The archive · Space, Robots, Defence · Strategic decision · 2024–2026
Spirit AI Raises ~5B Yuan in 3 Months, Valuation Exceeds 10B
Founded in 2024, Spirit AI raised ~5B yuan in 3 months, betting on models over hardware; v1.6 topped RoboArena.
Spirit AI (千寻智能)
What the business is
An embodied intelligence company. It self-develops a 'General Brain' for robots: Spirit, an embodied foundation model with a VLA (vision-language-action) and world-model architecture, and also develops the humanoid robot Moz1.
Starting capital：Cumulative financing since the seed round: two rounds totaling nearly 2 billion yuan (近20亿元) in February 2026 (valuation exceeding 10 billion yuan/百亿), 1 billion yuan (10亿元) in April, 1.5 billion yuan (15亿元) in June, totaling nearly 5 billion yuan (近50亿元) in 3 months.
How it started
In early 2024, Han Fengtao, former co-founder and CTO of Rokae (珞石机器人), founded Spirit AI with co-founders Gao Yang (Tsinghua University → UC Berkeley → assistant professor at Tsinghua's Institute for Interdisciplinary Information Sciences, known as one of the 'Four Berkeley Returnees' (伯克利归国四子)) and Zheng Lingyin. In the same year, it completed 3 funding rounds: a seed round by Shunwei Capital (顺为资本) and Oasis Capital (绿洲资本); a 200 million yuan (2亿元) angel round by Shunwei Capital together with Qiancheng Capital (千乘资本); and an angel+ round by Bairui Capital (柏睿资本).
What happened
In April 2025, it completed a 528 million yuan (5.28亿元) Pre-A round led by Prosperity7 Ventures, part of Saudi Aramco (沙特阿美); that year, it released the VLA Spirit v1 early access version and launched its first commercial-grade humanoid robot, Moz1. In February 2026, two funding rounds totaling nearly 2 billion yuan (近20亿元) pushed its valuation above 10 billion yuan (百亿元); in April, it received another 1 billion yuan (10亿元) round co-led by Shunwei Capital and Yunfeng Capital (云锋基金); on June 3, it announced a 1.5 billion yuan (15亿元) A+ round and said Spirit v1.6 ranked first globally in the RoboArena benchmark, surpassing NVIDIA (英伟达) Cosmos3 and Physical Intelligence Pi0.5.
How it ended up
It completed 4 funding rounds in 3 months, raising a cumulative total of nearly 5 billion yuan (近50亿元), setting a record for financing frequency in the embodied intelligence track and becoming a unicorn valued at over 10 billion yuan (百亿元). It has launched benchmark pilots with CATL (宁德时代), JD.com (京东), Bosch (博世), and others, but bulk procurement orders have not yet landed.
Background
Spirit AI, founded in early 2024 by Han Fengtao (ex-Rokae CTO) and Gao Yang (Berkeley returnee), raised 528M yuan Pre-A in April 2025. By June 2026, it completed 4 rounds in 3 months, totaling ~5B yuan, valuation over 10B yuan.
Its bet: embodied intelligence hinges on a 'General Brain', not hardware. Founders see AI models as source of generalization; hardware improves incrementally. So Spirit focuses on data and algorithms, using self-developed UMI devices to collect real data, targeting 1M hours by 2026 to train Spirit (VLA+world-model), adapting one model to various robots.
On June 3, 2026, Spirit v1.6 topped RoboArena, first Chinese model to do so, beating NVIDIA Cosmos3 and Physical Intelligence Pi0.5. It announced a 1.5B yuan A+ round. Collaborations with CATL, JD, Bosch are pilots; large-scale orders not yet landed.
What has to be true
- Large AI models gave robots generalizable skills, while hardware didn't leap; intelligence became the highest-leverage investment.
- Team is rare: Han Fengtao delivered 20,000+ robots; Gao Yang leads in pre-training to reinforcement learning.
- Data is the bottleneck; Spirit uses self-developed UMI devices to secure real data, betting data density sets model limits.
- Industrial investors like Shunwei and Yunfeng entered, signaling shift from validation to deployment; they bet on robots in their operations.
What can be applied
While hardware costs fall and many build robot bodies, Spirit AI invests in algorithms and data, making the model the adaptable layer.
Aftermath
As of June 4, 2026, Spirit AI's valuation had surpassed 10 billion yuan (百亿元人民币), Spirit v1.6 had topped RoboArena globally, and the humanoid robot Moz1 had been released. The company's collaborations with CATL, JD.com, and Bosch are still mainly benchmark pilots; a common industry problem is the gap between laboratory success rates and real production-line deployment results. The company's short-term performance target is to achieve a 95% task success rate after pre-training with only 1 hour of fine-tuning data.
Sources
spotted an error? The archive wants to know.
Your turn
You just read one. Describe what you are building, and see who is betting on the same thing.
Free account · 3 free questions · no card