_________

Published on Oct. 30, 2026

Physical AI aims to give machines and vehicles an understanding of the laws of the real world. That should allow them to predict what happens next and reach the next stage of automation. What they need most is data. It has become “the oil of the AI age.”

At China’s first major conference on physical AI, Momenta founder Cao Xudong explained that his company would need data from 100 billion kilometers on the road to achieve Level 4 autonomous driving. That would be equivalent to one million cars driving for an entire year.

Jiangsu province, north of Shanghai, sees this data bottleneck as a business opportunity for the entire Yangtze River Delta. The region aims to become a national powerhouse for data annotation and the training of robots and cars.

A package of measures to that end was presented at the 2026 Physical AI and World Models Conference, held in Nanjing, Jiangsu’s capital, from October 20 to 22, 2026. The event brought together China’s leading AI researchers.

Jiangsu presented an action plan for high-quality datasets. During the conference, it also formed a ten-city alliance for data annotation with neighboring Anhui province. Dozens of industrial applications were presented for testing. A dedicated investment fund was set up, and companies signed their first major cooperation agreements.

Nanjing, a city of ten million people, the surrounding province of Jiangsu and parts of neighboring Anhui belong to the highly industrialized Yangtze River Delta. They are already among China’s most advanced and prosperous regions. The new AI initiative is intended to ensure that they help shape the industrial transformation driven by AI from the outset.

Alongside the city alliance, an industry association with 40 founding members was established at the conference. Two rounds of cooperation agreements totaled more than 6 billion yuan (approximately 810 million euros), the news agency Xinhua reported on October 22, 2026 (in Chinese).

At the opening of the conference, leading researchers and entrepreners discussed the AI world models intended to connect existing factories, their wealth of data and new applications of physical AI.

“When a glass slides off a table, humans can naturally anticipate its trajectory, assess the risks and intervene to change the outcome. This is precisely one of the most difficult capabilities for AI to learn,” Dai Qionghai, president of the Chinese Association for Artificial Intelligence and a member of the Chinese Academy of Engineering, said at the conference.

Cao Xudong, founder and CEO of driver-assistance developer Momenta, explained why this approach requires so much data. Digital AI predicts the next token, compressing the knowledge of the digital world into a model, he said.

“Similarly, physical AI works by predicting what will happen in the physical world. By doing so, it compresses the underlying laws of the physical world into a model, enabling that model to develop physical-world common sense,” Cao said.

On its very first day, his company had calculated how much data Level 4 autonomous driving would require at scale. The answer was “100 billion kilometers,” Cao said. That would be equivalent to one million passenger cars driving for an entire year.

Only production vehicles can collect data at that scale. About 1.3 million vehicles equipped with Momenta’s systems were on the road and had covered 20 billion kilometers, vice president Gu Gongyao said at a parallel session on autonomous driving. After ten years, the company had reached only one-fifth of its own target.

Jiangsu’s senior Party officials now see this appetite for data as a major opportunity for economic growth in the years ahead. At the opening of the AI conference, Nanjing, Wuxi, Suzhou, Wuhu and six other cities in Jiangsu and Anhui formed an alliance for national data annotation pilot programs and other initiatives.

The ten cities plan to cooperate in five areas, including shared service platforms and mutual recognition of data standards. Their goal is a “Jiangsu-Anhui model for data annotation in industrial manufacturing” that can be replicated nationwide, speakers said during the conference.

Data annotation is the groundwork for machine learning. People or software tools label what appears in videos, sensor data and production records, and which actions lead to which outcomes. This turns raw data into training material that AI models can learn from.

The emerging ecosystem in and around Nanjing is seeking to pool resources and divide the work to produce large volumes of high-quality data more quickly. Annotation companies supply the training data, model developers provide the algorithms, and the government supports opportunities to put them to use.

The province’s new action plan sets targets for high-quality datasets. Those are accompanied by a call for 50 selected applications for 2026 across industry, government administration, research, public services and consumer goods.

The new industry alliance’s 40 founding members span basic research into world models, the development of embodied AI and autonomous driving, and industrial manufacturing. Preparatory work on a national standard for AI world models has also begun.

China’s first investment fund for physical AI and world models, formally established at the conference, has a planned size of 5 billion yuan (approximately US$750 million). It will invest in medical AI, industrial world models and autonomous driving, according to the announcement.

Using data resources for world models requires a healthy industrial ecosystem, an appropriate regulatory framework and cooperation among government, industry, universities and research institutes, Xia Bing, deputy director of the National Data Administration, said in his opening address.

Nanjing’s Qinhuai district already offers an example of how effective coordination between public authorities and innovative AI companies can be. More than 20 core companies in the connected-vehicle sector and nearly 100 of their suppliers have established operations there.

From January through August 2026, they generated revenue of more than 7 billion yuan (approximately 940 million euros), the local news portal Xinhua Baoye reported on October 28, 2026 (in Chinese).

Autonomous delivery vehicles are already operating commercially there. Connected buses have logged more than 15,000 test kilometers. The district plans to open its roads to trials in stages and at designated times of day, and to connect three previously separate test areas into a single network.

Changsha is following a similar approach, as we reported in our Week 42 Must Read. There, construction equipment manufacturers are making their factories available as training environments. Jiangsu is taking a slightly different route by organizing the data preparation itself.

As autonomous driving wins the race to bring AI into mass-produced products and everyday life, Jiangsu is already looking at the opportunities created by the convergence of the auto industry, robotics and factory automation. Again and again, the starting point is high-quality datasets.

“Moving autonomous driving toward physical AI is not just an algorithm competition. It is also a test of how effectively data, vehicle platforms, chips, real-world scenarios, and policy can work together as a system,” said Chen Songcan, president of the Jiangsu Artificial Intelligence Society.

The conference’s keynote speeches offered more than grand visions. They also acknowledged that China and the rest of the world are still at the beginning of this journey. “The physical world has a much lower tolerance for errors than the digital world,” said Gao Haichuan, founder and CEO of Qianjue Technology, as he presented his company’s Polibrain world model. It is designed to let robots predict the consequences of their actions directly and avoid “physical hallucinations,” predictions that appear plausible but violate the laws of physics.

Speaking by video, Turing Award winner Andrew Yao warned that models must recognize their operating limits or risk making inaccurate predictions in the real world. Yann LeCun, who also joined remotely, asked why AI could write code and pass the bar exam but could not clear a dinner table.

“Physical AI is still at the starting line of a 100-kilometer marathon,” said Chen Boyuan, founder of Inverse Matrix Technology and head of an innovation center at the Beijing Academy of Artificial Intelligence, or BAAI.

Yu Haibin, a member of the Chinese Academy of Engineering and director of the Institute of Industrial Artificial Intelligence at the Chinese Academy of Sciences, brought the discussion down to the factory floor with an example from car manufacturing. “A world model is AI’s worldview. It reflects AI’s understanding of the laws governing the physical world.

“By leveraging world models, AI can conduct mental rehearsals to make better decisions,” Yu said. When fitting car doors, manufacturers often have only limited measurement data, he explained. The door’s three-dimensional strain field is difficult to model. A world model of the assembly process could make real-time control more precise and flexible across different vehicle models.

The most compelling contribution came from Harry Shum, council chairman of the Hong Kong University of Science and Technology and a foreign member of the US National Academy of Engineering. Robots still lack truly dexterous hands, he said. A human hand has 27 degrees of freedom and nearly 30,000 tactile sensors, and more than a third of the neurons in the brain are involved in controlling the hands, he explained. Only the ability to grasp and manipulate objects can take robots “from being performers to becoming workers.”

“Large language models have arrived, and we can no longer outperform them in mathematics. Unitree robots have arrived, and we can no longer outperform them at jumping and moving around. So what do we have left? We still have our dexterous hands,” Shum said. In the AI age, dexterous hands are “the last bastion of human dignity,” while dexterous manipulation is “the Mount Everest of embodied intelligence technology,” he said.

He too ended his speech by paying homage to the holy grail of physical AI, large, clean datasets. “First and foremost, the most important thing is data. Only with data do we have the opportunity to build truly remarkable models,” Shum said. “And once we solve the problem of dexterous robotic manipulation, more and more business models will become possible.”

That is precisely what Jiangsu is hoping for.