In the first three months of 2026, disclosed funding in China’s embodied AI sector reached about 15 billion yuan, and the count of domestic embodied firms valued above 10 billion rose to seven, from zero a year earlier. Yet the same industry shows two faces: some are charging ahead, others are selling the training servers.
Xinding Capital put it bluntly: of more than 100 robot firms funded in 2025, perhaps only 10 to 20 survive in 2026, an elimination rate of 80 to 90 per cent. Survival turns on merging the brain and the body, not on a more agile frame.
A strange capital signal backs this. In the first half of 2026 total embodied funding was about 43.8 billion yuan, but only 12.8 per cent went to building robot bodies. More than half, some 22.253 billion, flowed to brain layer algorithm and control software firms. Capital no longer trusts that building a good robot scales, it is betting on a general AI brain for all of them.
The body has hit its limit. Unitree backflips and AgiBot arms impress in labs, but a slightly shifted parcel on a belt, a product a customer moved, or a clear glass door expose the gap. Two years pushed body capability far, then hit a thicker wall: robots have bodies without brains. Huawei, Tencent, Baidu and Alibaba all shipped embodied platforms with the same stance, no body, only brain, and Tencent moved from modular add on to full stack partner.
The borrowed brain fails on structure. Figure AI took OpenAI’s model in 2024, and Physical Intelligence, Nvidia GR00T and Google Gemini Robotics followed the same path, a general model fine tuned for robots. But a model trained to predict the next frame or parse language answers “what does this mean”, while a robot needs “what will my action change”, a causal gap no tune closes. Data is the deeper wall: by early 2026 real physical interaction data totalled about 500,000 hours, under one twenty thousandth of a large language model’s, a gap above 99 per cent. The good data only comes from your own body, and no one shares it, because it is the moat.
Stanford HAI’s AI Index 2026 shows the cost: embodied robots hit 89.4 per cent success in simulation but fell to 12 per cent in real homes. A reducer maker noted millimetre precision cannot be filled by upper layer algorithms, and load to weight and rough terrain are body problems. Brain and body shape each other, and when they sit in different firms the data loop breaks.
Editor’s note: This is an adapted translation of the original OFweek report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://robot.ofweek.com/2026-10/ART-898890-8420-30705578.html.