Nvidia unveiled three major components for physical AI at GTC Taipei: Cosmos 3 world model, Alpamayo 2 Super driving model, and an open humanoid robot reference platform. The announcement signals the company's push beyond data centers into robotics and autonomous systems.
Nvidia used its GTC Taipei event to consolidate its physical AI strategy, introducing tools designed for robots, autonomous vehicles, and video systems.
The flagship offering, Cosmos 3, represents a next-generation world model—a foundational AI system that can understand and predict physical environments. World models serve as a core component for training robots and autonomous systems by simulating real-world scenarios.
Alpamayo 2 Super is a substantially upgraded driving model that Nvidia positions for autonomous vehicle development. The scaled-up version builds on previous iterations to handle more complex driving scenarios and decision-making tasks.
The humanoid robot reference platform marks Nvidia's direct entry into robotics standardization. By providing an open reference design, Nvidia aims to accelerate development across the industry rather than competing as a hardware manufacturer.
The announcements reflect a broader industry shift toward physical AI—systems that operate in the real world rather than purely digital environments. Nvidia's strategy emphasizes software, models, and platforms rather than end-user robots, positioning the company as infrastructure provider.
The timing aligns with increased investment from roboticists and autonomous vehicle developers seeking unified AI frameworks. Partnerships, including ABB Robotics' collaboration with Nvidia, underscore growing adoption of these tools across manufacturing and logistics.
These releases follow Nvidia's recent expansion into consumer PC chips, demonstrating the company's broadening portfolio beyond its dominant GPU business. The physical AI push represents a longer-term bet on industrial and commercial applications where AI systems must interact with tangible environments.
Nvidia's reference platform approach suggests the company views standardization as essential for ecosystem growth. By opening the humanoid design, Nvidia enables developers to focus on AI software rather than mechanical engineering, potentially accelerating deployment timelines across industries.
Liquid AI has released an 8 billion parameter mixture-of-experts model trained on 38 trillion tokens. The model, detailed in the company's latest announcement, represents progress in efficient large language model development.
An unknown language model called Hy3 has surged to the top of OpenRouter's performance rankings, significantly outpacing established competitors. The sudden emergence has sparked speculation across the developer community.
Chinese AI startup MiniMax Group has filed for a domestic initial public offering as it escalates competition with rivals like DeepSeek in the country's booming artificial intelligence sector.
ElevenLabs has released Dubbing v2, an AI dubbing model that maintains original speaker emotion, tone, and pacing across 90+ languages while keeping audio synchronized with video content.