Cosmos Lab

Cosmos Lab is the Nvidia team led by [[LiuMingyu|Liu Ming-Yu / 刘洺堉]] in 150. 对英伟达研究副总裁刘洺堉的4小时访谈：Cosmos 3、世界模型、武术、黄仁勋影响我的，和你不需要击败所有对手. The source says Liu’s direct team is about 80-90 people, while the broader Cosmos project involves about 200-300 people across Nvidia.

The lab’s role is to build Cosmos as a [[WorldFoundationModels|World Foundation Model]] stack for Physical AI. In the source, that means helping developers with better data, better starting points, and better environments rather than only releasing a media-generation model. Cosmos 3 is the episode’s current focal point: it consolidates earlier predict, transfer, reason, and policy directions into a single Omni-Model with language, video, audio, and action.

Source Position

• Cosmos Lab is presented as an ecosystem-enabling internal team, not as a standalone startup or customer-facing application company.
• The source connects the lab to Large Company Open Source Strategy because open releases can reduce developer burden while also helping Nvidia learn future infrastructure requirements.
• The lab’s operating challenge is AI Organization Design at large-model scale: shared vision, high compute cost, many contributors, and enough autonomy for teams to make local decisions.

Connections

• [[LiuMingyu|Liu Ming-Yu / 刘洺堉]] — lab lead and source guest.
• Nvidia, Jensen Huang, and AI Infrastructure Full-Stack Moat — company and platform context.
• Cosmos 3, World Foundation Models, World Models, and World Model VLA Fusion — model-stack context.
• Physical AI, Robot Data Scale Up, and Robotics Simulation Evaluation — developer problems the lab is meant to support.