“We are indeed in contact with DeepSeek regarding funding,” a friend from a financial advisory firm told us.
Over the past two weeks, news of DeepSeek's fundraising has finally taken hold. When summarizing this "unusual" move, almost everyone inevitably mentions that Liang Wenfeng is going to give core employees a definite valuation.
After all, the competition for large-scale models became increasingly fierce that year, and the turnover of core talents at DeepSeek was a hot topic, with people such as Luo Fuli, Wang Bingxuan, and Guo Daya jumping ship to Xiaomi, Tencent, and ByteDance.
Beyond the noise, a set of data paints a more peaceful picture—DeepSeek V4 was finally released in late April, and a long list of author acknowledgments in the technical report shows that 10 out of the approximately 270 research engineers left during the development process . This translates to a technical R&D staff turnover rate of less than 4%.
In fact, Leung Man-fung retained the vast majority of people.
DeepSeek's employee exodus has been triggered, with 10 employees leaving.
Starting in 2023, a strong feeling of being pushed back came.
Following ChatGPT's success, star companies such as Luna's Dark Side, Leap Star, and MiniMax were established around the same time, while large-scale model products from major manufacturers like Doubao, Tongyi Qianwen, and Wenxin Yiyan emerged in large numbers.
At just the right time, Liang Wenfeng also launched DeepSeek in Beijing and Hangzhou that year.
In a rare public sharing session, he discussed his talent philosophy: Most of DeepSeek's developers are recent graduates or have not been working in AI for long. If you are pursuing short-term goals, it is certainly right to hire experienced people, but in the long run, basic skills, creativity and passion are more important.
That's true. In early 2025, DeepSeek R1 demonstrated its strength, and people began to really pay attention to this team of about 150 people, many of whom were young people who had just graduated or were still in school from top universities in China, with a very high proportion of Tsinghua and Peking University graduates.
Almost inevitably, the topic of talent mobility began to plague DeepSeek in the following year.
Starting in 2025, DeepSeek saw reports of departures from its core team, including Luo Fuli, Wang Bingxuan, Wei Haoran, and Ruan Chong. Many of these individuals moved to other companies to become heads of core businesses. A popular, easy-to-understand joke about this was: "When DeepSeek employees see that people of similar skill levels can earn so much more by leaving, why can't I?"
Until early 2026, when Guo Daya jumped ship to ByteDance's Seed team, discussions about DeepSeek's talent drain reached a high point. At that time, coupled with the long delay in the release of DeepSeek V4, it was inevitable to raise some concerns about a talent gap.
But reality isn't so disheartening. DeepSeek V4 has finally been released, and the accompanying technical report includes a list of author acknowledgments. A closer look reveals that its Research & Engineering team comprises approximately 270 people, considered the core R&D team of any AI company, along with 48 members in Business & Compliance.
During the development of DeepSeek V4, only 10 members of the research engineering team left.
In other words, out of a 270-person R&D team, 10 chose to leave, resulting in a turnover rate of less than 4% in the core department—which is low enough. Data shows that OpenAI lost over 25% of its key research talent in the past two years, most of whom jumped ship to competitors like Meta or started their own businesses.
Opening the door to financing for the first time, stabilizing morale.
The venture capital community is currently particularly eager to see: who will participate in DeepSeek's first funding round?
Starting in April, DeepSeek was first reported to be launching its first round of external financing at a valuation exceeding $10 billion. Just a week later, reports surfaced that DeepSeek was in talks with Tencent and Alibaba regarding investment. Later, industry insiders circulated that DeepSeek's pre-investment valuation was 300 billion RMB.
As of now, DeepSeek has not responded to the funding news.
An investor advisor (FA) told us that they have recently been in talks with investment institutions regarding potential collaborations for DeepSeek's funding round, but very few financial investors are involved in this round. This has also been confirmed: Tencent and DeepSeek have communication regarding routine business operations, but there has been no substantive discussion about funding.
Everything remains shrouded in mystery.
On April 27, DeepSeek's registered capital increased from 10 million yuan to 15 million yuan. Liang Wenfeng's subscribed capital increased from 100,000 yuan to 5.1 million yuan, and his direct shareholding ratio rose from 1% to 34%. Simultaneously, the shareholding ratio of Ningbo Cheng'en Enterprise Management Consulting Partnership, controlled by Liang Wenfeng, decreased from 99% to 66%. After these changes, Liang Wenfeng holds approximately 84.29% of DeepSeek's equity through both direct and indirect means.
It is worth noting that previously, Liang Wenfeng held the vast majority of DeepSeek's shares through Ningbo Cheng'en, with very few direct shares. However, after this change, Liang Wenfeng's direct shareholding has increased to 34%. As a result, Liang Wenfeng's controlling stake is now more readily apparent— if due diligence for financing is conducted, DeepSeek's shareholding structure will become clearer.
"It's not something most people can participate in," investors remarked sincerely. Indeed, after a fierce battle in China's large-scale model market, DeepSeek remains very attractive.
Just as the DeepSeek V4 preview has finally been released, the Pro and Flash versions come standard with millions of contexts. The Pro version boasts up to 1.6 trillion parameters, and the price is impressive: Pro costs 1 yuan (cache hit) or 12 yuan (cache miss) per million token inputs, and 24 yuan for outputs, while Flash costs 0.2 yuan, 1 yuan, and 2 yuan respectively.
Meanwhile, the rumored compatibility with domestic chips has been confirmed. In the DeepSeek V4 technical report, although it can be seen that the model training part still most likely uses NVIDIA chips, Huawei Ascend and NVIDIA are listed side by side in the verification platform. "It is expected that after the Ascend 950 supernode is launched and deployed in batches in the second half of the year, the price of the Pro version will also be significantly reduced."
This move signifies that DeepSeek has created a crack in NVIDIA's robust CUDA ecosystem. The implications are self-evident.
Liang Wenfeng's conviction signifies the true beginning of the era of domestically developed AI.
Liang Wenfeng and DeepSeek are on a path that goes against consensus.
Typically, the timeline for a star tech company goes like this: seizing funding opportunities as it begins to show its potential, accelerating talent expansion and product iteration, quickly capturing the market, and seeking an IPO. Once this sequence of actions is established, it's difficult to stop.
However, each step of DeepSeek was unexpectedly slower.
When DeepSeek R1 was released in early 2025, Liang Wenfeng had virtually no rivals, but at the height of his success, he rejected all investors who came knocking on his door. Ironically, in today's fiercely competitive environment, with numerous competitors, DeepSeek announced its first round of funding. Most external discussions attributed this to two main points: research and development requires funding , and on a deeper level, DeepSeek needs to provide its internal talent with a definite valuation.
Product iterations have also been slow to arrive. It's been 15 months since the last major update to DeepSeek V4, and despite much anticipation, DeepSeek only released a preview version initially, neglecting the previously considered missing multimodal capabilities. It wasn't until April 29th that DeepSeek launched a gray-scale image recognition mode, signaling the introduction of multimodal capabilities.
"Unmoved by praise, unfazed by slander," this is DeepSeek's stance. And the market's feedback seems to confirm that patience can lead to success.
On the day DeepSeek V4 was released, domestic AI chips from companies such as Huawei Ascend, Cambricon, Hygon, Moore Threads, Muxi, Kunlun Chip, Pingtouge Zhenwu, and Tianshu Zhixin completed their adaptation. This immediately sparked a surge in the secondary market price of domestic chip products.
Meanwhile, reports indicate that market demand for Huawei's Ascend 950 series AI chips has surged, with ByteDance, Tencent, and Alibaba, three of China's leading internet companies, already in talks with Huawei regarding new chip orders.
Thus, a contrarian triggered an industry resonance with its slow pace—when underlying chip manufacturers and leading companies began to evolve and converge around DeepSeek's benchmark, DeepSeek may have already leaped out of the original competitive arena.
As the saying goes: slow is fast. However, not many people truly believe it.
This article is from the WeChat public account "Investment Community" (ID: pedaily2012), author: Feng Yuchen
