DeepSeek, finally took on this 50 billion

In the past two years, DeepSeek has been an outlier in the AI circle in China and even around the world.

While other large model companies are busy raising funds, increasing their valuations, and telling stories to investors, DeepSeek doesn't participate in any of these. Its funds come from Magic Square Quantitative, a quantitative private equity fund founded by Liang Wenfeng, which had an annualized return rate of 56.55% in 2025 and managed assets of over 70 billion yuan. All R & D funds come from here. It doesn't need external capital and doesn't need to explain to any investors why there is no commercial progress this quarter.

This logic has led to a well - known "hidden master" scenario: In January 2025, DeepSeek - R1 was released, sweeping various evaluation lists and causing the US stock technology sector to fluctuate for a whole day. But after that, Liang Wenfeng still didn't raise funds. The outside world's impression of this company has gradually solidified into one word: "not short of money".

In May this year, this impression began to waver. The media started reporting that DeepSeek was in contact with external investors, and the rumored amount was getting larger. The news fermented in the circle for nearly two months, and all parties were speculating about who had invested, who hadn't, and what the valuation was.

Recently, foreign media The Information reported that DeepSeek has officially completed its first - round financing, raising a total of over 50 billion yuan (about 7.4 billion US dollars). Currently, neither DeepSeek nor the investors have responded to this financing information. If the final official announcement is true, it will set a new record for the highest single - round financing in China's AI industry so far.

This is the same Liang Wenfeng and the same DeepSeek. Only this time, he decided to accept external funds.

Where does the money come from?

There are no more than 10 investors in this round of financing, but the lineup is quite impressive.

The most significant one is the National Integrated Circuit Industry Investment Fund, also known as the "Big Fund". The Big Fund has been focusing on semiconductor and chip investment for 12 years since its establishment. This is the first time it has cross - border invested in a large model enterprise. Among other investors, Tencent invested 1 billion yuan, making it one of the largest external institutional investors in this financing (Liang Wenfeng personally invested 2 billion yuan, being the largest single investor in this financing); CATL followed with an investment of 500 million yuan, and NetEase, JD.com, and IDG Capital each invested about 300 million yuan. In addition, institutions such as Monolith Capital are also on the investor list.

This list itself says a lot.

The entry of the Big Fund means that at the national level, large models are being incorporated into strategic infrastructure as important as semiconductors. Previously, it only focused on chips, but now it also focuses on what runs on the chips. Tencent's entry as an institutional investor is to leave an ecological entrance for itself outside of competition. The appearance of CATL is even more unexpected. Why does a battery company invest in AI? The answer may lie in the energy consumption demand of computing data centers, which is a new direction that CATL is betting on.

The national team, gaming, social media, e - commerce, logistics, industrial manufacturing, and capital - this list covers almost half of China's economic growth engines, indicating the significance of this financing.

The money is in, but the key is still in his hand

Before this round of financing was officially launched, Liang Wenfeng did something first.

On April 27, 2026, DeepSeek completed a business registration change, increasing its registered capital from 10 million yuan to 15 million yuan. The newly added part was all subscribed by Liang Wenfeng personally. According to media reports, this operation increased his direct shareholding ratio from 1% to 34%. Coupled with the indirect equity he held through Magic Square - related companies, Liang Wenfeng's total control over DeepSeek reached about 84%, and his voting rights were close to 100%.

This action took place before the financing was made public, and the timing was very precise.

After the investors enter, they will find that they are buying a rather special asset: The funds need to be injected into a limited partnership managed by Liang Wenfeng, rather than directly into the main body of DeepSeek; all investors do not have voting rights in DeepSeek; the shares they hold have a five - year lock - up period, during which they cannot be transferred on the secondary market or privately to avoid short - term cashing out. Liang Wenfeng's team also requires strict verification of the real identities of the LPs behind all investors to avoid the risk of shares flowing into unknown entities.

In other words: You can share the company's growth, but you can hardly influence its decision - making. The key is still in Liang Wenfeng's hand. And you have to be a "long - termist" with him.

It's worth noting that the Big Fund is the only exception. According to Reuters, the Big Fund's funds are directly injected into the main body of DeepSeek, and it retains voting rights and is not subject to the five - year lock - up period. Although they are all "investors", the treatment is very different. This is no longer an ordinary financial investment, but a direct endorsement at the national strategic level.

So, why did Liang Wenfeng decide to raise funds at this particular time?

The answer is that three things came together at the same time.

The first is talent. Since 2025, at least five core researchers of DeepSeek have left one after another. Among them are Guo Daya, a core researcher of R1 (who joined ByteDance), and Wang Bingxuan, the core author of the first - generation large model (who joined Tencent). For a non - listed company, the options in employees' hands have no market - based pricing and no exit channels, making it increasingly difficult to retain talent.

The second is computing power. The focus of AI competition is shifting from chatbots to AI Agents. The scale of computing power required for training and inference is on another level, and the profits of Magic Square Quantitative are no longer sufficient to support it alone.

The third is the valuation window. DeepSeek's valuation soared from 10 billion US dollars to over 50 billion US dollars in just two months. Such windows don't come often, and if you miss them, you miss them.

With the superposition of these three factors, raising funds has changed from "unnecessary" to "a must - do now".

What will this money be used for?

According to media disclosures, over 60% of this financing will be used for the construction of computing power infrastructure, about 25% for model R & D and iteration, and the remaining about 15% for core talent incentives.

There is a longer logical line behind the issue of computing power.

The V4 preview version released by DeepSeek in April this year is the first open - source model with a trillion parameters that has completed full - stack adaptation to Huawei's Ascend platform. This is not an ordinary technological change. The top - level computing power chips of NVIDIA are still subject to export controls. Chinese AI companies have long been in a passive position in terms of computing power. DeepSeek has proved with V4 that domestic chips can run top - level large models. In this round of financing, over 60% of the funds will continue to be invested in domestic computing power such as Huawei's Ascend, with a very clear direction.

It's not hard to understand why the National Big Fund chose to lead the investment at this time. The Big Fund has always invested in the chip industry chain, from design to manufacturing to packaging and testing. Now it has entered the large model field, and the investment logic is actually the same: For domestic computing power to really work, it needs both good chips and good models running on them. DeepSeek is the missing piece in this closed - loop.

The chain reaction triggered by this financing at the industry level is also worthy of attention. In the past, the valuations of large model companies were often linked to their revenue scale. DeepSeek has almost no business model, but its valuation has increased six - fold in two months. The market's pricing is aimed at another thing: technological sovereignty and ecological control. A company that can break the dependence on NVIDIA's computing power and output open - source standards globally has strategic value that cannot be measured by quarterly revenue. Once this valuation logic is established, the financing coordinate system of the entire domestic AI industry will shift.

With the financing in place, there are also new developments at the product level. According to reports, DeepSeek plans to launch the V4.1 version in the near future, which will fully support the MCP protocol and integrate image and audio input capabilities, officially moving towards multi - modality. With the influx of funds, the iteration pace may also speed up.

Conclusion

Although DeepSeek has raised the largest amount of money in the domestic AI circle so far, it is still the same DeepSeek.

It is still open - source. The V4 preview version was released in April, and the weights were open - sourced simultaneously. This will not change because of the financing. According to reports, Liang Wenfeng's stance at the investor meeting is also very clear: the main task is to promote technological upgrading, rather than short - term monetization.

What has changed is the scale and source of the money. The logic of doing things remains the same.

People outside send money in, but find themselves standing outside the door, unable to see what's going on inside or get involved. It's quite rare for a company to maintain such a high degree of independence in the largest - scale financing. As for how far this logic can go, we'll leave it to time to answer. But the past has given us a sufficiently optimistic answer.

This article is from the WeChat official account "Slow - motion Manfon". Author: Gu Zi, Editor: Liu Han. Republished by 36Kr with permission.