Chinese Team Brings World‑Model AI to Mass Production – The Physical‑World Anthropic

Physical AI has reached a critical point: unlike large language models that predict the next token, world‑model AI predicts the next frame of the physical world, enabling a deeper understanding of physics for autonomous driving.

For the past decade, the dominant approach in autonomous driving has been imitation learning—recording human drivers and training AI to mimic their actions. This method hits a ceiling defined by human performance because it learns only "what" to do, not "why"; it captures actions without causal reasoning, and cannot handle rare edge cases such as a heavy truck failing to stop on a wet road.

Momenta’s solution is a world‑model stack called R7, organized into three layers:

Pre‑training: Leveraging data from a fleet of 900,000 production cars that have collectively logged over 100 billion kilometers, Momenta extracts roughly 100 million high‑quality segments that encode fundamental physical laws such as mass, inertia, and collision dynamics.

Simulation: Because extreme scenarios (e.g., a pedestrian suddenly stepping onto the road) are extremely sparse in real‑world data—occurring once per 100,000 km—Momenta generates virtual scenes using the learned physics, dramatically narrowing the gap between simulation and reality and achieving a simulation‑to‑real efficiency gain of more than ten‑thousand‑fold.

Reinforcement learning: In these simulated extremes, the model iteratively trials and errors under reward‑penalty signals, evolving beyond mere imitation to discover optimal actions constrained by physics.

The R7 loop is self‑reinforcing: production‑fleet data continuously feed the pre‑training stage, the evolved model is pushed back to vehicles via OTA updates, and improved performance attracts more OEM customers, expanding the data pool further.

Industry momentum is evident. In June, NVIDIA open‑sourced Cosmos 3; Tesla released FSD V14.3.4; a Chinese company valued at over ¥100 billion passed its Hong Kong Stock Exchange hearing, with 65 % of its assisted‑driving solutions deployed in mass‑produced cars (over 900,000 units). Meanwhile, AI‑focused startups such as AMI Labs (founded by Yang Li‑kun) raised $1.03 billion on the JEPA architecture, and Li Fei‑fei’s World Labs secured $1 billion from NVIDIA and AMD. Google’s Gemini Omni integrated a world model into its Gemini system.

Anthropic, though not the earliest large‑model player, found a high‑value niche in programming—Claude Code reached $2.5 billion ARR within a year, and Anthropic’s total ARR surged from under $1 billion to over $30 billion, later expanding into finance, healthcare, and enterprise services.

Momenta mirrors this trajectory in the physical‑AI domain: it first proved the world‑model concept in autonomous driving, then leverages the same foundation for Robotaxi, Robovan, and Robotruck. The platform’s hardware‑software integration creates a thicker moat than pure‑software APIs, with OEM relationships and OTA pipelines serving as critical iteration channels.

According to CIC Zhuo Shi consulting, the combined market for mass‑produced autonomous driving, Robotaxi, Robovan, and Robotruck could exceed $500 billion globally by 2030. As the physical‑AI foundation solidifies, the market is expected to become a “red sea” for digital AI, while physical AI remains an emerging “blue ocean” with potentially even larger economic impact.

In summary, the convergence of massive real‑world data, world‑model pre‑training, high‑fidelity simulation, and reinforcement learning positions Momenta’s R7 as a platform that could usher in a “GPT‑moment” for the physical world, much as large language models did for digital AI.