Is physical AI heating up or cooling down: primary market cools while big tech accelerates deployment

From the past of "showing off athletic muscle" to today's "showing off the brain," among this year's hottest terms in the AI world, Physical AI certainly occupies a place. Looking across this year's robotics exhibitions, concepts centered on the "brain"—such as "world models," "one brain, multiple forms," and "eyes-brain-hands"—have been pushed to the forefront by major companies as core selling points.

At the World Physical AI Conference in October, Wang Xingxing, founder of Unitree Robotics, said that with the release of GPT-6, Physical AI has now shown signs of being broken through, and a key breakthrough is likely to come in the next 1–2 or 2–3 years. At present, the core challenge for embodied intelligence to pass the ChatGPT moment is aligning the input and output of AI models with real robots.

As "can it be deployed" and "can it do work" become the focus of the industry, related concepts around "Physical AI" carry a great deal of industry expectation, and the world model is one of them. As the "brain" through which robots understand the physical world and interact with the external environment, the world model is held to high expectations.

But the other side of high expectations is the possibility of falling short. Since the second half of the year, a number of "brain"-related companies have still secured large single-round financings, but there is increasingly more news of cooling in the entire embodied intelligence primary market, and the risks, bubbles, and anxiety surrounding Physical AI have followed.

Is Physical AI actually heating up, or is it "cooling down"? Reporters found that beneath this boom, primary-market capital, startups, tech giants, and the A-share industrial chain are showing markedly different temperatures.

2026 is the year of the "brain": hot money floods into "Physical AI"

According to incomplete statistics from Cailian Press Chuangtoutong, before 2025, primary-market financing events in Physical AI/world models were mostly in the single digits, rising to 23 in 2025. So far, 2026 has already reached more than 100, and multiple financing rounds for a single company have become the norm.

Judging from the distribution of financing timing, June to September 2026 was the period when the track "absorbed large amounts of money" most intensively. For example, LiberAI focuses on physical world models and completed four financing rounds from May to September, totaling nearly 1 billion yuan; Liu Songming is regarded as having the "most top-tier" founder profile among the "young geniuses" currently sought after by VCs; Manifold Space raised from a seed round in July 2025 all the way to a Pre-A+ round in June 2026, completing six financing rounds in a year, with Pre-A rounds totaling nearly 1 billion yuan.

In the world model field, it is not uncommon for startups to complete several financing leaps in a short period. Although seed rounds, angel rounds, Pre-A rounds, and other pre-Series A projects currently account for more than 70%, mostly early-stage projects, there are quite a few cases with single rounds exceeding 100 million yuan or even 1 billion yuan.

These large financings are concentrated in several types of companies: one type builds the foundation for world models/embodied brains, such as LiberAI, Octopus Dynamics, Inverse Matrix Technology, Manifold Space, Moushen Intelligence, and Westlake Shuzhi; another type builds data infrastructure and simulation engines, such as Guanglun Intelligence, Hive Connect, Feijie Kesi, and Songying Technology.

There is also a category focused on combining the brain with dexterous tactile software and hardware, or companies with core hardware mass-production barriers. For example, Lingchu Intelligence completed angel and Pre-A rounds in March 2026, officially announcing a total of 2 billion yuan received; in April, state-owned capital such as Guotou Xiandao and Jingxi Ruiling invested again; in July it completed an A round of over $100 million. Another example is Paxini, which focuses on tactile sensors and dexterous hands: in March 2026 it secured a B round of over 1 billion yuan, in August it completed a strategic financing of 1 billion yuan, and in September it received another B++ round of several hundred million yuan, bringing cumulative financing to over 4 billion yuan.

Second-half primary market: financing turns cold, LPs begin asking about real scenarios

According to incomplete statistics from Cailian Press Chuangtoutong, in October alone there were 5 primary-market financing events in Physical AI/world models, indeed a decrease from before. Entrepreneurs and investors active on the front lines have already felt the cooling ahead of time.

"Everyone was firing ammunition too fast," lamented Wang Sheng, angel partner at Inno Fund. He said many top VCs emphasized in May and June that the market was too hot and needed to slow down. "But this is a train; stepping on the brakes won't stop it all at once. Inertia caused the number of projects closed in July to drop by about 20–30%, and August fell another 20% from July. From the outside it feels like September was a cliff, but the braking had already begun earlier."

"Generally speaking, after the tide goes out, you see who is swimming naked. There is a lot of money in the market, that's true; people may just feel the pace is a bit slower, but that doesn't mean there is essentially no money," said Yu Kaicheng, director of the Westlake World Model Research Center. In his view, the commercialization of new technology is a long process: "A batch surges up, then it cools for a while, then an even bigger batch surges up—this is normal." This year, Yu Kaicheng's Westlake Shuzhi Awomo was founded and launched the "Concept World Model." Recently, this startup has cumulatively obtained over 100 million yuan in seed and angel financing, becoming one of the members of the world model startup wave.

Looking back over the past three years, financing for embodied intelligence in the primary market has made great strides, with capital continuously being released. In 2024, embodied intelligence financing was still 14.05 billion yuan; by 2025, total domestic financing in embodied intelligence and robotics was 73.543 billion yuan, with 744 investment and financing events for the year. In the first half of 2026 alone, total domestic financing in the embodied intelligence track reached 93.5 billion yuan, five times that of the first half of 2025; the number of financing events reached 322, up 137% year on year.

Many investors also said that they have indeed become more cautious recently, with the pace slowing somewhat, and their investment focus shifting from the imaginative potential of technology to investment risk and the ability to deliver on business results. However, the slowdown in dealmaking takes some time to show up in the written financing data.

In the first half of the year, a joke circulated in investment circles: "Little genius + physical AI + top-journal paper = money wired directly." Now that the market has cooled, the "little genius" in that joke is being reexamined. But in conversations, some investors cautioned that in the AI era, "little genius" is not a derogatory term, and being labeled as such does not mean a project is not a good one.

Fan Chenchao, deputy general manager of Zhejiang Venture Technology, believes that compared with the previous mobile internet era, when most startups were built on new business models, this generation of AI startups is mostly technology entrepreneurship. Moreover, "little genius" is only one point of attention in investment decisions, a "people" factor. Whether a startup ultimately succeeds depends on the combined result of many factors.

He summed up a profile formula for a foundation-model startup team: the core founder's sensitivity to advanced technology routes × team leadership/organizational capability × talent density (industry appeal) + (funding + computing power + engineering capability, etc.). He believes that "if any one of the first few capabilities is a weak link, or is zero, the final result will be very poor, or even directly drop to zero." Much of the "addition" part in parentheses can be solved with money.

At the capital roundtable at the founding conference of the Westlake World Model Research Center, Wang Sheng, a partner at Inno Fund, said bluntly that the biggest challenge for "little geniuses" comes from "future corporate governance and operations, as well as the ability to grasp business opportunities amid development and change—this is indeed what they need to learn."

Big tech jumps in: focus on foundations and ecosystems

If we shift our gaze to big tech, the situation is quite different. Google, Nvidia, AMD, ByteDance, Alibaba and others are moving quickly in layouts related to world models and physical AI. Compared with startups, big tech pays more attention to building foundations and ecosystems, and to finding new points of connection between physical AI and existing businesses, rather than directly entering the embodied intelligence "brain" track.

First, they are competing for the next-generation interaction "entry point." Take Google's and ByteDance's world models as examples; the core is real-time interaction with the virtual world.

At the end of October, Google DeepMind officially announced its latest-generation world model Genie 4 and demonstrated an experiment in which it worked with the large model Gemini 4 Argon to simulate a humanoid robot performing tasks.

ByteDance's world model benchmarked against Genie has also frequently been in the news. For example, in June, reports said Wu Yonghui set a goal for the world model to release at least one version benchmarked against Genie 3 by the end of the year; in September, ByteDance founder Zhang Yiming personally led the project, coordinating internal cross-departmental teams and investing computing power and AI R&D resources.

It is understood that the core of this R&D effort is to try to link ByteDance's AI models, cloud computing, content platforms and Pico hardware, making the world model a new flywheel. In October, Tim, founder of Media Storm, had already previewed ByteDance's world model in advance; ByteDance has not yet officially released it.

Second, hardware giants have begun to "build computing platforms for the next generation of AI," with AMD and Nvidia as examples.

On September 28, AMD announced an all-stock acquisition of World Labs, founded by Fei-Fei Li, for about $8.2 billion. Fei-Fei Li will join AMD as executive vice president and chief scientist. This is also seen as a landmark event marking physical AI as a new battleground for computing power.

AMD Chair and CEO Lisa Su framed this as a long-term bet on the next decade: "Building computing platforms for the next generation of AI requires a deep understanding of how models will evolve. We can use that insight to develop hardware, software and systems."

As another major hardware chip company, Nvidia officially released its open world foundation model for physical AI, NVIDIA Cosmos 3, in June this year. At the same time, Nvidia has partnered with Unitree Robotics to launch a new-generation humanoid robot reference design, H2+, also known as the Isaac GR00T system. In October, Nvidia also announced an expansion of the Cosmos world model platform, adding training tools, simulation checkpoints and Cosmos Data Factory blueprints to help physical AI developers accelerate data pipeline construction.

At this year's GTC conference in Berlin, Nvidia CEO Jensen Huang explicitly said that "the next wave is physical AI," which also sparked a wave of discussion about physical AI in China at the time.

Third, the foundation models deployed by big tech are beginning to move toward the physical world.

In September, OpenAI released GPT-6 Astra. The focus was not on building a dedicated embodied model from scratch, but on bringing the accumulated capabilities of general models in perception, reasoning and planning into the robotics field, which triggered industry discussion about "whether general models constitute a dimensionality-reduction strike against specialized models."

For a time, the core barriers and room for survival of specialized embodied intelligence models came under scrutiny. The core of this shock is not that Astra can immediately replace specialized models, but that it has shaken the logic that previously supported the valuations of independent embodied "brain" companies. Some overseas commentary said: "The OpenAI of the robotics field will be OpenAI itself, not Physical Intelligence or Skild."

Other major companies have also made related arrangements. Alibaba released three major models: the VLA manipulation model Qwen-RobotManip, the VLN mobility model Qwen-RobotNav, and the world model Qwen-RobotWorld, attempting to use a single model system to cover the manipulation, navigation, and prediction of physical laws for robots of different forms. In addition, Anthropic released the "Model Hardware Standard" (MHS), which overseas media interpreted as its first public entry into the fields of embodied intelligence and physical AI, with its entry point being the formulation of interface standards for AI-controlled physical devices.

The physical AI industry chain in A-shares: most are still "shovel sellers"

In the secondary market, in the just-passed October, the physical AI industry chain in A-shares set off multiple waves of limit-up surges. At present, the layout of A-share companies in physical AI basically unfolds along the chain of "perception—physical simulation/world models—computing power—embodied bodies and supporting components—industrial AI/industry scenario applications." Most companies still remain in the position of "selling shovels." Very few A-share companies truly work on physical AI simulation platforms and world models, and Suocheng Technology is one of them.

In March this year, Suocheng Technology released the world physical model "Yingzao·Wanxiang" and the agent "Dasheng," benchmarked against NVIDIA's Omniverse/Isaac Sim system. In the first half of 2026, the "Kaiwu" physical AI business had achieved revenue of 6.59 million yuan, accounting for 9.06%; R&D investment was 84.8 million yuan, up 67.53% year on year, accounting for 116.59% of operating revenue. It is worth noting that the company's total revenue in the first half of the year was 72.74 million yuan, with a net loss attributable to shareholders of 78.66 million yuan. The company also explicitly stated in its announcement that the physical AI business is currently "in the early stage of layout, and commercialization is uncertain."

Xie Jun, investment manager at Hang Seng Qianhai, believes that physical AI extends AI capabilities into the real physical world and, relying on world models, simulation, perception, and execution hardware, opens up new demand for computing power and continues the computing power narrative. He believes that physical AI is currently in the early growth stage of transitioning from technical validation to small-batch pilot delivery, and he is optimistic about three major directions: simulation and world model toolchains, perception hardware, and execution components.

For physical AI to be implemented, the "foundation" of perception, computing power, components, and simulation tools is indispensable, which is also the consensus of many practitioners in interviews. But for the secondary market, in the conceptual frenzy it is even more necessary to distinguish who is "truly laying out" and who is "riding the hype." What must be guarded against is that before embodied intelligence complete machines usher in large-scale volume growth, it is difficult for the upstream of the industry chain to have substantive performance support. From concept to orders, the industrial long run of physical AI has only just begun.

This article comes from the WeChat public account "STAR Market Daily," author: Wang Nai, published by 36Kr with authorization.