彭博称DeepSeek拟领投机器人具身智能初创Winverse AI种子轮，投后估值约4亿美元，交易尚未敲定

DeepSeek is expanding its reach beyond digital chatbots into the physical world, agreeing to lead a seed financing round for Winverse AI. According to sources familiar with the matter cited by Bloomberg , the investment would value the Chinese robotics startup at approximately US$400 million.

Other institutional backers, including Meituan's venture arm Long-Z Investments, have also committed capital to the transaction. The deal has not yet been finalized.
If completed, the transaction would mark one of DeepSeek's largest startup bets since hedge fund manager Liang Wenfeng founded the research lab in 2023. The deal would allow DeepSeek to deploy capital from its massive US$12 billion-plus financing round to build out an ecosystem of hardware and embodied-AI partners. Bloomberg reported that DeepSeek is close to securing CNY80 billion (US$12 billion) from backers including Tencent Holdings and battery giant CATL ahead of a targeted 2027 market debut.
Closing the gap between robot hardware and intelligence
DeepSeek's move on Winverse follows its first verified entry into embodied intelligence: purchasing a 2% stake in Unitree Robotics' Shanghai initial public offering. According to stock exchange filings cited by Reuters , DeepSeek invested CNY140.8 million (US$21 million) in Unitree's CNY6.1 billion (US$.5 million) IPO, which valued the mainland's premier humanoid robot maker at nearly US$9 billion.
The Unitree deal was structured around a mutual procurement agreement: Unitree receives priority access to DeepSeek's model-training services and technical architecture, while DeepSeek gains preferential access to Unitree's quadrupeds, humanoids, and hardware platforms for co-developing embodied models.
While China leads globally in manufacturing physical robot components, endowing robots with the spatial intelligence to navigate real-world environments remains a major technological hurdle. Standard large language models (LLMs) excel at predicting text sequences, but translating human instructions into precise, real-time joint angles and physical actions requires a specialized class of vision-language-action models.
By pairing its high-level reasoning models with physical hardware producers like Unitree and specialized software developers like Winverse, DeepSeek aims to convert mechanical hardware into practical labor for factory floors and logistics hubs.
As performance gains in pure text models slow across the broader industry, combining digital intelligence with physical systems is rapidly becoming the next critical front in global AI competition.