Qianxun Smart announced on 3 June that its self-developed Spirit v1.6 embodied base model ranked first on the RoboArena benchmark, ahead of NVIDIA Cosmos3 and Physical Intelligence’s Pi0.5, and closed a 1.5bn yuan Series A+ round. In just over two years the company has completed four rounds in three months, raised nearly 5bn yuan and pushed past a 20bn yuan valuation, a funding pace the embodied-intelligence sector has never seen.
A price-to-sales multiple near 200 times
The other side of the headline matters. Public figures put Qianxun’s 2026 revenue target at about 100m yuan on roughly 200 units. A 20bn yuan valuation against 100m yuan of sales is a price-to-sales ratio close to 200 times. A company just starting to earn revenue is being priced as if the category’s winner is already decided. The real question is what the capital is betting on, and whether that bet holds.
Three logics behind the panic pricing
The first is path dependency from the large-model era. Before ChatGPT, OpenAI was worth about 20bn dollars; after the launch it leapt past 100bn. That success story is now being copied directly onto embodied AI: claim the ecosystem slot first, wait for the explosion later. Qianxun’s backers are betting on exactly that logic, racing to occupy the ‘general brain’ position.
The second is the rarity of the team. Founder Han Fengtao, a Huazhong University of Science and Technology graduate and PhD under robotics academician Ding Han, co-founded ROKAE and as CTO shipped more than 20,000 industrial robots, living through China’s industrial-robot localisation rate climbing from 3 per cent to above 50 per cent. Co-founder Gao Yang, a Tsinghua assistant professor and UC Berkeley PhD advised by Pieter Abbeel, proposed the ViLa and CoPa models. The ‘industry veteran plus AI scientist’ pairing is uncommon in this field.
The third is the time window. Han’s view is that 2026 for embodied AI is what 2023 was for language models, with a ‘ChatGPT moment’ possible between late 2027 and early 2028. Because the field looks winner-takes-most, 2026 is the positioning window, so capital is buying the seat, not the present.
Where the bull case breaks
But large models and embodied intelligence have a different commercial path. A model is software-defined with nearly zero marginal cost; a robot is software and hardware together, and every additional unit adds cost. That difference may rewrite the valuation model’s base assumption. And the winner-takes-most claim is unproven: the model layer still has Gemini, Claude and others competing, so the network effects of a robot ‘brain’ may be far weaker than in software. As one institute fellow notes, industrial investors entering usually marks a shift from tech validation to scene deployment, which means top firms from 2026 to 2027 must move from continuous fundraising to continuous delivery.
Read the original report (OFweek Robotics)
Translated and adapted from OFweek Robotics (robot.ofweek.com).