As the tech industry races to master physical automation, Chinese startups and state-backed funds are shifting focus away from text-generation models toward world models and physical task execution. According to CNBC’s The China Connection newsletter, companies like QJ Robots and Lumos are capitalizing on a manufacturing ecosystem edge to build foundational technologies that help robots learn human skills.
The Shift from Large Language Models to World Models
While U.S. giants chase human-level artificial general intelligence with massive capital expenditures—such as OpenAI’s projected need for $115 billion through 2029—Beijing is taking a more targeted approach. According to CNBC, China quietly launched a 60.06 billion yuan ($8.42 billion) national AI fund at the start of the year. Rather than prioritizing consumer-facing chatbots, this state-backed capital is flowing into hardware, sensors, and foundational physical models.
This ecosystem focus aims to overcome the primary bottleneck identified at Beijing’s World Robot Conference: teaching humanoids how to interact fluidly with the physical world. Unitree’s founder noted at the conference that the biggest challenge for humanoids is learning human skills. In response, local developers are focusing on world models capable of environment prediction and high-efficiency task execution.
Data Acquisition and Industrial Ecosystems
QJ Robots, a three-year-old startup founded by Tsinghua University automation PhDs Haichuan Gao and Tianren Zhang, represents this lean operational strategy. Backed by Temasek, the company’s AI model predicts the world around them to streamline robotic task execution. According to CEO Gao and CTO Zhang, the startup has deployed its model to over 100,000 robots primarily across mainland China, pulling in more than 100 million yuan ($15 million) while continuously harvesting operational data.
However, Zhang emphasized that more kinds of data are needed to train the model effectively. Other firms are tackling this data deficit through specialized hardware capture. Ace Robotics, chaired by SenseTime co-founder Wang Xiaogang, utilizes wearable sensors to capture precise human activity data. Wang argues that relying on online training videos introduces unrealistic special effects into models, making real-world data essential. He predicts the industry will reach a breakthrough moment by the end of next year.
Platform Integration and Real-World Friction
Deploying physical AI in industrial environments requires more than raw data; it requires a platform capable of translating data and AI instruction into task completion. Chao Yu, another Tsinghua alumnus and founder of industrial robotics firm Lumos, launched the NexCore system to bridge this gap. Designed to work with a range of robots rather than just Lumos machinery, the platform aims to streamline factory automation.
Despite these architectural strides, commercial realities remain stark. Conference attendees observed that humanoid machines folding clothes or vending water still move slowly and stiffly. One observer noted that consumer patience exists purely out of novelty, remarking in Mandarin Chinese that a human worker would definitely be scolded for such slow service. Meanwhile, market sentiment remains volatile; Unitree shares tumbled after leadership sought to lower expectations for humanoid commercialization, and Pop Mart saw shares fall following declining ex-China sales, prompting a price target cut from Citi.
Strategic Resourcefulness Amid Chip Constraints
Despite strict U.S. export controls limiting access to advanced hardware like Nvidia’s best chips, domestic firms continue to iterate. According to S&P Global Ratings Director Clifford Kurz, China’s strategy relies on resourcefulness and adaptation, leveraging deep financial resources, an expansive talent pool, and existing infrastructure. State backing remains robust; X Square Robot secured a raise of around $100 million led by Alibaba Cloud, with additional backing from state-backed entities like CDB Capital and CAS Investment, according to Chief Operating Officer Yang Qian.
Whether this localized push will replicate the state-subsidized success of China’s electric vehicle sector or echo the struggles of its earlier semiconductor ambitions remains an open question. Yet, with state-backed funds actively funneling capital into industrial hardware and smart infrastructure, the race to ground large-scale AI in physical reality is accelerating across factories and utility grids alike.
