Tech • AI • Robotics • Game
NVIDIA GTC Berlin 2026 Keynote With CEO Jensen Huang
9.4/10TL;DR
NVIDIA used its Berlin event to frame AI as an industrial production system, unveil new data center and PC products, and argue that Europe is becoming a major base for AI infrastructure, software and robotics.
KEY POINTS
AI as a manufacturing industry
Jensen Huang described AI not as conventional software but as a production stack that turns computing power into tokens, or units of digital intelligence. He presented a five-layer model: energy, advanced computing, AI factories, models, and applications. The central argument was that countries and companies should not only consume AI services, but also produce AI locally as part of a broader AI economy.
Europe positioned as a strategic AI market
NVIDIA said it has worked in Europe for 33 years and invested nearly $10 billion across the region in the last year through startups and partnerships. The company said it now has more than 12,000 partnerships in Europe and recently announced investments in 14 startups. Huang highlighted customers and partners including L’Oréal, Revolut, Spotify, Merck, Vodafone, Zara, Porsche, Nestlé and HSBC as evidence that AI adoption is spreading far beyond large cloud providers.
New AI factory economics
The company argued that AI infrastructure now resembles heavy industry, with a 1-gigawatt AI data center costing roughly $50 billion when land, power, computing, cooling and networking are included. Such facilities are designed to produce tokens at industrial scale, with expected annual revenue measured in the tens of billions of dollars. Huang said delays are increasingly expensive, making faster deployment a major commercial priority.
DSX Pod aims to cut deployment times
NVIDIA introduced DSX Pod, a factory-integrated AI infrastructure system built around Vera Rubin NVL72. Instead of assembling supercomputers largely on site, the company plans to pre-integrate, cable and test systems in the supply chain, then ship them in transportable units. Huang said that could reduce the time from delivery to token production from four to five months to about one month, lowering idle capital and speeding revenue generation.
Vera Rubin and network upgrades
The company said Vera Rubin is now in full production and represents a major step beyond Blackwell. A single system links up to 72 GPUs through NVLink, with each GPU rated at 20 petaflops, effectively creating one giant computing domain. NVIDIA also showcased a new Spectrum-X Ethernet switch with co-packaged optics, a design intended to improve bandwidth and efficiency by bringing optical interconnects directly onto the chip.
Performance claims far beyond Moore’s law
Huang said NVIDIA has improved AI performance by 1 million times over the last 10 years, from the Pascal era through Volta, Ampere, Hopper, Blackwell and now Vera Rubin. He contrasted that with the roughly 100x improvement expected over a decade under a classic interpretation of Moore’s law. The gain, he said, came from “extreme co-design” across chips, systems, networking, software and algorithms.
Security push for autonomous agents
A major theme was the rise of AI agents that can reason, use tools, access files and interact with networks. NVIDIA introduced Open Shell, a secure runtime meant to constrain agents through policy-based controls over system calls, file access and network connections. The design pairs with hardware monitoring to enforce least-privilege access, reflecting growing concern that autonomous software systems need browser-like sandboxing before broad enterprise deployment.
Agent architecture extends beyond language models
NVIDIA described an agent as a five-part system: the model, a harness that orchestrates work, a secure sandbox, access to knowledge, and access to tools. Huang argued this architecture requires more than a single model server, creating demand for multiple computing layers optimized for reasoning, memory, data processing and accelerated tool use. That framing supports NVIDIA’s broader pitch that enterprises need full-stack AI infrastructure rather than isolated chips.
RTX Spark brings AI agents to PCs
The company also unveiled RTX Spark, a new PC line developed with Microsoft over four years and powered by the new N1X chip. NVIDIA said the system can deliver 1 petaflop and 128 GB of memory in a compact desktop form, while a higher-end DGX Station version for Windows reaches 20 petaflops and 754 GB of memory. The products are aimed at running AI agents and accelerated software locally, not just in the cloud.
Robotics and physical AI as the next platform
Huang said robotics could become the next high-volume computing platform, especially in Europe, where manufacturing and mechatronics remain core strengths. NVIDIA said physical AI needs three computing environments: one for model training, one digital twin for simulation through Omniverse, and one onboard robotics computer. It announced the NVIDIA Agile One reference robot with Agile Robots, due in the first quarter of next year, as a pre-integrated starting point for robotics developers.
CONCLUSION
NVIDIA’s message was that AI is moving from software feature to industrial infrastructure, with data centers, secure agents and robots all part of the same stack. The company is betting that Europe will be a major builder, not just a buyer, of that next computing economy.
Ask a question