DeepSeek is about to get another huge round of financing.
On October 22, multiple foreign media reported, citing people familiar with the matter, that DeepSeek’s latest round of financing raised approximately 100 billion yuan, becoming one of the largest private equity financing transactions in China. CATL, Tencent and other companies are among the major investors; this round of financing is expected to be completed in the next few days, and specific transaction terms may still be adjusted.
The scale of this financing far exceeds the original plan.
According to reports, DeepSeek raised its financing target at least twice during negotiations with investors, from the initial 50 billion yuan to 80 billion yuan, and finally reached about 100 billion yuan.
At the end of June this year, DeepSeek had just completed its first round of external financing, with a scale of more than 50 billion yuan.
If the above news is finally confirmed, within half a year, DeepSeek’s two rounds of financing will total more than 150 billion yuan.
China News Weekly asked DeepSeek for confirmation, but as of press time, had not received a reply.
Picture/picture insect creativity
密集融資
More than two years ago, DeepSeek had no plans to raise funds.
In July 2024, DeepSeek founder Liang Wenfeng was asked in an interview whether the company had a financing plan. He replied: “There is no financing plan in the short term. The problem we face has never been money, but the embargo of high-end chips.”
When asked whether artificial intelligence research and development requires larger capital investment, Liang Wenfeng said: “More investment does not necessarily produce more innovation. Otherwise, large manufacturers can take over all innovations.”
At that time, DeepSeek relied on quantitative private equity and had accumulated a certain amount of computing power and financial resources. Unlike other large-model startups that rely on external financing to support research and development, DeepSeek has mainly relied on itself and related parties to provide financial support in the long term.
But this year, things have changed.
According to a financial report on October 13, DeepSeek launched its first round of external financing in April this year and completed delivery at the end of June. The financing scale exceeded 50 billion yuan and the company valuation exceeded 350 billion yuan.
The first round of financing had just ended, and the second round followed.
In July this year, DeepSeek launched a new round of financing, with a pre-investment valuation of approximately 500 billion yuan. It was initially planned to raise 50 billion yuan, but the scale has since been increased, and the planned financing scale has been expanded to about 100 billion yuan.
At the same time, DeepSeek was also revealed to be preparing to be listed on the Science and Technology Innovation Board. Financial reports show that the company has previously contacted a number of investment banks and plans to submit a listing application as soon as the end of this year, with the goal of listing in 2027. This plan has also not been officially confirmed.
However, accepting external financing does not mean that DeepSeek is willing to hand over more control.
In June this year, foreign media reported that DeepSeek adopted a relatively special transaction structure for its first round of financing. Except for individual investors, funds mainly enter through limited partnerships managed by Liang Wenfeng, rather than investors directly holding equity in DeepSeek’s main company. Relevant investors do not have voting rights and are subject to a five-year lock-in period.
What’s even more special is that Liang Wenfeng himself is said to have invested approximately 20 billion yuan in this round of financing, becoming the largest single investor.
It is not yet known whether the latest round of financing will continue this arrangement.
From not rushing to raise funds to intensive financing, why did DeepSeek’s capital strategy suddenly change?
You still have to invest in computing power
At the beginning of 2025, DeepSeek shocked the world with a number: US$5.576 million.
This is the GPU computing cost of the formal training phase of V3 calculated in its DeepSeek-V3 technical report based on the assumed rental price of NVIDIA H800 GPU of $2 per hour.
This figure has triggered a re-examination of the development costs of large models. In the past, hundreds of millions of dollars in training investments were seen as an important threshold for competition in leading large models, but DeepSeek has proven that algorithm innovation and engineering optimization can also significantly improve the efficiency of computing power usage.
For a time, low-cost training was even interpreted by some public opinion as saying that large model research and development no longer requires huge financial investments.
But just because it costs less to train a model once, it doesn’t mean that it’s just as cheap to develop the model and let it run continuously.
DeepSeek clearly pointed out in the technical report that the US$5.576 million only covers the GPU computing expenses in the formal training phase, and does not include previous research experiments around model architecture, algorithms and data, nor does it represent the company’s entire R&D and operating expenses.
Zhang Xiaorong, president of the Deepin Technology Research Institute, said that the outside world actually misinterpreted DeepSeek’s efficiency advantages in specific links as the low-cost operating capabilities of the entire company.
In his view, DeepSeek still faces computing resource constraints. If more sufficient computing power can be obtained, there is room for further improvement in model performance and service capabilities.
Especially as large models move from technology research and development to large-scale application, the computing power needs of enterprises are also changing.
In the past, the outside world paid more attention to how much computing power is needed to train a large model; but after the model is put into use, it still needs to continue to provide inference services to users. The more users and the more frequent the calls, the more computing resources are usually consumed.
Zhang Xiaorong pointed out that for the current leading large model companies, inference services, computing power infrastructure, inference framework optimization and energy security are becoming important directions for capital investment. As applications scale, ongoing inference expenses may become an increasingly important part of the overall cost.
What’s more noteworthy is that technological progress may not necessarily reduce the total demand for computing power.
Xiaorong Chang explained this phenomenon with the “Jevons Paradox”: when the efficiency of use of a resource increases and the unit cost decreases, it may stimulate more demand and promote an increase in total resource consumption.
Specific to the field of AI, algorithm optimization makes a single model call cheaper, which may attract more users, generate more application scenarios, and bring greater Token consumption. Ultimately, even as unit computing costs continue to decrease, enterprises’ overall computing power and power needs are likely to grow.
In other words, DeepSeek has previously demonstrated that it can use computing power more efficiently, but this does not mean that it no longer needs more computing power.
誰在押注
As SpaceX goes public, OpenAI plans to go public within the year, Anthropic launches its IPO process, and global technology giants continue to expand their investment in computing power, the valuation and return on investment issues of AI companies have attracted more and more attention.
Two large domestic model listed companies are also experiencing valuation corrections. On October 22, MiniMax’s stock price fell below the issue price for the first time, falling approximately 88% from the high point in March this year; on the same day, Zhipu closed at HK$460.80, down approximately 85% from the historical high point in June.
Just as the stock prices of two large listed model companies have fallen sharply, DeepSeek, which has not yet been listed, was revealed to have raised approximately 100 billion yuan.
Judging from the list of investors currently disclosed, the participants include not only professional investment institutions, but also many industrial capitals.
According to reports, CATL, Tencent and other companies are the main investors in DeepSeek’s current round of financing. In addition, Meituan, Geely, Midea, GigaDevice and other companies were also exposed to be involved.
Why is there still so much capital willing to bet on DeepSeek?
According to Liu Chunsheng, associate professor at the Central University of Finance and Economics, DeepSeek’s technical strength and open source route are important foundations for attracting capital. For industrial capital, the value of investing in DeepSeek goes beyond financial returns.
For example, Tencent can use this to further improve its AI and cloud computing ecosystem; CATL may explore computing and power synergy opportunities in areas such as computing power supply and energy storage, while using large model technology to promote the digital upgrade of its own industry chain.
Financial commentator Guo Shiliang also believes that in addition to financial returns, industrial capital participation in investment may also be based on considerations such as resource allocation and complementary advantages.
“Industrial capital is optimistic about ecological position and synergy.” He said.
On the other hand, the continued increase in capital is expected to provide DeepSeek with more industrial resources besides money to participate in the global AI competition.
Zhang Xiaorong believes that DeepSeek has previously relied on the efficiency advantage formed by algorithm innovation to improve the unit economic model, but it is temporarily unable to eliminate the capital and computing power scale gap between it and international giants such as OpenAI and Anthropic. “For some time to come, competition will rely on both capital efficiency and capital scale, with scale determining the upper limit of service capacity.”
However, more investment does not necessarily mean that better technical results will be achieved.
Guo Shiliang believes that the real test of large-model enterprises in the future will still be technology iteration capabilities and cost management capabilities. Whoever can continue to improve technology while effectively controlling costs will have a better chance of consolidating its competitive position in the industry.
More than two years ago, Liang Wenfeng said that more investment does not necessarily produce more innovation.
Nowadays, DeepSeek is about to receive hundreds of billions of yuan in financing, and this sentence has new practical significance.
For DeepSeek, perhaps more important than raising 100 billion yuan is how to spend the 100 billion yuan well.
References:
“Revealing DeepSeek: A more extreme story of Chinese technological idealism”, July 2024, 36 Krypton
Reporter: Shi Hanxu
Related Reading
- Original Zhang Hongchao’s Mixue Bingcheng is quietly turning into a supply chain company2026-10-24
- Original Brokerage firms cut salaries, down to investment banking department2026-10-24
- Hongguo, which has a monthly activity of 368 million, no longer needs “wool users”?2026-10-24
- Behind the five reminders, what calculation did Starbucks make?2026-10-24
- U.S. deficit nears $2 trillion as government spending keeps climbing2026-10-17
