Overview
On October 22, the 2026 World Physical AI Conference and Unitree Partner Ecosystem Conference opened in Hangzhou. Unitree founder Wang Xingxing argued that embodied intelligence’s “ChatGPT moment” could arrive as soon as next year, with physical AI already showing signs of a breakthrough and a window of roughly one to three years. He demonstrated GPT-6 directly controlling a robot to draw and play Go with zero training, and outlined directions including ten-million-unit annual hardware output, deployed physical-AI models, and self-sufficient compute and chips. In parallel, the nation’s first physical-AI and world-model investment fund was signed in Nanjing; Jiangsu released a high-quality dataset plan and launched a cross-provincial data-standard city alliance, with a second batch of industry deals exceeding 6 billion yuan.
Background & Interpretation
1. Technical Background & Evolution
The embodied-AI race is shifting from “parameter contests” to “putting models into mass-producible, trustworthy robots”. Large models gave AI cognition in the information world; bringing AI into factories, homes and streets depends on mass-produced humanoids. Unitree’s G1, with tens of thousands of units shipped, validates real demand, and its expected STAR Market listing provides capital for sustained R&D. World models and physical AI have become the spine linking models, robots, chips and data.
2. Core Drivers & Mechanisms
Physical AI rests on three pillars: large models, real-world data and on-device compute. On-device compute is the most fundamental yet weakest link—most robots still rely on 8 nm or older dual-chip split architectures (brain reasoning plus cerebellum motion control), causing latency and power waste. Fusing “brain and cerebellum” on a single chip is the breakthrough direction. On the data side, quality datasets and cross-provincial standard alliances lay the training foundation; on capital, dedicated funds supply ammunition.
3. Market Structure & Industry Chain Impact
Industrial capital is racing to stake out physical-AI infrastructure—from domestic compute chips, optical interconnects and high-density power to algorithm efficiency—forming a chain-style layout. Over half of leading embodied-AI players use D-Robotics’ platform in production, showing domestic on-device “brain” chips have entered scaled validation. Competitive focus is extending to standards, safety norms and mass-production capability.
Implications & Outlook
If the “ChatGPT moment” arrives on schedule, humanoids could move from demos to scaled delivery, driving volume in joints, reducers, sensors and dexterous hands. Watch the pace of on-device inference cost decline, mass-production yield and reliability, and real-scenario willingness to pay.
Takeaways for Industry Participants
- For robot OEMs: use real-scenario constraints to force model pruning and on-device deployment; prioritise producible delivery.
- For chip and component makers: position around brain-cerebellum fusion and low-power inference; bind to leading OEM ecosystems.
- For investors and industrial capital: lay out early along the compute-power—power—interconnect—algorithm chain and embed into real order ecosystems.
This column compiles industry information and shares technical perspectives; it does not constitute investment advice.