Nvidia CEO Jensen Huang unveiled AI infrastructure and robots at GTC Berlin, emphasizing Europe's role in AI production and showcasing cutting-edge tech.
Nvidia CEO outlines AI strategy at GTC Berlin
At the GTC Berlin conference on Wednesday, October 21, Nvidia CEO Jensen Huang presented a condensed version of the company's approach to artificial intelligence, emphasizing heavy infrastructure, geopolitics, and a strong sense of spectacle. Huang, who had recently completed a tour of Paris, London, Stockholm, and Dublin, highlighted the need for Europe to not only consume AI but also produce it. Nvidia has already committed nearly 10 billion dollars (approximately 8.9 billion euros) to the tech ecosystem in Europe.New AI data center module introduced
Huang addressed the issue of infrastructure, introducing the DSX POD, a prefabricated AI data center module that can be operational in 14 days once placed on a site already connected to electricity and cooling circuits. The DSX POD includes seven Vera Rubin NVL72 bays, each containing 72 GPUs, as well as all the network equipment, management electronics, and liquid cooling.Cost estimates for DSX POD
The DSX POD is expected to be costly, with each NVL72 rack estimated to cost between 5 and 7 million dollars, leading to a total cost for a single container ranging between 35 and 49 million dollars (approximately 30 to 42 million euros).New AI agent regulation framework
In parallel, Nvidia is addressing the regulation of AI agents through a new reference architecture called OpenShell, based on the open-source software of the same name. This tool confines AI agents in a "sandbox," a strictly isolated virtual space to prevent destructive actions or data leaks.Hardware stack for secure AI execution
Nvidia recommends combining its Vera Rubin GPUs for model reasoning, its Vera processors for agent execution, and its programmable BlueField-4 network cards to enforce security rules on the fly. This is not a new product sold separately but a turnkey method to encourage the adoption of the entire hardware stack.Microduck robots demonstrate AI capabilities
To illustrate the practical applications of these infrastructures and agents, four Microduck robots, small bipedal robots shaped like ducks, were showcased on stage. These robots, designed by Pollen Robotics, a Bordeaux-based company acquired by Hugging Face in 2025, are available for pre-order at a price of 399 dollars before tax (approximately 355 euros) with a delivery promised before Christmas.AI-generated walking path for robots
The robots demonstrated how a local AI agent running on an RTX Spark PC generated a walking path and modeled the small duck in the house simulator Isaac Sim. The intensive training passed through Nvidia's cloud servers before being directly injected into the real Microduck present in Berlin.Humanoid robot for AI research
For industrial research, Nvidia unveiled Agile ONE SN, a reference humanoid robot for researchers expected to be available next year. This robot has 31 degrees of freedom to test and train the next generation of physical artificial intelligences.