DeepSeek Leads Winverse AI Seed Round at $400 Million Valuation

A model lab starts investing in other startups

DeepSeek is best known in China for building large language models cheaply. It is now putting money into the next layer of the AI race. The Hangzhou lab has agreed to lead the first outside funding round for Winverse AI, a startup building AI for robots, according to Bloomberg, which cited people familiar with the matter.3 The seed round is worth tens of millions of dollars and would value Winverse at about $400 million after the investment. Meituan's investment arm, Long-Z, has also committed.2

The deal has not closed, and both the valuation and the list of investors could still change.36 DeepSeek, Long-Z and the founder did not respond to Bloomberg's requests for comment.2 Still, the main point is already clear. This would be one of the largest startup investments DeepSeek has made since hedge fund manager Liang Wenfeng founded it in 2023.8 It also points to a strategy that goes beyond releasing new models.

Who is behind Winverse

Hu Han founded Winverse this year. He was most recently lead scientist and head of Tencent's vision-language and embodied foundation model teams.3 He has a PhD from Tsinghua University and has worked at Microsoft Research Asia and Baidu.6 He left Tencent while the company was reorganizing its AI research group to catch up with rivals such as DeepSeek and Moonshot.8

That background is ironic. Tencent rebuilt its AI team partly because DeepSeek was pulling ahead, and now one of the scientists who left in that shake-up is taking DeepSeek's money. Tencent itself sits on the other side of the table: Bloomberg-based reports describe it as a backer of DeepSeek's own fundraising round.2 This kind of overlapping ownership has become common in China's AI scene, where the same few large companies end up funding many of the competitors.

Hu is part of a larger trend. Chinese AI researchers are leaving big tech companies to start specialized labs, many of them working on world models for autonomous driving, interactive video or robotics. Other examples include former Alibaba Qwen tech lead Justin Lin and former ByteDance intern Tian Keyu.8

The $400 million number has been misreported

The biggest problem in the coverage is the size of the deal. Several outlets wrote headlines describing a "$400 million seed round" or a "$400 million funding round."16 That figure is the post-money valuation, not the amount raised. Bloomberg's reporting, as repeated by Investing.com, The Next Web and Techmeme, puts the actual investment at tens of millions of dollars.32 In other words, investors would own a minority stake in a company valued at roughly $400 million. They are not putting $400 million into a startup that is only a few months old.

The gap matters for anyone tracking AI funding. A $400 million valuation at the seed stage is very large by normal venture standards, especially for a company that is still in stealth.4 But it is a long way from the multibillion-dollar valuations of US robot AI companies. Taking the headlines literally would exaggerate how much capital is flowing into Chinese embodied-AI startups.

A second difference is about timing. Some crypto-focused news services said Winverse had "completed" the round.5910 The underlying Bloomberg reporting says DeepSeek has agreed to lead the round and that terms could still move.3 The more careful version is the right one: this is an agreed deal, not a finished one.

Where DeepSeek's money comes from

The Winverse investment depends on DeepSeek's own fundraising. Bloomberg has reported that DeepSeek is close to raising at least 80 billion yuan, or about $12 billion, which is more than its target, ahead of a planned stock market listing in early 2027.3 Most of that money was intended for model training and data centers. It could also let the lab start backing a new group of startups that develop technology alongside it.8

Reports differ on how far that raise has gone. Startup Fortune said DeepSeek had already closed its first outside round at more than $12 billion, at a valuation above $50 billion, with Tencent and battery maker CATL among the investors, citing an October 6 Bloomberg report.1 Most other coverage still says the raise is close to done rather than finished.62 Either way, DeepSeek began spending new capital on outside bets within days, which is unusual for a lab that built its reputation almost entirely on open-weight models.1

This looks like DeepSeek turning into a strategic investor, much like OpenAI and Nvidia have done in the US. When a model lab backs startups that will build on its research, it can lock in partners, customers and access to the data that physical-world AI needs.

The Unitree deal set the pattern

Winverse is not DeepSeek's first move into robotics. The lab bought a stake in humanoid robot maker Unitree when Unitree listed in Shanghai in August, with the stated goal of working together on embodied-intelligence models.3 Quartz reports the stake as 2.31% and says DeepSeek agreed to a 36-month lock-up, three times the holding period most other strategic investors accepted. Analysts saw that as a partnership rather than a bet on returns.6 Startup Fortune, citing an exchange filing reported by Reuters, put the investment at 140.8 million yuan, about $20.8 million. It added that the deal came with a purchasing arrangement: each company gets priority access to the other's training services or AI models.1

Unitree's listing, valued at about $900 million in Investing.com's account, made China's lead in humanoid robot hardware more visible.8 Put together, the two deals look like a plan. Unitree gives DeepSeek access to robot bodies. Winverse, led by a specialist in vision-language and embodied models, could give it the software that controls them.

The race to build robot brains

The coverage agrees on the main reason for the deal. China leads in building the arms and legs of humanoid robots. But getting those machines to understand speech, solve real-world problems and do more than acrobatics requires different AI models from chatbots like ChatGPT or DeepSeek's own.8 Startup Fortune makes the same point more directly: language models are good at predicting the next word but much worse at predicting the next joint movement, and that limitation is holding humanoid robots back.1

The valuation gap with the US is large. Skild AI and Physical Intelligence, two US startups building general-purpose AI for robots, are valued at $14 billion and $5.6 billion, with backing from companies including Nvidia and Alphabet.6 Figure AI is reportedly in talks to raise $1.5 billion at a $39.5 billion valuation.1 Next to those numbers, a $400 million seed valuation is a modest place to start.

Money is also tighter in China. Unitree and its partners keep looking for capital to pay for their next stages of research.8 That helps explain why DeepSeek's involvement matters beyond the dollar amount. In a market with less risk capital, a well-funded and well-known lab leading a round can help a startup attract other investors, as Long-Z's participation already shows.2

What this signals

The main conclusion is that DeepSeek no longer wants to be judged only by its models. The Unitree stake, the planned IPO, and now the Winverse round point to a company building a group of partners in embodied AI.36 There is also a geographic angle. Hangzhou is home to DeepSeek, Unitree and chipmaker Manycore, and local players increasingly treat embodied intelligence as the next real contest.1

Whether this works is uncertain. Winverse is still in stealth, the deal could change, and DeepSeek has no hardware or factories of its own, so it depends on partners for any real presence in physical AI.13 The deal is a cheap option on a field where US rivals are valued in the billions. If Hu Han's team makes progress in robot perception and control, DeepSeek will have bought in early at a low price.

For investors watching China's AI unicorns, Winverse is worth tracking less for the size of its first check than for who wrote it. A model lab preparing to go public is starting to back other startups, and in Chinese AI right now, the companies that build the models are increasingly also the ones funding new ventures.