Liang Wenfeng used profits from his hedge fund, High-Flyer, to launch DeepSeek in 2023 and chase artificial general intelligence.
DeepSeek’s low-cost, open models, led by R1, showed that smart design can rival huge budgets, and the shock hit NVIDIA’s market value hard.
In 2026, DeepSeek took outside money for the first time, with a first round of about USD 7.4 billion and a second round reportedly near USD 15 billion.
Liang Wenfeng built one of the most talked-about AI companies in the world without a big marketing push or a Silicon Valley address. The founder of DeepSeek showed that a smart, lean approach can compete with models that cost far more to build.
DeepSeek has around 173 million downloads since January 2025, according to Fortune, and a funding round this year valued it at about USD 50 billion. This article looks at how Liang Wenfeng, a hedge fund founder turned AI leader, is shaking up the industry with efficient models.
Liang was born in 1985 and studied electronic engineering at Zhejiang University. He later co-founded High-Flyer, a hedge fund that uses math and AI to pick investments. Profits from that business paid for DeepSeek, which Liang launched in 2023 to reach artificial general intelligence.
In a 2023 interview, he said the push into large models was not directly tied to finance. Reports say High-Flyer also bought thousands of NVIDIA chips before export limits tightened, which gave the new lab a head start.
DeepSeek’s main idea is simple: use less computing power for each answer. Its V3 model uses a design called mixture of experts, which splits the model into many small expert parts and uses only a few at a time.
The model has 671 billion parameters in total, but it activates only about 37 billion for each token, or small piece of text. This cuts the cost of running the model.
The company said the final training run of V3 cost about USD 5.6 million in GPU time, though that figure leaves out earlier research and testing. A 2025 paper co-authored by Liang also described sparse attention, a method that helps models handle long text with less memory.
The company also built its models to run well on modest hardware, a smart choice after US export rules limited its access to top chips.
In January 2025, DeepSeek released R1, a reasoning model that matched top rivals on many tests at a much lower cost to run. It was also open source, so anyone could download it and build on it. Scale AI CEO Alexandr Wang praised R1 as a top performer on Humanity’s Last Exam, a test built from hard expert questions.
Investors worried about demand for costly AI chips, and NVIDIA lost nearly USD 600 billion in market value in a single day. A Nature paper later reported that the extra training behind R1 cost about USD 294,000. The episode led many analysts to question whether huge budgets were the only path to strong AI.
Liang runs DeepSeek differently from most tech firms. In a leaked staff talk reported by Fortune, he said, “We generally don’t work overtime.” He added that research needs a relaxed setting and that the team stays focused on only a few tasks. DeepSeek also keeps prices low, and Liang has said it aims to earn only a reasonable profit.
Its pricing helped set off price cuts among Chinese rivals. Liang’s control over the company is strong as well. He owns about 81.7% of DeepSeek after its first outside round, according to Qichacha data cited by Bloomberg.
For years, Liang refused outside money. This changed in 2026. DeepSeek raised about USD 7.4 billion in a first round that closed in June, and Liang put in around USD 3 billion himself. A second round has reportedly raised about USD 15 billion from investors such as CATL and Tencent, Bloomberg reported on October 22.
SCMP reported that Liang took outside capital partly since rivals were pulling away staff. One investor told the paper that Chinese AI models are becoming alike, with no clear leader.
People familiar with the matter say Liang is also weighing a move into robotics, though DeepSeek has not commented. Challenges remain, including tough competition and the need to keep top researchers.
Liang Wenfeng has shown that AI progress does not depend only on money and giant data centers. His team cut costs with smart design, shared its work openly, and forced rivals to rethink their plans. Now the test is harder.
DeepSeek has billions in fresh funding, a large user base and growing competition. If it keeps turning clever research into cheap, strong models, the founder of DeepSeek could keep shaping how the whole industry builds AI.
1. Who is Liang Wenfeng?
Liang Wenfeng is the founder and CEO of DeepSeek. He studied at Zhejiang University and co-founded the hedge fund High-Flyer before starting DeepSeek in 2023.
2. How is DeepSeek funded?
High-Flyer’s profits paid for DeepSeek in its early years. In 2026, it raised about USD 7.4 billion in a first outside round at a valuation near USD 50 billion, and a second round has reportedly raised about USD 15 billion.
3. What is a mixture-of-experts model?
It is a design that splits a model into many small expert parts and uses only a few for each task. This lowers the cost of running the model without cutting its overall size.
4. Why did DeepSeek R1 shake the market?
R1 matched top rivals on many tests at a much lower cost. Investors feared less demand for costly AI chips, and NVIDIA lost nearly USD 600 billion in market value in one day.
5. Is DeepSeek open source?
Many DeepSeek models, including R1, were released as open source. Developers can download them and build on them.