DeepSeek is leading a $400 million seed round for a stealth robot AI startup

DeepSeek, the Chinese AI company that upended the global model race with R1, is leading a $400 million seed round for a stealth robotics AI startup called Winverse. 

The deal was first reported by Bloomberg, but it is not yet closed. The roster of participating investors and the final terms remain subject to change. The direction, however, is unambiguous. DeepSeek is moving into embodied AI, and it is doing it with serious capital at a serious valuation for a team that has never publicly shipped a robot product.

Who is Hu Han?

Winverse was founded in 2026 by Hu Han. Before Winverse, he was lead scientist and head of Tencent Holdings' vision-language and embodied foundation model teams. Before Tencent, he worked at Baidu. Before that, Microsoft Research Asia. He holds a PhD from Tsinghua University.

That biography follows a well-worn path through China's AI research hierarchy: top-tier university, elite research lab, hyperscale tech company, and then out on his own. What's different is the exit timing. Hu is not leaving to start another LLM lab. He's starting a robotics foundation model company, which is a more specific and more capital-intensive bet on where the frontier moves next.

Meituan's Long-Z Investments is also in the round, alongside DeepSeek. Meituan is China's dominant food delivery and local services platform, with deep interest in last-mile logistics automation. Both investors have obvious strategic reasons to want better robot AI.

DeepSeek's physical AI strategy

This is not DeepSeek's first move into robotics. 

In August 2026, DeepSeek took a 2.31% strategic stake in Unitree Robotics ahead of Unitree's Shanghai listing. That stake was described as a collaboration on embodied-intelligence models, rather than a purely financial investment. DeepSeek, the company that built R1 by applying aggressive reinforcement learning and efficiency techniques to language models, is now asking whether the same approach scales to physical world models.

The Winverse bet extends that logic. A seed investment at $400 million implies that DeepSeek is not buying access to a product. It's backing Hu Han's thesis about how to build the AI layer that makes robots generally capable, the same way R1 was a thesis about how to make language models generally capable at a fraction of the usual cost.

DeepSeek itself is raising more than $12 billion, which gives it deployment capital that most AI companies don't have. 

The Winverse round, if it closes at these terms, would represent a meaningful allocation of that capital into physical AI infrastructure.

The seed valuation requires explanation

$400 million at seed is extraordinary by any standard. 

For context: Odyssey-3 world model, a world model that already controls humanoid robots across a range of tasks. Physical Intelligence, the San Francisco company building manipulation policies, raised $400 million at its Series B after shipping demonstrable capabilities. Winverse is raising $400 million at seed, with no public product.

The price is buying Hu Han's team and their thesis. In China's AI labor market, a former Tencent foundation model head with a Tsinghua PhD and Microsoft Research Asia background clearly commands a premium. The investors are paying for the probability that this particular person, with this particular background, can build the right technical approach before someone else does.

That calculus is becoming more common as foundation model talent disperses out of China's hyperscalers. Tesla Optimus chief's new startup The pattern holds across both Chinese and American markets: the premium for proven deep technical leadership in physical AI is high, and investors are paying it early.

What Winverse is building

Winverse is building AI for robots. The company is in stealth, so no architecture has been published, no demo shown, no customer announced. What's publicly known is Hu Han's background in vision-language models and embodied foundation models. And the core capability gap he's trying to solve: a robot that can see the world, understand language instructions, and then move through physical space to execute them reliably.

China's 70 robot training schools that provide the data infrastructure for exactly this kind of research. Those facilities run robots through repetitive task demonstrations, generating the training data that foundation models need to learn physical manipulation. 

If Winverse's approach requires that data, the infrastructure is already there.

The competitive set is formidable. AWS physical AI toolchain NVIDIA is funding model development through its robotics research program. Mecka AI training data Every major player in AI is racing to own some part of the stack between "language model" and "robot that reliably does things in the real world."

Why DeepSeek specifically

DeepSeek's core capability is efficiency: building models that match frontier performance at dramatically lower cost, through better training algorithms and more aggressive use of reinforcement learning. 

(R1 was not just a capable model. It was a demonstration that the prevailing assumptions about model cost were wrong.)

If the same efficiency thesis applies to physical world models, DeepSeek is exactly the right institutional backer for Winverse. An embodied AI foundation model that achieves strong manipulation performance at a fraction of the compute cost would do for robot AI what R1 did for language AI: make the technology accessible to hardware companies that can't afford frontier model training runs.

China's humanoid hardware supply chain With DeepSeek building physical AI (along with all the other Chinese companies working on it as well) China is invested on both sides: hardware and software..

Sources

• Quartz — DeepSeek leads $400 million seed round for robotics AI startup
• The Next Web — DeepSeek to lead seed round in $400mn robot AI start-up, Bloomberg reports
• Investing.com — DeepSeek to lead funding round for robotics AI startup Winverse at $400 mln
• Investing.com — DeepSeek backs ex-Tencent scientist's robotics startup Winverse AI