China's Smart-Driving Race Faces a "Physical AI" Reckoning

China's smart-driving sector is being pulled into a new arena. Algorithm providers that built their business around hardware sales and development fees are now being measured on whether they can deliver foundation models for the physical world. KrAsia reports that Zhuoyu Technology, an independent DJI spinoff, unveiled a native multimodal foundation model for mobile physical AI at the latest Beijing Auto Show.

Yu Beibei, vice president of Zhuoyu, put it bluntly to 36Kr: "If you don't get on this technology path, you might not be able to break through in the future." The pressure comes not from autonomous-driving peers but from larger AI players moving in from embodied intelligence, where vision, language, action and reasoning are trained together.

From expert models to foundation models

Yu describes the industry as sitting at an inflection point: continue with specialized expert models or switch to a large-model approach where vision, audio, action, rules and reasoning are all treated as modalities during pretraining. Zhuoyu's training mix today is roughly 30 percent vehicle data, 30 percent robot data and 40 percent first-person-view video from the internet. The commercial model shifts too. The first growth curve sells hardware and dev fees. The second extends the stack into Level 4 robotaxis and robovans, with revenue from profit sharing and subscriptions.

The macro frame

Humanoid Daily, citing the new Capgemini Research Institute report, places this inside a larger industrial pivot. The survey of 1,678 senior executives across 15 industries found nearly 80 percent of organizations already engaging with physical AI. The addressable sectors represent an estimated 50 to 80 trillion US dollars of global GDP. Labor shortages were cited by 74 percent as the top driver. Capgemini's view: growth over the next three to five years will be led by intelligence embedded into proven form factors like autonomous mobile robots, not general-purpose humanoids, where average scaling timelines now sit at about seven years.

Sources: KrAsia, Humanoid Daily