Founders steering world models toward robotics and industrial simulation, away from gaming and film, point investors to where early revenue would have to come from in a crowded, fast-funded field.
The founders of two Chinese physical-AI startups have told an audience that world models, AI systems meant to predict how the physical world behaves, are far from ready for industrial use. Speaking at a forum during video platform Bilibili’s Super Science Night, Chen Boyuan, co-founder and chief research and development officer of Inverse Matrix, compared the field’s progress to a long race. "If we compare physics to a 100-kilometer marathon, we are only at the five-to-10-km mark," he said, according to Yicai.
Chen said polished demonstrations are possible in laboratories, but a wide gap remains before the technology can be deployed in factories. He expects technical approaches to converge and applications to spread across many real-world settings, following a path similar to large language models.
The caution comes from founders who have raised money quickly. Inverse Matrix, founded earlier this year, raised more than US$10 million in March and more than US$100 million in June, according to media reports cited by Yicai.
Physical accuracy matters more than visual polish
The founders pushed back on the idea that gaming, film and television are the best first markets. Chen and Jiang Zheyuan, founder and chairman of humanoid robot maker Songyan Dynamics, argued that world models fit better where physical correctness matters, such as autonomous driving, robotics and industrial simulation, Yicai reported. Jiang said gaming is feasible but rewards visual polish over strict physics.
Chen used a factory example to explain the standard. A robot that plugs in cables with a 90% success rate still fails one time in 10, and each failure can damage cables, create safety risks and cause operating losses. Inverse Matrix is running deployment trials in data centers and chip manufacturing, according to Yicai.
Training robots is the most promising near-term use, the founders said. A reliable world model would let developers train robots in simulation, replacing the cost of human operators, equipment wear and spare parts with cheaper computing time. Chen also pointed to work in places people cannot easily reach, such as high-temperature underground mines.
Chen has argued this line for months. In a July speech reported by Beijing News, he said video models aim for smooth, realistic images, while world models must be physically correct, and that future models would predict the next physical state of the world rather than the next word. He also proposed a W0 to W5 grading scale for world models, modeled on the L0 to L5 levels used for autonomous driving. Chen also leads the Behavioral World Model Innovation Center at the Beijing Academy of Artificial Intelligence (BAAI), according to Beijing News.
Money is arriving faster than products
Investors have moved quickly into the category. Thirty-nine world-model companies were founded and completed a first funding round in China this year, half of them within two months, according to an October ITjuzi report. In 2025, 30 such companies were founded, the report said.
Global deals have set high reference points. AMD acquired Fei-Fei Li’s spatial intelligence company World Labs for US$8.2 billion in September, and Yann LeCun’s Advanced Machine Intelligence Labs, founded late last year, has reached a valuation of US$3.5 billion, Yicai reported. EastFrontier has tracked the Chinese side of the trend, including GigaAI’s plans for a Hong Kong IPO.
Chen said large language models already handle many tasks in the digital world through AI agents, but applications in the physical world remain limited. That gap is the opportunity investors are funding.
Factory deployments will separate the leaders
The founders gave no timeline for commercial products and no figures for current revenue or customers. Yicai’s report left open whether world models will create value in real industries or simply add cost, and said only real deployment will answer that.
The milestones to watch are concrete ones: an Inverse Matrix customer moving its data-center or chipmaking trials into routine operation, published success rates for robot tasks trained mainly in simulation, and whether any of this year’s 39 new startups reach a second large round on the strength of paying customers.