Liang Wenfeng has invested in a post-1980s Tsinghua PhD, with the valuation reaching 2.7 billion.

DeepSeek is going to invest in another robotics company.

On October 12, according to foreign media reports, DeepSeek plans to lead the seed round financing of robotics AI startup Winverse AI, with an amount of tens of millions of US dollars, and a post-investment valuation of about 400 million US dollars (equivalent to about 2.677 billion yuan). Institutions including Dragon Ball Investment under Meituan have also committed to participate in the financing.

The transaction has not been finally completed yet.

Surprisingly, this company was established only a few months ago and has not made its public debut so far. Its founder Hu Han, who used to be the head of Tencent Hunyuan Vision and Language Model, just left his job to start a business in July this year.

Two months ago, DeepSeek also spent 140.8 million yuan to participate in the IPO strategic placement of Unitree Technology, and reached an agreement with it to jointly develop the AI model for humanoid robots.

From Unitree to Winverse, DeepSeek has continuously invested funds in the robotics industry. Shi Yunlei, founder of Embodied Robotics, told Pencil News that DeepSeek is likely targeting two core points: "the learning capability in the physical world, and the new industrial entry point."

- 01 - Tsinghua PhD, Left Tencent to Launch Startup

Who is Hu Han?

His resume is very outstanding: he holds a doctorate degree from Tsinghua University, is a former principal researcher at Microsoft Research Asia, former scientist at Tencent Hunyuan, and the winner of the Best Paper Award at the top international computer vision conference.

In 2004, Hu Han was admitted to the Department of Automation of Tsinghua University, where he studied for 10 years and obtained his doctorate degree in 2014. During his PhD study, he also went to the University of Pennsylvania in the United States as a visiting scholar, studying under computer vision scholar Shi Jianbo.

After graduation, Hu Han joined Baidu Deep Learning Institute. In December 2016, he joined Microsoft Research Asia, and gradually grew into an important researcher in the field of visual computing.

It was a paper that made Hu Han really famous.

In 2021, he and his team published Swin Transformer, trying to make the Transformer architecture process images more efficiently. Before that, this architecture had shone brilliantly mainly in the field of natural language processing.

This research quickly attracted widespread attention in the academic circle.

In that year, Swin Transformer won the Best Paper Award of International Conference on Computer Vision (ICCV) — the Marr Prize. The conference received 6152 papers in total, only 1612 of which were accepted, and finally only one paper won the Best Paper Award.

Hu Han is the co-first author and the only corresponding author of this paper.

After that, Hu Han participated in researches including Swin Transformer V2 and SimMIM, exploring how to train larger and more efficient visual models. In 2022, he was also selected into the "35 Innovators Under 35" list of the MIT Technology Review China region.

It is worth mentioning that there is no lack of intersection between Hu Han and DeepSeek.

Looking through his paper list, you will find a familiar name: Xie Zhenda.

During their time at Microsoft Research Asia, the two co-authored papers including PixPro, Swin Transformer V2 and SimMIM. Among them, SimMIM was accepted by CVPR, the top computer vision conference, in 2022, with Xie Zhenda as the first author and Hu Han as one of the paper authors.

Xie Zhenda later joined DeepSeek, becoming a basic model researcher there and participating in the R&D of multiple generations of DeepSeek models.

After leaving Microsoft, Hu Han joined Tencent Hunyuan Large Model team.

At Tencent, he not only participated in the R&D of large language models, but also took charge of directions such as vision-language models and embodied foundation models.

In April 2026, Tencent released the research results of the HY-Embodied-0.5 embodied foundation model, and both Hu Han and Shun-Yu Yao, Chief AI Scientist of Tencent, appeared on the paper author list.

The problem this research aims to solve is how to enable AI to understand the spatial relations and object movements in the real world, and provide reasoning capabilities for robots to complete tasks.

However, only three months later, Hu Han chose to leave.

On July 23, Intelligent Emergence under 36Kr reported that Hu Han, head of Tencent Hunyuan Multimodal Understanding, submitted his resignation to start a business. Before that, he had been transferred to the Frontier Technology Research Group under the Large Language Model Department, in charge of multimodal understanding research and reporting to Shun-Yu Yao.

His position was later taken over by Tian Yonglong, a former OpenAI researcher.

On the same day Hu Han resigned, Tencent announced the adjustment of Hunyuan organizational structure, merging the large language model department and the multimodal model department to establish a new Basic Model Department, which is uniformly managed by Shun-Yu Yao.

Hu Han embarked on the road of starting a business.

However, Winverse AI is still quite mysterious so far. The outside world has no clear information about its team size, product form and commercialization progress.

Foreign media described it as a company that develops robot AI. But Phoenix Tech learned from a number of industry insiders that Winverse may not be a strict embodied intelligence company, but more inclined to underlying infrastructure such as model acceleration.

- 02 - DeepSeek Starts to Invest in Robotics

Liang Wenfeng has shown strong interest in robotics recently.

On October 12, Phoenix Tech cited reports from multiple insiders that Liang Wenfeng has intensively contacted embodied intelligence entrepreneurs recently, discussing directions such as the interaction between robots, large models and the real world.

DeepSeek has considered two paths internally: one is to form its own R&D team, and the other is to invest in embodied intelligence enterprises. Judging from the transactions that have taken place, the latter is being promoted.

In August this year, Unitree Technology was listed on the Shanghai Stock Exchange. DeepSeek participated in the strategic placement with 140.8 million yuan, obtained 933,399 shares, and signed a cooperation arrangement with Unitree.

The two sides agreed that Unitree will give priority to DeepSeek when purchasing AI development services; when DeepSeek needs robots and embodied intelligence solutions, it will also give priority to Unitree.

One company is good at manufacturing robots, and the other is good at training large models. The two companies are trying to combine their respective advantages.

Now, Winverse AI has entered DeepSeek's investment vision.

Behind this, there is another important change: DeepSeek itself has begun to have more sufficient funds.

On October 6, it was reported that DeepSeek is pushing forward a new round of financing with a scale of at least 800 billion yuan. It originally planned to raise about 500 billion yuan, but the scale was increased because investors' subscription enthusiasm exceeded expectations.

Tencent and CATL are important investors among them. Reports say DeepSeek is preparing for a possible IPO in early 2027.

This round of financing has not been finally completed, but the use of funds has attracted attention.

Part of the funds will be used for model training and data center construction, and the other part may flow to AI startups to expand technical capabilities through investment and joint R&D.

From a laboratory focusing on large model R&D to starting to invest in upstream and downstream startups, DeepSeek's role is changing.

Why choose robotics?

Shi Yunlei, founder of Embodied Robotics, analyzed to Pencil News that an important reason is that large models are looking for new data and application scenarios outside the Internet.

In the past, large models learned from web pages, books, codes and videos, and were good at answering questions and generating content. But even the smartest model may not be able to stably control a robot to pick up a cup.

Because there are too many uncertainties in the real world.

The cup may slip, the table may be moved, and the weight, material and position of objects are all changing. Robots not only need to understand instructions, but also must adjust their actions according to the actual situation.

Shi Yunlei said: "The hardest part is to transform the understanding and reasoning at the language level into a stable and generalizable execution capability in the physical world."

This requires a large amount of real operation data.

In the past, the training materials of large models mainly came from the content already created by human beings. Robots need to walk into factories, warehouses and homes by themselves, continuously perform tasks, and accumulate experience of success and failure.

Every operation of the machine may become the material for the next round of training.

In October this year, British media reported that a number of enterprises that specifically provide training environments for robots have emerged overseas. Companies such as Generalist, Neura and Skild have built facilities similar to "robot gyms", allowing robotic arms to repeatedly practice grasping, carrying and sorting items to collect operation data.

Other enterprises deploy data collection equipment in different countries, allowing human operators to perform tasks repeatedly, so as to train robots.

These are all fields that large model enterprises were relatively unfamiliar with in the past.

Shi Yunlei believes that another opportunity DeepSeek values is the new industrial entry point. Robots with autonomous execution capabilities are expected to enter the manufacturing, logistics, elderly care and other service industries.

Embodied Robotics completed a 100 million-yuan-level Pre-A round of financing this year, and is promoting the elderly care robot Qijia Q1 to enter real elderly care scenarios.

According to Shi Yunlei, the company has collected more than 1000 hours of operation data in elderly care institutions and families. As of July this year, the robot has been continuously operated and verified for more than 45 days in some elderly care institutions in Shanghai and the Beijing-Tianjin-Hebei region.

But Shi Yunlei believes that the more obvious shortboard of current robot commercialization is "high-quality operation data, reliable hardware and real scenarios".

Simply installing a stronger large model into a robot cannot solve all problems. This may explain why DeepSeek tends to enter the field through investment.

Unitree already has robot ontology, manufacturing and motion control capabilities, and the Winverse team has model research experience. Compared with building a complete robot industrial chain from scratch, entering a new field through investment and cooperation may be more efficient.

- 03 - The Robotics Track Welcomes a Group of New Investors

It is not only DeepSeek that is becoming interested in robotics.

Global capital is pouring more and more funds into companies that specifically develop robot intelligent systems (the "brain" of robots).

In January this year, American robotics AI company Skild AI announced the completion of a 1.4 billion US dollar financing, with a valuation of more than 14 billion US dollars.

The lead investor was SoftBank, with NVIDIA, Bezos's investment institution and others participating.

Skild AI was founded in 2023, and its founders come from robotics and artificial intelligence research fields of institutions such as Carnegie Mellon University. It hopes to develop a general robot model that can be used by robots of different brands and different forms.

This business has begun to generate revenue.

In September this year, Deepak Pathak and Abhinav Gupta, co-founders of Skild AI, publicly disclosed that the company's annualized recurring revenue has exceeded 100 million US dollars.

It only took 10 months from the first commercial deployment.

According to the data disclosed by the company, Skild has more than 60 paying customers, covering warehouses, factories, data centers, logistics distribution, security, catering services and other fields.

Skild also cooperates with NVIDIA and Foxconn to deploy its robot intelligent system on the robotic arms of the Blackwell server production line, trying to automatically complete high-precision assembly tasks.

Another star company, Physical Intelligence, is also attracting large amounts of capital.

This company was founded in 2024, and its team comes from institutions such as Google DeepMind, Stanford University and the University of California, Berkeley.

In November 2025, it obtained 600 million US dollars in financing, with a valuation of 5.6 billion US dollars. By March 2026, Bloomberg reported that the company is negotiating a new financing of about 1 billion US dollars, with a proposed valuation of more than 11 billion US dollars.

Its research direction is also general robot intelligence: let one set of models learn multiple actions, and then adapt to different robots.

From Skild AI to Physical Intelligence, and then to Winverse AI, which was established only a few months ago, capital is not only betting on robot ontology, but also on the underlying models, data and software infrastructure behind it.

However, there are obvious differences between robot AI and traditional software business.

Software can be copied quickly, but when robots enter the real world, they need hardware, sites, maintenance personnel and long-term testing.

For example, in an automobile factory, even if the robot has a 99% grasping success rate, the remaining 1% failure may lead to production interruption. Even if the success rate is very high, if each operation takes too long, it still cannot meet the production rhythm.

Shi Yunlei believes that after large model enterprises enter the market, "the differentiation of simple integration and demonstration will be narrowed".

In the future, the real value of a robotics startup will lie in task success rate, manual intervention frequency, fault recovery time and overall usage cost.

Embodied Robotics' own business also reflects this change.

On the one hand, the company lets Qijia Q1 receive real scenario tests in elderly care institutions, and on the other hand, it promotes the relatively mature "Gewu" rehabilitation robot product line to the market.

According to Shi Yunlei, Gewu has achieved mass production of 1000 units and has been exported overseas.

But for capital, the robotics industry has another layer of attraction. Over the past decade, the main battlefields of Internet entrepreneurship have been traffic, advertising, e-commerce and software. After the rise of large models, capital began to invest in computing power, data centers and foundation models.

Now, investment is further extending to the real world.

NVIDIA invests in robotics AI companies, SoftBank bets heavily on general robot intelligence, Google develops Gemini Robotics, and DeepSeek has successively chosen Unitree and Winverse.

A new group of industrial capital is entering the robotics track.

This article does not constitute any investment advice.

This article is from the WeChat official account "Pencil News" (ID: pencilnews), author: Song Ge, editor: Huang Xiaogui, published with authorization from 36Kr.