This article is reprinted from Cailian Press, all rights reserved. Click to view the original link.

---

Is the Physical AI Explosion Imminent? NVIDIA Releases "All-Powerful" World Model, Boosting Robotics and Autonomous Driving

During the 2026 Taipei GTC conference held today, NVIDIA released Cosmos 3—a physical AI world model based on a breakthrough Transformer hybrid architecture.

Cosmos 3 is reportedly the world's first fully open-source Omni-Model (also known as a multimodal model). It can natively understand and generate text, images, video, ambient sound, and actions with leading physical accuracy, shortening the training and evaluation cycles for physical AI from months to days.

"Thanks to breakthroughs in multimodal reasoning languages, vision, and world models, the explosive growth of physical AI is imminent," said Jensen Huang. "Cosmos 3 will enable developers to make generational leaps in building robots, autonomous vehicles, and visual AI, allowing them to perceive, reason, plan, and act in the physical world."

Previously, the Cosmos series platforms were integrated from multiple specialized models, requiring multiple models to run to complete the process of inference, generation, and action. In contrast, Cosmos 3 addresses a fundamental challenge in physical AI: enabling robots and autonomous vehicles to work stably in the real world despite limited training data and fragmented simulation stacks.

Among open-source world models, Cosmos 3 ranks first in world generation accuracy in Artificial Analysis, Physics-IQ, PAI-Bench, and R-Bench tests; in terms of action policies, Cosmos 3 ranks first in RoboLab and RoboArena tests.

To date, Cosmos has numerous users in the robotics and autonomous driving fields. According to a list disclosed by NVIDIA, these include Agile Robots, Doosan Robotics, LG Electronics, Samsung, Skild AI, and Li Auto. Notably, NVIDIA announced today the launch of the Isaac GR00T humanoid robot reference design, which will be introduced by Unitree by the end of 2026, alongside collaborations with humanoid robot manufacturers in the US, Europe, and South Korea to create research-purpose robots.

Going forward, the Cosmos series models will continue to develop. NVIDIA announced today the launch of the NVIDIA Cosmos Consortium, a global collaborative organization composed of world model builders and AI developers, including Agile Robots, Black Forest Labs, Generalist, LTX, Runway, and Skild AI, who will jointly strive to advance the development of next-generation world models.

SDIC Securities believes that physical AI is the next evolution direction for artificial intelligence, currently facing multiple challenges including high data costs and difficulties with physical laws. World models and spatial intelligence will provide the foundation for physical AI to perceive and cognize the 3D world.

Zheshang Securities points out that there is no fixed implementation paradigm for physical AI yet. Currently, physical AI requires the joint realization of world models + VLA (Vision-Language-Action), and the implementation method for world models has not yet converged. According to NVIDIA's calculations, the reshaping space for manufacturing and logistics industries corresponding to physical AI is approximately $50 trillion; according to investment firm Coatue Management's calculations, the baseline market size for physical AI is $6 trillion.