Liang Wenfeng goes down the mountain, what is DeepSeek looking for?
Key Takeaways Shifting from focusing solely on model development to intensively reviewing investment projects, what is Liang Wenfeng thinking about?
The reclusive master is stepping down from seclusion.
Recently, foreign media exposed the full details of DeepSeek's second round of financing, which has raised a total of 1055 billion RMB to date, with Tencent and CATL as lead investors. Having previously stuck to the stance of "we have sufficient capital and will not accept external financing", Liang Wenfeng changed his past posture by launching two consecutive rounds of financing within half a year.
After stepping out of seclusion, the master has also begun to set his sights on areas beyond large models. It is understood that Liang Wenfeng has intensively met with embodied intelligence entrepreneurs, and spent a large amount of time studying the embodied intelligence field.
Based on his phased learning outcomes, DeepSeek has started to make exploratory investments with its capital. At present, DeepSeek has confirmed investment plans for two projects, namely embodied brain company Winverse and embodied motion control company Ligewanwu.
Considering the increasingly close connection between large models and embodied intelligence, these two rounds of investment may only be the beginning of DeepSeek's layout in the embodied field, with a larger industrial landscape likely behind it. The reason for planning such a landscape is related not only to the general environment, but also to DeepSeek's own situation.
Starting to calculate long-term accounts
Before the summer of 2026, Liang Wenfeng's public image on the internet was similar to a hidden master in the AI industry: he did not take external investment, did not touch the most popular embodied intelligence in recent years, and DeepSeek's pricing was so low that it was almost like running a charity. As a result, he is often called "Saint Liang" in developer communities, and people half-jokingly say "his kindness can never be repaid".
However, within this half year, DeepSeek has successively pushed forward two rounds of financing, completely changing its past operation mode of relying on internal funds and focusing on model research.
In June this year, DeepSeek completed its first round of financing at a scale of about 51 billion RMB, led by Liang Wenfeng himself, the CATL ecosystem and Tencent. Among them, Liang Wenfeng personally contributed about 20 billion RMB, making him the largest single investor; Tencent contributed about 10 billion RMB, and the CATL ecosystem contributed about 5 billion RMB.
Shortly afterwards, DeepSeek launched its second round of financing, with the scale raised from the initial 500 billion RMB to over 1 trillion RMB. The list of investors was further expanded to cover chip, automotive, consumer electronics and internet enterprises. In addition to Tencent and the CATL ecosystem each contributing 10 billion RMB, other industrial investors include ChangXin Memory Technologies, SMIC, vivo, Meituan, Geely, Midea Group, etc. The capital supporting DeepSeek's development has expanded to more industrial investor groups.
Behind this is the requirement for investment scale and sustainability brought by the competition of cutting-edge models.
In the past, the most noteworthy thing about DeepSeek was how it trained models with higher efficiency and provided capabilities to the market at more cost-effective prices. But after a breakthrough in one training session, the company still needs to bear two types of costs at the same time: one is for exploring next-generation models, where the results and returns are full of uncertainty; the other is for supporting already launched services, which requires continuous provision of stable and available inference capabilities. R&D will not stop just because the product is launched, and services cannot be suspended for the next training session.
Higher technical efficiency can reduce the cost of a single task, but it does not necessarily lead to a decline in the company's total investment. When model capabilities expand and service types increase, the saved resources may be quickly consumed by new demands. For DeepSeek, the test of long-term competition lies in whether it can keep capital, computing power and revenue in pace with the R&D rhythm while making continuous investments.
In the second half of this year, DeepSeek's adjustments to its charging methods have already revealed such operational considerations.
In mid-August, when the official version of V4-Pro was launched, peak-valley pricing was introduced for APIs, and the overall price was raised. Adopting different prices for peak and off-peak hours can guide time-insensitive demands to use services at staggered times, so that limited computing power can undertake more tasks. By November 1st, the newly launched Search and Fetch APIs charge for successful requests, and DeepSeek has established a charging entry for search and crawling capabilities beyond models.
From model invocation to supporting tools, charging has begun to cover more service links. In the past, low prices helped DeepSeek quickly enter the selection range of developers; next, how to turn these usages into revenue sufficient to support continuous supply will increasingly affect its operation arrangements.
Along with revenue, another issue that needs to be solved is where the computing power comes from.
The report released by a16z in October shows that the rental prices of NVIDIA H200 and B200 have risen significantly compared with the beginning of the year. For enterprises that continuously train and deploy models, the acquisition cost, supply rhythm and available scale of computing resources will all affect when the next-generation product can be launched. Even with sufficient capital on the books, enough chips and software systems that can make chips operate efficiently are still needed.
DeepSeek promotes the adaptation of domestic computing power, which is exactly expanding its options in this regard.
Recent media reports say that DeepSeek's next-generation model V5 is planned to carry out pre-training based on Ascend chips. Previously, Ascend chips only undertook the inference tasks of DeepSeek models. In the V4 stage, the two sides have completed a large number of underlying operator adaptations, and V5 will be the first large model natively trained based on Ascend 950, which is scheduled to be launched next year.
Looking at the shareholder lists of the two rounds of financing, the changes have another meaning. In addition to Tencent and the CATL ecosystem, the entry of enterprises such as ChangXin Memory Technologies, SMIC, vivo, Geely and Midea has established capital ties between DeepSeek and storage chips, terminals and manufacturing industries. These enterprises are distributed in different links of model supply and application, and may carry out more industrial linkage with DeepSeek in the future.
What is worth noting here is ChangXin Technology. In fact, when ChangXin conducted its IPO earlier, institutions under Liang Wenfeng, such as High-Flyer Quantitative Investment, appeared in ChangXin's placement list as a group to participate in the new share subscription. The statement "you invest in my AI, and I invest in your chips" shows that ChangXin and DeepSeek have actually carried out deep ecological binding.
Why bet on embodied intelligence?
After its capital reserve expanded, DeepSeek also began to look for new investment directions beyond models. Embodied intelligence is one of the noteworthy choices.
Liang Wenfeng first participated in the share placement of Unitree, then intensively contacted embodied entrepreneurs, and DeepSeek later appeared on the investor list of two early-stage companies.
At first, Liang Wenfeng participated in the placement of Unitree through three entities: DeepSeek, Ningbo High-Flyer Quantitative Investment and Zhejiang Jiuzhang Asset, which was widely regarded as a financial investment. This is because Liang Wenfeng once said "the embodied industry is still too immature".
But after that, he began to meet with some embodied intelligence entrepreneurs and started intensive learning as a layman. Soon, the two projects invested by DeepSeek surfaced: Winverse AI and Ligewanwu.
Hu Han, the founder of Winverse AI, once led the visual language model and embodied intelligence basic model team of Tencent Hunyuan, and his entrepreneurial direction is embodied brain and world model. Yu Tao, the founder of Ligewanwu, once engaged in reinforcement learning research and led the team at Dynamic Robotics to build a whole-body control system. At present, his company is exploring the combination of large language models and embodied intelligence.
Bloomberg reported on October 12 that DeepSeek led the seed round investment of Winverse AI with an amount of tens of millions of US dollars, and the post-investment valuation was about 400 million US dollars, with Meituan Longzhu participating in the investment; the first round of financing of Ligewanwu was completed in the form of a "friends and family round", and the investors include HSG, Lisi Capital and DeepSeek.
From the perspective of technical direction, the two projects focus on the key issues of robots understanding the environment and executing actions respectively. They are complementary to Unitree's ontology business, and also correspond to the capabilities that large language models need to make up for when entering the physical world.
In his communication with investors in July 2026, Liang Wenfeng presented a six-stage roadmap for AGI: Language Model → Chain of Thought → Agent → Continuous Learning → Self-Iteration → Embodied Intelligence. DeepSeek's continuous promotion of Agent iteration in the second half of the year and its actions to penetrate into the embodied field are exactly in line with this vision.
On August 13, when the official version of V4-Pro was released, DeepSeek open-sourced the agent operation framework Harness. On October 31, DeepSeek's Harness team even disclosed the Super-agent design prototype, enabling Agent to evolve from a tool waiting for instructions to a collaborator that can perceive the working environment, execute tasks continuously, and ask humans for help only when necessary.
This series of actions is strengthening the autonomous execution and continuous operation capabilities of AI, so that AI can first learn to "do things by itself and keep doing things" in the digital world, and will not be at a loss when entering the physical world.
With one company focused on robot brain, another focused on robot motion control, plus Unitree that develops robot ontology, as well as its own large language model and Agent, DeepSeek's investment has included all key links of embodied intelligence from perception and understanding, motion control to hardware carrier in the same layout.
From this perspective, DeepSeek's investment in the embodied field may serve two goals at the same time: to find application space for existing models, and to access the problems required for next-stage model research. The former is related to who will use DeepSeek's capabilities, while the latter is related to how these capabilities will evolve in the future.
When it comes to the combination of large models and embodied intelligence, DeepSeek is not the only player. Its rivals across the ocean are also carrying out systematic operations.
For example, OpenAI established OpenAI Robotics in June this year, led by Aditya Ramesh, the core developer of Sora. Its official website has posted 27 robot-related positions, and it is expected to launch a demonstration prototype in 2027.
Can embodied intelligence feed back to DeepSeek?
DeepSeek is mobilizing resources for longer-term competition. But whether this layout can be established ultimately depends on one result: whether the newly added capital, computing power options and external cooperation can help it continue to improve its model capabilities.
Investing in embodied companies is only the starting point of this process. DeepSeek needs to find specific tasks with partners, let models participate in tests, and then convert failures in execution into analyzable and trainable feedback. After the model is improved, it needs to go back to the robot to verify the effect. Only when this process can operate continuously, can the equity relationship be transformed into technical accumulation.
At present, how close this connection is remains the most noteworthy part of DeepSeek's layout.
Invested companies have their own R&D goals, and they will also select models according to task performance. For example, Unitree is currently trying to deploy OpenAI models, which shows that the technical choices of robotics enterprises are always open.
Not long ago, Wang Xingxing revealed at the Industry Fair that the test effect of GPT-6 deployed on Unitree robots is good, and it is expected that after the release of GPT-7 next year, the current bottleneck of embodied intelligence will be further broken through.
On September 3, OpenAI released GPT-6 Astra; in simulation tests, GPT-6 Astra directly acts as an action strategy, achieving an average success rate of 22.48% in 42 robot simulation operation tasks and 2100 tests, surpassing 40 public robot strategies in the evaluation.
The capability demonstrated by leading large language models to swallow the embodied base is actually the source of a certain anxiety for DeepSeek. For DeepSeek, the most important thing is not to widen the gap with cutting-edge models. Embodied intelligence may not only be its second growth curve, but also a technical defense battle.
At the same time, large models and embodied intelligence are two mirrored commercialization paths, both of which have the problems of excessive early investment scale and immature commercialization.
Early investment allows DeepSeek to gradually understand the industry and decide whether to dig deeper according to progress. But as joint R&D, computing power support and personnel participation increase, the cost of exiting or adjusting direction will also rise. If the investment expands faster than technical verification, the company may face the pressure of capital and R&D on its main model business while waiting for returns from embodied cooperation. The exploration that was originally intended to expand the capability boundary may also slow down the iteration of existing capabilities.
After "stepping down from seclusion", DeepSeek has obtained more resources, and also faced more directions worthy of investment. The real test is to make these directions form accumulation around model capabilities. If cooperation continues to increase but technical returns are always vague, the hundreds of billions of financing may only prolong the waiting time.
This article is from the WeChat official account "Baobian" (ID: baobiannews), written by Zhang Jingwei, and authorized for release by 36Kr.