The OpenAI of Physical AI Won't Be OpenAI, Says a Founder Who's Already Sold Two Companies

Two minutes. That is how long Yao Song (姚颂) felt pure joy after selling his first company. In 2018, at 26, he sold DeePhi Technology (深鉴科技), an AI chip startup he had co-founded while still at Tsinghua, to Xilinx for roughly $300 million. The "joy lasted two minutes," he said in a 2h55m 晚点 LatePost interview published July 22, 2026. Then the gray set in. "Life went dim," he told host Cheng Manqi (程曼祺).

A decade later, Yao Song is betting his third company on a contrarian thesis about the next AI cycle. His new venture, Striding AI (正行创新), closed a near-$100 million angel round in June 2026, anchored by Thailand's CP Group (正大集团) and Tsinghua researcher Yu Chao (于超), to build what he calls "physical AI": machine learning for robots, autonomous systems, and other machines that act on the physical world, not chatbots.

The bet is that the LLM-lab consensus is structurally wrong about who wins this cycle. "OpenAI and Anthropic will not pick the fruit of physical AI," Yao Song said in the LatePost interview. "All the manufacturing and internet giants will pour in. This is the future pillar of the economy."

The bottleneck, he argues, is not model capability. It is the algorithm-to-value chain: the gauntlet of customer co-development, real-world data flywheels, and deployment economics that any physical-world AI product must clear before it earns revenue. A better benchmark score does not shrink that chain. A shorter translation chain does.

Striding AI has built its own test for that. The company is betting on a "latent space" approach, where a model learns compressed representations of the physical world that can be acted on directly, rather than chasing the world-model (WAM) approach that dominates current physical-AI research. The difference matters for end-side deployment, where model size and inference speed are constraints from day one. Yao Song puts the trade-off in one line: "If you don't want to be boring, then be happy while suffering."

The credibility he is asking the reader to weigh is concrete. DeePhi was an early mover in compressing deep-learning models onto FPGAs (reconfigurable chips used as an alternative to GPUs for AI inference), and its sale to Xilinx at the peak of the CNN-era buildout (the 2012–2018 wave of investment in convolutional neural networks for computer vision) validated a pattern Yao Song had spotted at 22: the real moat was in inference, not training. He has since watched that lesson generalize. "In the CNN era," he said, "the names were Alex [Krizhevsky], Kaiming He (何恺明), Shaoqing Ren (任少卿), Jifeng Dai (代季峰). No one leads forever."

His second company, 东方空间 (Oriens Space), a commercial launch provider, executed the first private solid-on-solid booster launch in China in 2024, a single rocket stacking two solid-fuel stages, the configuration that makes small payloads cheap to orbit, and the first livestreamed private launch in the country's commercial space sector. That round of capital, talent, and pacing taught him that every milestone needed a commercial landing, or the company would stall. "Intelligence-first works only for a very few," he said. "At every milestone, you need commercial traction."

He is applying the same yardstick to physical AI. The 24-month falsifier, the test that would prove him wrong, is clean: an LLM lab ships a physical AI product with non-trivial deployment economics, unit economics that hold up at scale rather than pilot deployments or demo reels, inside two years. If OpenAI or Anthropic does that, the bet collapses.

The bet is also a critique of founder-as-celebrity culture. Yao Song draws a sharp line between entrepreneurs ("ten attack, zero defense") and CEOs ("eight attack, two defense"). He named the new company Striding, "to stride, to walk properly," and gestured at CP Group's 80-year-old chairman Dhanin Chearavanont (谢国民), who once told a meeting, "Don't any of you stand up; you are all the teachers," as the model of a long-game operator. The contrast with the late-2010s mythology of the "prodigy founder" is intentional.

The cohort Yao Song came out of at Tsinghua (the same network that produced Moonshot AI's Yang Zhilin / 杨植麟, Infinigence's Gao Jiyang / 高继扬, and the academic researchers Wang He / 王鹤 and Xu Huazhe / 许华哲) used to time-block eleven hours a day of study. None of them, in his telling, can solve physical AI alone. "One or two geniuses cannot solve it," he said. The century's economic pillar, if his read is right, will be built by teams that look more like the CP Group elder than the late-2010s founder myth.

The bet was placed in June 2026. It has 24 months to prove itself.