China’s New AI Gold Rush: World Models
The Takeaway
- Chinese entrepreneurs launch AI labs focused on world models.
- World model startups attract significant venture capital funding.
- Chip shortages and founder experience pose challenges for neolabs.
This year, the environment is completely different, he said. “World model” has become a new buzz phrase in China’s startup sector, and Lu, who has co-authored academic papers exploring how to make AI understand a virtual spatial environment, started receiving friend requests on the WeChat messaging app from venture capitalists. His startup—FaceMind, which is based in Shanghai and Hong Kong—has raised $20 million in seed money this year from investors including Chinese cybersecurity firm Qihoo 360’s venture fund.
“If I work for big companies, I may not have the freedom to explore world models, because those companies have their own strategies and many other priorities,” said Lu, who turned down a job offer from ByteDance to work on FaceMind.
Lu is one of a huge number of Chinese entrepreneurs, including many university professors or doctoral students, rushing to launch their own AI labs. And while that trend is replicated in the U.S. to some degree, more U.S. neolabs tend to focus on diverse research labs and are run by researchers who have spent time at or even held executive positions at established companies.
What’s different about this new generation of Chinese AI labs is that they’re mostly focused on world models, which simulate real-world environments and aim to approximate the physics of how objects move and interact with the surroundings.
That’s because world models are considered the key to unlocking breakthroughs in humanoids and autonomous vehicles, two areas where China has the world’s broadest and deepest supply chain. The Chinese neolabs think they have a shot, because the race to build world models is still in the early stage, with no front-runners yet, in contrast with the well-beaten path of large language models.
Local venture capitalists say that has produced a frenzy of interest in world models reminiscent of the large language model boom in 2023, when countless new Chinese startups raced to develop large language models in what local media described as the “war of 100 models.”
“The hype around world models right now is unprecedented. Big tech companies, startups and venture investors are all rushing in,” said Jianke Zhang, a second-year AI and robotics doctoral student at Tsinghua University, who co-founded world model startup LatentVerse with his classmate two months ago. “There are so many Ph.D. students launching their own startups now,” Zhang said.
LatentVerse raised tens of millions of dollars in June from blue-chip investors like Hillhouse and is now looking at raising a new round of funding.
The gold rush is so intense that all kinds of startups are starting to adopt the world model label based on their own interpretation of the concept. For instance, PettiChat, a Chinese startup founded in January that develops wearable AI devices for pets, said in April that it is building what will be the first “animal behavior world model,” an AI system aimed at simulating and predicting a pet’s behavior.
Researchers launching their own startups for world models this year have a new formula for raising money, which they’ve dubbed “1-3-5-7.” That means they want to be valued at $100 million in their first funding round, $300 million in their second, and so forth. Some of them are even raising the first three rounds in the same discussion, asking investors to commit to all of them.
Some venture capitalists caution it could be a risky bet to invest in researchers who lack the track record of building frontier models at established labs or tech companies. By comparison, many U.S. neolab founders are researchers who have spent time or even held executive positions at established companies such as OpenAI and xAI.
The list of challenges for Chinese neolabs doesn’t stop here. All of them face a congenital weakness: a shortage of advanced chips, imposed by U.S. export restrictions. Skeptics say the Chinese neolabs will struggle to secure enough chips and talent, once the hype wears off.
While newly founded world model startups are quickly raising money, they will have to compete with deep-pocketed tech giants and well-established AI unicorns for limited chip supplies. Major Chinese AI firms Moonshot AI and DeepSeek, both of which have raised billions of dollars from investors, used Nvidia’s advanced Blackwell chips to train their most recent large language models, as The Information reported.
Researchers at new Chinese startups working on world models say they are renting chips from tech giants that operate cloud services, such as Alibaba and Baidu, to train the models. But they also acknowledge that most Chinese neolabs haven’t found a long-term solution for the compute problem.
War of 100 World Models
In the U.S., a few new world model startups founded by well-known AI pioneers have raised billions of dollars. World Labs, a San Francisco–based startup led by Stanford professor Fei-Fei Li, a pioneer in spatial intelligence research, raised $1 billion at a $5 billion valuation in February. In March, former Meta Platforms chief AI scientist Yann LeCun raised more than $1 billion in a seed round for his new startup, Advanced Machine Intelligence Labs.
What’s happening in the U.S. has sparked intense discussion in China’s tech and investment circles about the importance of world models as the critical step toward large-scale commercialization of humanoids and other automation technologies. The concept of world models resonates particularly strongly in China because the country has massive manufacturing and logistics industries, which provide real-world data from factories and warehouses, as well as strong government support for robotics.
One of the world model startups attracting a lot of attention from investors in recent months is Ziran Yizhi—“natural will” in English. Its founder, Ning Ding, an assistant professor at Tsinghua University, was the main author of a highly influential 2023 academic paper about a technique to fine-tune massive AI models by updating a tiny fraction of their parameters. The Beijing-based startup, founded earlier this year, is valued at about $570 million, and its backers include HongShan, Ant Group and battery giant Contemporary Amperex Technology Ltd., according to a person with knowledge of the matter.
A growing number of established companies from traditional industries are also backing world model startups. Last week, Shenzhen Edge Medical, a Hong Kong–listed medical device company, announced that it has invested in PrimalVerse, a Beijing-based world model startup founded in January by another assistant professor of Tsinghua University. Shenzhen Edge Medical said in a regulatory filing that it will work with the startup to develop AI-powered surgical robots.
Some Chinese AI unicorns that develop video-generation models have shifted their focus to world models. Beijing-based ShengShu Technology used to be best known for its AI video business, Vidu, which competes with ByteDance’s Seedance video models. In April, ShengShu released Motubrain, a world model that aims to serve as the brain for robots to navigate the physical environment. Shengshu now describes itself as a “general world model company.”
In January, PixVerse, another AI video-generation startup that competes with ShengShu, launched a “real-time world model,” R1, which allows users to create a realistic virtual world and keep making changes to it in real time, all with text or image prompts. Earlier this month, PixVerse announced PixVerse Game Engine, a new software framework powered by the R1 model for building interactive videogames.
While some investors are starting to describe the world model movement as a bubble, founders say the rapid concentration of capital and resources could lead to real advancements.
“Without this bubble, it may not be possible for breakthroughs to happen,” Zhang said.