and actual deployment, Huang proposed that Physical AI requires three computing systems working together, respectively for model training, physical simulation, and real-time operation of the robot itself. Rather than repeatedly testing directly in real environments, developers can first train and validate robot actions in a simulated environment, then deploy the models to physical devices.

Nvidia that day also announced updates to the Isaac open robotics development platform, adding a humanoid robot reference platform, datasets, a runtime framework and robot learning tools, covering training data generation, model training, simulation testing and deployment.

Nvidia has partnered with German robotics company Agile Robots to launch the Agile ONE SN humanoid robot reference platform, which uses the Isaac GR00T model and the Jetson Thor computing platform. The robot is 1.4 meters tall and weighs 40 kilograms, with 31 degrees of freedom in its body and 20 degrees of freedom in each hand. It is planned to be made available to researchers and developers in early 2027.

Huang Renxun believes that a unified software and hardware reference platform can reduce the time developers spend integrating sensors, controllers and simulation tools on their own.

Image source: Nvidia

However, moving from a research platform to actual production, the return on investment of humanoid robots remains to be verified.

Gartner predicted in January this year that by 2028, fewer than 20 companies will truly put humanoid robots into manufacturing and supply chain production operations. The institution pointed out that humanoid robots still face limitations in dexterous manipulation, battery life, system integration and maintenance costs, and in some warehousing tasks, wheeled multi-functional robots are actually more efficient.

As an "easter egg" during this year's keynote, the robots that took the stage this time were four robotic ducks, Microduck. The four robotic ducks walked onto the stage and, following Huang Renxun's instructions, walked, stood and sat down. Microduck was launched by Pollen Robotics, a subsidiary of Hugging Face. It is about 25 centimeters tall, uses an open-source design, and is priced at 399 US dollars. This robotic duck became a hit overseas as soon as it was released.

Image source: livestream screenshot

The training process of Microduck was demonstrated on site. Developers can build a simulation environment in Isaac Sim, train robot actions through the cloud, and then deploy the trained policies to the physical robotic duck. Nvidia also used this to demonstrate the robot development process from simulation training to physical execution.

When talking about humanoid robots, Huang Renxun said that if robots can quickly learn new tasks, then for small and medium-sized enterprises without dedicated robot development teams, in the future they will only need to show the robot the work to be completed, and the robot will be able to learn and adapt.

This article is from "Jiemian News", reporter: Xu Meihui, published by 36Kr with authorization.