The ceiling of language models, and Liang Wenfeng's next bet

Liang Wenfeng has been meeting people frequently recently.

The people he meets are all entrepreneurs in the embodied intelligence sector. This matter itself is not complicated, what makes it special is the timing of its occurrence. DeepSeek has just completed a financing round of at least 800 billion RMB, pushing its valuation to a new record high, and the plan for its listing in early 2027 has been put on the agenda. At this juncture, the founder's intensive engagement with a field he is not familiar with points to one fact: the growth story of large language models may be approaching its end.

A Distinctive Investment Signal

Liang Wenfeng's first public investment in the embodied intelligence sector goes to Unitree Robotics.

Shortly before Unitree's listing, he purchased approximately 2% of its shares for 141 million RMB, and the two sides plan to cooperate on embodied intelligence models. The amount of this investment is not large, but the signal it sends is clear. Unitree is one of China's most mature humanoid robot hardware companies, which completed a 900 million USD IPO in August. Liang Wenfeng's decision to bet on this target is equivalent to acknowledging one point: having only a "brain" is not enough, an intelligent system also needs a physical "body".

What deserves more attention is his second investment. DeepSeek plans to lead the seed round of Winverse AI, with a post-money valuation of around 400 million USD. This company was founded by Hu Han, a former Tencent AI scientist, and has been established for only a few months. Hu Han has a very special background — he does not work on robot hardware, and he previously led Tencent's team responsible for vision-language and embodied foundation models. In other words, he is a professional who works on the "brain" part of AI systems.

However, industry insiders have a calm judgment on this investment: strictly speaking, Winverse AI is not an embodied intelligence company. Its business is more focused on underlying infrastructure, with a specialization in model acceleration technology. This judgment is of great importance. It shows that Liang Wenfeng is not blindly chasing the current trend, and he is investing in areas he fully understands — model efficiency, training acceleration, and underlying architecture, which are exactly DeepSeek's core competencies.

GPT-6 Forces All Players to Recalculate Their Strategies

The direct trigger for DeepSeek's shift to embodied intelligence at this moment comes from OpenAI.

GPT-6 Astra has achieved a 95% success rate in robot control tasks, while its predecessor Fable 5.1 only had a 40% success rate. This leap is not a gradual improvement, but a change of order of magnitude. Multiple test demonstrations show that GPT-6 has delivered industry-remarkable execution performance in the embodied intelligence field.

Liang Wenfeng recently publicly admitted that DeepSeek still lags behind GPT-6 in certain capabilities. This statement carries far more weight than it appears on the surface. A founder known for his technical idealism publicly acknowledging the gap means that the internal anxiety about the company's roadmap is real. People are re-examining the question: can simply scaling up large language models maintain a competitive edge in the AGI race?

What is more alerting is that OpenAI is also separately training a dedicated embodied intelligence foundation model. This means GPT-6 may only be a starting point. Even though it is already impressive enough, it still takes 40 minutes to complete the action of picking up a cup — while the current industry requirements for embodied intelligence are still at a very low level.

The Difficulty of Embodied Intelligence Does Not Lie in Hardware

Many people understand embodied intelligence as "installing a robot body for AI". This understanding is far too simplistic.

The real core problems of embodied intelligence are: how to make the model understand space? How to make robots learn continuous actions? How to make artificial intelligence continuously revise its decisions through real-world feedback? Each of these three problems is far more complex than the challenges faced by large language models. Language models process symbols and probabilities, while embodied intelligence deals with physical contact, torque control, and real-time feedback.

China is in a world-leading position in the field of humanoid robot hardware manufacturing. Unitree Robotics' IPO has further consolidated this advantage. However, leading in hardware does not equal leading in intelligence. To make robots intelligent enough to respond to voice commands and solve real-world problems, a new generation of artificial intelligence systems is required, which is completely different from the existing models of ChatGPT or DeepSeek.

American companies have taken earlier steps in this direction. Skild AI and Physical Intelligence are developing general-purpose robot brains, with valuations as high as 140 billion USD and 5.6 billion USD respectively, backed by NVIDIA and Alphabet. They do not manufacture hardware, and focus exclusively on the "brain" part. This division of labor is very clear: China makes the "bodies", while the United States makes the "brains". What Liang Wenfeng aims to do is to evolve DeepSeek from a language model "brain" into a "brain" that can command physical bodies.

Two Paths, Two Types of Risks

DeepSeek has two optional paths to enter the embodied intelligence sector.

The first path is to build a dedicated team directly. This path allows maximum control over the technical roadmap, but requires adding brand new capabilities in robot hardware, motion control and data collection. DeepSeek's strengths lie in algorithms, model training and engineering efficiency, while embodied intelligence requires a large amount of real-world data and long-term adaptation between models and hardware. This is not a problem that can be solved simply by recruiting a few more people.

The second path is to enter the sector through investment or strategic cooperation. This path is more in line with the current division of labor logic of the AI industrial chain: model companies provide the "brains", and robot companies provide the "bodies". For large model companies, their march into embodied intelligence is not only to manufacture robot products, but also to seize the key entry point for future artificial intelligence capabilities. Liang Wenfeng's current moves in this direction — investing in Unitree and leading the investment in Winverse AI — show that he tends to choose the second path.

However, the second path has a hidden risk: it relies on the technical roadmap and commercialization pace of the invested companies. If the development of Unitree or Winverse AI fails to meet expectations, DeepSeek's embodied intelligence strategy will fall into a passive position. Liang Wenfeng has always been known for prudence. Compared with some large companies that quickly launch applications to seize market share, DeepSeek is more focused on underlying technological breakthroughs. This style was an advantage in the language model era, but it remains to be seen whether it will still hold true in the embodied intelligence era.

Talents Like Hu Han and a New Talent Migration

Hu Han belongs to the new generation of Chinese AI researchers.

They left large technology companies to found their own professional labs, focusing on world model research in fields such as autonomous driving, interactive video and robotics. Other researchers who founded their own labs include Justin Lin, former head of Zhiyuan Technology under Alibaba, and Tian Keyu, a former intern at ByteDance. The common feature of this group of people is that they have accumulated sufficient technology and resources on large platforms, and then choose to place independent bets on more vertical tracks.

This talent migration is both an opportunity and a challenge for DeepSeek. The opportunity lies in the fact that these labs may become DeepSeek's investment targets or cooperation partners. The challenge is that if Liang Wenfeng does not develop this business on his own, these talents may be poached by other large companies or capital players.

For Liang Wenfeng, who has long been bullish on AGI, embodied intelligence may be a problem that DeepSeek has to solve in its next development stage. From language models to intelligent agents, and then to robots capable of taking physical actions, the next battle in the artificial intelligence industry may no longer be just the competition between different models, but a contest of who can truly bring intelligence into the real world.

Liang Wenfeng has started looking for a "body" for his AI system. But it remains unanswered whether what he finds will be a new growth curve, or a new battlefield that he is not good at.

This article is from the WeChat official account "Tech Business", authored by Tech Business, and authorized for release by 36Kr.