China’s embodied-AI boom is often told through the most visible products: humanoid robots walking on stages, robot arms sorting packages, and companies raising large rounds to build machines with legs, hands, and cameras. New industry data suggests that investors are increasingly looking past the body. The larger shift is toward the software, data, components, and control systems that could make those machines useful outside a carefully prepared demonstration.
An IT Juzi analysis republished by 36Kr says China had 425 embodied-AI startups as of August and that 321 of them were founded between 2023 and 2026. It reports 466 financing events worth 124.51 billion yuan in the first eight months of 2026. Most strikingly, the analysis says embodied-brain systems accounted for 38.8% of financing events this year, ahead of humanoid robots at 21.1%.
Those figures are the research firm’s own market analysis, not official national statistics, and they should be read as a measure of funding activity rather than proof of commercial success. But they capture a real strategic issue. It is easier to see a robot’s body than its intelligence. Yet the intelligence layer may determine whether a robot can handle a new object, adapt to an unfamiliar room, recover from an error, or work safely around people.
Why Capital Is Moving Beyond the Humanoid Body
Humanoid robots still attract the most attention because they are tangible and easy to compare with human movement. They also remain the largest recipients of funding in absolute terms, according to the IT Juzi analysis. Building an entire robot is expensive. It requires motors, sensors, batteries, cameras, materials, manufacturing capacity, and the integration work needed to make those parts operate together.
But a body without adaptable software has limited value. Traditional industrial robots can repeat a fixed task with great precision, yet they often require expensive programming and a carefully controlled environment. Embodied AI aims at a more flexible goal: a system that can perceive its surroundings, interpret instructions, learn from data, and alter its actions when conditions change.
That is why “brain” systems are attracting a growing share of investment events. The phrase includes perception models, world models, robot control software, simulation environments, data tools, and the infrastructure used to train or evaluate physical AI. These systems are less visible than a polished humanoid, but they may be more reusable across many kinds of machines.
EastFrontier’s coverage of 51World’s data platform for training embodied AI illustrated this point. A robot needs far more than hardware to learn how to operate in the physical world. It needs examples, simulation, testing environments, and feedback loops that help it connect perception with action. Investors are increasingly directing money toward those less glamorous but essential layers.
A Crowded Market With a Concentrated Geography
IT Juzi’s data also highlights how geographically concentrated China’s embodied-AI startup scene has become. The analysis counts 285 companies in Beijing, Guangdong, and Shanghai combined. That concentration is not accidental. Beijing has a deep pool of AI research talent and national institutions. Guangdong benefits from dense hardware supply chains and manufacturing expertise. Shanghai combines research, finance, industrial customers, and access to capital markets.
The same geography helps explain why an embodied-AI company can form quickly but still struggle to scale. A founder may be able to find sensors, engineering talent, contract manufacturers, and early customers more easily in one of the major hubs. Outside them, the costs of building a supply chain and recruiting specialized workers can be much higher.
The industry is also very young. IT Juzi says three-quarters of the 425 startups were created between 2023 and 2026. That means many companies are still in early rounds, still refining a technical approach, or still searching for a problem that customers will pay to solve. A surge in startup formation is evidence of interest, not a guarantee that the market can support every entrant.
Recent funding has reinforced the sense of momentum. EastFrontier reported that XPeng’s Dogotix raised more than US$900 million, a round that gave physical AI a prominent new valuation benchmark. Yet large rounds can cut in two directions. They provide companies with the capital needed for data collection, hardware iteration, and deployment. They also raise expectations that companies will move from prototypes to repeatable products quickly.
The Hard Problem Is Still Generalization
The capital shift toward embodied brains reflects an unresolved technical problem: generalization. A robot may perform well when it has been trained on a particular warehouse layout, set of objects, or task sequence. It may perform poorly when lighting changes, an object is new, or an unexpected person enters its path. Closing that gap requires better models, better data, and better ways to test systems before they touch the real world.
The IT Juzi analysis frames the transition as a move from building bodies to building brains. In reality, the two cannot be separated. A control model needs to understand the limits of the hardware. A hardware team needs sensors and actuators that can generate reliable data. A company that builds a robot hand must consider the software required to use it. The more useful framing is that investors are broadening their attention from the final machine to the entire chain that makes the machine intelligent.
That broader view is already visible in China’s training infrastructure. EastFrontier noted that embodied-AI training grounds are expanding beyond pilot projects. These facilities can create data, simulate working conditions, and give startups a place to test systems before a commercial customer takes on the risk. They are part of the “brain” story even though they may look like warehouses, factories, or test spaces.
The sector’s challenge is to turn investment into systems that perform reliably in ordinary settings. The market has already shown that investors will fund the idea. The next question is whether robots can handle the last difficult steps: moving safely through a changing environment, manipulating unfamiliar objects, and completing tasks without constant human rescue.
For China’s embodied-AI companies, that may be the real meaning of the new capital mix. The race is no longer only to build the most convincing robot body. It is to build the intelligence and infrastructure that makes a body economically useful.