nvidia-cosmos/cosmos-predict2

Cosmos-Predict2 is a cornerstone of NVIDIA's 'Physical AI' strategy. While LLM labs like OpenAI and Anthropic focus on digital reasoning, NVIDIA is uniquely positioned to dominate the world-modeling layer due to its control over the underlying hardware (H100s/B200s) and simulation ecosystem (Omniverse/Isaac Sim). The defensibility is high (8) because world models require massive compute and specialized datasets that bridge the gap between video and physical telemetry—a niche NVIDIA is aggressively filling. The project has steady traction (763 stars, 102 forks) which, for a highly technical robotics framework, indicates strong institutional and researcher interest rather than amateur experimentation. The primary threat is not from other startup 'competitors' but from market consolidation; as compute costs for world models are prohibitive, the field is likely to be dominated by NVIDIA, Wayve (for driving), and potentially Google DeepMind. Platform domination risk is high because this tool effectively serves as a funnel into the NVIDIA hardware and software stack. It is a category-defining project for the 'Sim2Real' pipeline.