NVIDIA Showcases Physical AI Hardware at GTC Berlin: DSX PODs, Humanoid Robots, and Container-Based AI Factories

NVIDIA unveiled a comprehensive physical AI hardware stack at its GTC Berlin conference, featuring containerized AI factory systems, humanoid robotics, and accelerated computing platforms designed to shrink AI deployment timelines and scale production. The company demonstrated its vision of AI as a manufacturing industry through the DSX POD reference design, a fully integrated server container delivering 25 exaFLOPs of AI inference and capable of deployment in just 14 days—reducing traditional on-site setup from six months. The system integrates seven Vera Rubin NVL72 racks per container, with integrated networking, cooling, and management hardware. NVIDIA also highlighted RTX Spark, a portable Windows laptop featuring RTX graphics that delivers 1 petaflop of performance with 128GB memory—equivalent to the 2016 DGX-1 supercomputer but small enough to hold in one hand. On the robotics front, NVIDIA showcased multiple physical systems. The Agile ONE SN humanoid robot, developed with Agile Robots, stands 1.4 meters tall with 31 degrees of freedom and can lift up to 3 kilograms in each arm, designed to reduce researcher setup time from weeks to days ahead of its 2027 launch. During the keynote, four Microduck bipedal robots demonstrated on stage performed tasks including sitting, standing, and tumbling—trained entirely using RTX Spark and cloud resources. The company also introduced the OpenShell reference architecture, combining Vera Rubin hardware with OpenShell runtime software and BlueField-4 DPU security enforcement to establish best practices for building secure agentic AI systems. The announcements reflect NVIDIA's strategy to address supply chain constraints. The company outlined a three-year roadmap to expand semiconductor production capacity: a 2.5× increase in substrates, 4× growth in advanced packaging interposer wafers by end of 2028, 2× expansion in testing capacity by 2027, and a 15× to 30× increase in co-packaged optics capacity. Developer sentiment toward these announcements has been mixed. While the DSX POD's rapid deployment model and the RTX Spark's portability address long-standing pain points in AI infrastructure buildout, some observers note that DSX POD availability starting in early 2027 means customers must still wait for production ramp from partners including Gigabyte, HPE, and Dell. The humanoid robotics demonstrations generated significant interest, though questions remain about real-world task versatility beyond controlled stage environments. The OpenShell framework's focus on sandboxing and dynamic privilege management addresses enterprise security concerns around autonomous agents, a critical requirement as organizations move toward agentic AI deployments.