On October 22, 2026, at the Hangzhou International Expo Center, during Unitree Technology’s 10th Anniversary World Physical AI Conference, founder Wang Xingxing took the stage. Out of his 36-minute speech, he spent most of his time doing one thing: urging everyone to believe in AI. His own products appeared only at the beginning and end.
Wang Xingxing listed recent AI advancements—solving top-tier math problems, reverse engineering breaking down traditional software barriers, video generation where character expressions are "much better than those acted out by real humans," and top large model companies achieving internal model self-evolution—and then offered a judgment: with current model capabilities, even if they do not improve further, the information world has already been "completely penetrated." He then said something unlikely for a hardware company founder: "I feel that I don't believe enough yet, I'm not firm enough."
This was not improvised. Ten days earlier at the Shanghai Industrial Fair, he had publicly admitted, "The biggest mistake I made over the past decade or so was not believing in AI enough." More intuitive changes were reflected in timelines: At the World Robot Conference in August this year, his expectation for the "ChatGPT moment" of embodied intelligence was two to three years fast, five to ten years slow; two months later, he compressed this window to "one to two years, or two to three years if slow." He even stated that when GPT-7 comes out next year, it might directly solve the embodied intelligence problem.
A hardware-born engineer became the most radical AI evangelist in the room within two months. This looks like a cognitive mutation, but looking at Unitree's situation over the last half-year, his "belief" was almost a mandatory question.
Wang Xingxing was very frank on stage: "After all, we are a new company," and it is "not realistic" to build exactly the same models as the world's top multimodal large model companies. Unitree will conduct independent research and also cooperate with domestic and international large model companies. The most impressive demo he showed was connecting GPT-6 to a robot, which could wash brushes, change paint colors, and precisely place stones on a Go board without any training. He judged that planning and navigation issues have been "resolved in a sense," and what remains is aligning the input and output of multimodal large models with real robots—sensor errors, actuator errors—which "I think is also a problem that can be solved soon."
Connecting these sentences reveals the business logic behind this "faith." If the brains of robots are ultimately provided by companies like OpenAI and Google, and become stronger and cheaper as he says, then the deciding factor in the race for embodied intelligence will return to who can manufacture reliable bodies at the lowest cost and largest scale. This is precisely what Unitree excels at.
Therefore, among the most important things Wang Xingxing listed for the future, the first is not models, but the mass manufacturing capability of hardware. Unitree currently produces tens of thousands of robots and hundreds of thousands of joint motors annually, which he considers "far from enough"; true popularization requires at least "tens of millions of units per year." He even discussed computing clusters, chip capacity, and mines. As for Unitree's own model capabilities, he bets on a "self-evolution" cycle: letting AI retrieve papers, write control code, train in simulation, test on real machines, and feed back scoring results. The premise of this method is also that base models improve every month—in his words, "the more the model capability improves," the stronger this cycle "will become."