Liang Wenfeng, founder of DeepSeek—a leading Chinese developer of large-scale AI models—has recently been focusing intensely on the field of “Physical AI” (also known as Embodied AI). He has been engaging with various entrepreneurs to explore how robots and large models can interact with the real world. He is currently weighing whether DeepSeek should make a deeper foray into embodied AI, considering options ranging from building an in-house R&D team to entering the sector through strategic investments.
Liang has previously invested in Unitree Robotics. Bloomberg, citing sources, reported that DeepSeek plans to lead a funding round for Winverse AI, a robotics company founded by Hu Han, a former AI scientist at Tencent. Although Winverse AI was established only a few months ago, it aims to raise tens of millions of dollars in a seed round, with a projected post-money valuation of approximately $400 million.
DeepSeek has agreed to spearhead this financing round, with other institutions—such as Meituan’s Long-Zhu Capital—also committing to invest.
However, industry insiders have indicated to Phoenix Technology that the company cannot be classified as an embodied AI firm in the strict sense; instead, it focuses more on underlying infrastructure, specifically model acceleration technologies.
A key factor driving DeepSeek’s interest in embodied AI is the pressure posed by GPT-6. GPT-6 Astra achieved a 95% success rate in a robotic control task—far surpassing Fable 5.1’s 40%—and various demonstrations have shown that GPT-6 delivers performance in the embodied AI space that has astonished the industry.
Sources close to DeepSeek revealed that Liang recently acknowledged a gap between DeepSeek’s capabilities and those of GPT-6 in certain areas. This realization has prompted DeepSeek to re-evaluate its technological roadmap for the next phase: is relying solely on the scaling of language models sufficient to compete in the race toward AGI (Artificial General Intelligence)?
Furthermore, industry observers note that DeepSeek’s interest in embodied AI comes as no surprise, given that other AI companies—such as Zhipu AI and Kimi—are also building up capabilities in this sector. Embodied AI has emerged as a new battleground for global AI companies.
“Embodied AI” refers to giving AI a “body”—using a robot as a physical vessel to perceive the environment, comprehend tasks, and execute actions in the real world. Major overseas tech companies have already begun positioning themselves in this area. OpenAI has focused consistently on robotics and world model technologies in recent years; according to an industry insider, in addition to GPT-6, the company is independently training a foundation model for embodied AI. Google DeepMind is also advancing research aimed at enabling AI to comprehend the physical environment. Meanwhile, Tesla’s humanoid robot, Optimus, is being trained to develop its capabilities within real-world manufacturing settings.
In the Chinese market, embodied AI has long been a hot sector. Companies such as Galaxy General, Xinghaitu, and Deta Intelligence are exploring new approaches that integrate robotic hardware, motion control, and AI models.
For DeepSeek, entering this field does not represent a complete departure from its existing focus. In past public statements, Liang Wenfeng has consistently emphasized that DeepSeek’s core objective is AGI, rather than short-term commercialization.