The quantitative history of Liang Wenfeng, founder of DeepSeek | DeepNet

According to Bloomberg, seven startup founders and artificial intelligence experts estimate that DeepSeek's valuation is between about $1 billion and $155 billion. According to the Bloomberg Billionaires Index, based on the midpoint of the range, DeepSeek's valuation is between $2 billion and $30 billion, so Liang Wenfeng's 84% ​​stake is worth between $1.68 billion and $25.2 billion, which will make him one of Asia's richest technology tycoons.

Chanakya Ramdev, founder of Sweat Free Telecom, a telecommunications company in Waterloo, Ontario, values ​​DeepSeek at about half of OpenAI's valuation ($300 billion), or around $150 billion. This means that Liang Wenfeng's holdings are worth $126 billion, which may make him richer than Nvidia CEO Jensen Huang.

In the field of technological innovation, there is no shortage of capital myths that shine instantly. Behind the legendary stories of wealth, people inside and outside the industry are also thinking: How can an institution that is deeply engaged in quantitative investment surpass technology giants and the "AI Six Little Tigers" in the arena of AI big models for a short period of time and breed an inference engine with performance comparable to OpenAI?

Superficially, quantitative investment naturally combines the dual attributes of capital allocation and model development. Its high-frequency trading system needs to process massive market data flows and build a full-link infrastructure from data collection, cleaning to storage - this is the same as the data preprocessing process for large model training.

"Since its establishment, Huanfang Quant's investment strategy has undergone great changes, but in essence it has always been centered around the AI ​​structure. The core is the use of deep neural networks to train models, so we define ourselves as a hedge fund that relies entirely on artificial intelligence for investment." Huanfang Quant CEO Lu Zhengzhe once summarized the core characteristics of Huanfang Quant.

In essence, relying on artificial intelligence for investment is the concrete practice of Liang Wenfeng's firm belief that "artificial intelligence will definitely change the world. "

Looking back, Liang Wenfeng's career in the field of quantitative investment is like a hidden river, calm on the surface, but deep down there is a question about the essence of technology. Before this company stirred up the industry with AI big model technology, Liang Wenfeng's quantitative investment career had already laid the groundwork for his technology philosophy.

Liang Wenfeng, a believer in quantitative investment

"Like many new technologies, quantitative investment was also ridiculed when it first appeared. No one believed that computers could invest like humans. But Simons keenly predicted that with the development of computer technology, one day the 'impossible' would become a reality." This is what Liang Wenfeng said in the preface to James Simons' only biography, "The Man Who Conquered the Market" (Chinese version).

Before 2023, Liang Wenfeng was extremely low-key and rarely made public statements. One of his few public statements was writing a recommendation for James Simmons' biography . Another public speech was a keynote speech Liang Wenfeng gave at the 10th China Private Equity Golden Bull Award Ceremony on the "Future of China's Quantitative Investment".

James Simons is known as the "Father of Quantitative Investment". Renaissance Technologies, founded by him, established its first fund product, the Medallion Fund, in March 1988. Data shows that in the 30 years from 1988 to 2018, after deducting 5% management fees and 44% performance commissions, the fund achieved an annualized compound rate of return of 39.1%, earning the company more than $100 billion in profits.

James Simons is Liang Wenfeng's guide into the field of quantitative investment. Liang Wenfeng said, "Whenever I encounter difficulties at work, I will think of Simons' words: "There must be a way to model prices."

Public information shows that when Liang Wenfeng was a graduate student at Zhejiang University in 2008, he teamed up with his classmates to use machine learning technology for fully automatic quantitative trading and earned his first pot of gold. According to the official website of Huanfang Quantitative, "Since 2008, we have begun to accumulate market data, other relevant data in the financial market, macroeconomic data, etc., with a cumulative data volume of more than 10PB."

On April 16, 2015, the Shanghai Stock Exchange 50 and the China Securities Index 500 stock index futures were officially listed on the China Financial Futures Exchange, which created more room for quantitative funds to operate. Two months later, Liang Wenfeng and his college classmate Xu Jin founded Hangzhou Huanfang Technology Co., Ltd., which was later renamed Zhejiang Jiuzhang Asset Management Co., Ltd.

It should be noted that Huanfang Quantitative mainly consists of Zhejiang Jiuzhang Asset Management Co., Ltd. (Jiuzhang Asset) and Ningbo Huanfang Quantitative Investment Management Partnership (Limited Partnership) (Ningbo Huanfang Quantitative). Both companies are registered with the Fund Industry Association, and the actual controller is Liang Wenfeng. Among them, Liang Wenfeng holds 85% of the shares of Jiuzhang Asset and 76.2684% of the shares of Ningbo Huanfang Quantitative.

Seizing the dividend period of domestic quantitative trading, Huanfang Quantitative quickly entered the fast lane of quantitative investment. According to data from Private Equity Ranking Network and Huanfang Quantitative's official website, in 2016, Huanfang Quantitative managed funds of approximately 1 billion yuan; in 2019, Huanfang Quantitative ranked among the top 10 billion private equity funds; and in mid-2021, it exceeded the 100 billion yuan mark.

2021 is also the dividing line for Huanfang Quantitative to shift from rapid expansion to stable development. Since the end of 2021, there has been an endless stream of information about Huanfang Quantitative's sharp withdrawal, public apologies, and redemption advice.

According to data from Private Equity Ranking Network, Huanfang Quantitative's performance took a sharp turn for the worse since September 2021. By April 2022, the company's maximum dynamic drawdown was as high as 23.48%.

Regarding the changes at the end of 2021, Huanfang Quantitative staff told the "Daily Economic News" that "the excess returns in 2021 have indeed returned to zero, mainly because the strategy was not well implemented. The large scale is partly the reason, and the improved market efficiency, frequent style switching, and insufficient strategy models are all reasons."

According to Cailianshe, by early March 2022, Huanfang Quantitative had proactively reduced its scale to around 50 billion yuan. In terms of yield, according to data from Private Equity Ranking Network, the average yields of Huanfang Quantitative were 0.38% and 4.46% in 2022 and 2023, respectively. Compared with the annualized yield of over 30% of the "Jiuzhang" series products from 2016 to 2020, it has declined significantly.

According to data from 10Jun, as of December 20, 2024, among the 65 funds under Huanfang Quantitative that have publicly disclosed their net value, 29 funds have achieved an increase of more than 10% this year, and 36 funds have been in a declining state this year.

The decline in Huanfang Quantitative's annualized rate of return is not only affected by changes in the stock market and the interweaving of various domestic and foreign market factors, but also by another important reason, namely, "too many monks and too little porridge."

Data shows that from 2017 to 2020, the proportion of quantitative funds in all securities private equity funds increased from less than 5% to 15%, the scale of management increased by 4 times, and the proportion of quantitative trading volume in the entire A-share market trading volume exceeded 15%.

When the management scale of quantitative private equity reaches a certain value, excessively large funds will cause violent market fluctuations, volatility will increase accordingly, and performance will be difficult to maintain at a high level. The difficulty of making profits for quantitative private equity funds that rely solely on technical aspects will increase significantly.

Cracking the “ Dark Forest ” Law

In the field of quantitative investment, different quantitative investment institutions usually have their own unique strategies (such as high-frequency trading, statistical arbitrage, machine learning driven, etc.) and factors. The uniqueness of the strategy is one of the important sources of its competitiveness.

In order to ensure that their strategies continue to be effective in different market environments, quantitative institutions not only need to strictly keep the strategy logic confidential, but also need to build a strong technical infrastructure, dynamic optimization capabilities, unique data advantages and a complete risk control system.

Such a quantitative investment market is like the "Dark Forest" in the Three-Body World. Every civilization must hide itself. Once exposed, it will be destroyed by a more powerful civilization. The same is true for the quantitative market.

The core of quantitative investment is "information gap" and "speed gap". Once exposed, the strategy is imitated or reversely cracked by competitors, and its excess returns may disappear quickly. Therefore, the factors of quantitative investment institutions must be invisible like "Smart Son".

When quantitative investment enters "extraordinary competition", quantitative investment institutions need to ensure the continuous evolution of strategies, and they need to ensure the comprehensive iteration of technical engineering, cognitive frameworks, and organizational capabilities. At this time, machine deep learning and general artificial intelligence become one of the paths for quantitative investment institutions to break the "dark forest" law.

Magic Square Quant is one of the earliest quantitative institutions to explore the use of machine learning to replace traditional quantitative strategies.

According to the information on Huanfang’s official website, on October 21, 2016, Huanfang’s first stock position generated by a deep learning algorithm model went online for real trading, using GPU for calculations. Prior to this, the algorithm mainly relied on linear models and traditional machine learning algorithms, and model calculations mainly relied on CPUs.

An interesting anecdote that has not been officially confirmed by Huanfang Quant is that in order to emphasize the importance of data and AI models and avoid interference from traders' subjective experience, there is a saying circulating within Huanfang Quant: "Engineers' beliefs must be pure - the strategy code is the only source of truth. If a signal cannot pass the Monte Carlo test, it will be thrown into the recycling bin even if it is personally recommended by the chairman."

As of the end of 2017, almost all of Huanfang’s quantitative strategies have been calculated using AI models. In 2018, Liang Wenfeng established Huanfang Quantitative as the main development direction of AI. In 2019, Huanfang AI Lab was established, and recruitment of chip front-end design engineers/architects, deep learning scientists/engineers, and hardware engineers began, with annual salaries ranging from 300,000 to 2 million (according to Huanfang Quantitative’s official account).

To develop the Magic Cube AI Lab, computing power is the foundation. To meet the increasing demand for computing power, Liang Wenfeng began to seek large-scale computing power solutions.

In May 2020, Huanfang Quant launched the deep learning training platform "Firefly No. 1". The computing cluster is equipped with 1,100 high-end graphics cards, which can perform 184 quadrillion floating-point operations per second (18.4PFLOPS, 32-bit precision), equivalent to the computing power of 40,000 personal computers, with an average utilization rate of over 90%.

In January 2021, Magic Square AI Lab's second-generation supercomputer "Firefly 2" was officially delivered and put into use. The delivered AI computing power is 325PFLOPS (TF32), which is 18 times that of "Firefly 1". According to some information, Magic Square Quant's investment in "Firefly 2" has increased to 1 billion yuan, equipped with about 10,000 NVIDIA A100 graphics cards.

Just when many AI researchers have not yet realized that "Wanka" has become an insurmountable software and hardware barrier to general artificial intelligence, Huanfang Quantitative has already set no restrictions on employees' use of computing power.

Regarding the question of why Magic Square Quant was able to reserve high-end GPUs before NVIDIA graphics cards became "hard currency", Liang Wenfeng explained to "Undercurrent Waves": "For many outsiders, the ChatGPT wave has a particularly big impact; but for insiders, the impact brought by AlexNet in 2012 has ushered in a new era. AlexNet's error rate was much lower than other models at the time, and it revived neural network research that had been dormant for decades. Although the specific technical direction has been changing, the combination of model, data and computing power remains unchanged, especially when OpenAI released GPT3 in 2020. The direction was very clear and required a lot of computing power; but even in 2021, when we invested in the construction of Firefly 2, most people still couldn't understand it."

DeepSeek Blaster

Reserve computing power is only the first step. Against the backdrop of declining alpha (excess returns) in the quantitative investment market, Magic Square Quant has quietly embarked on a new journey.

On April 14, 2023, Magic Quant announced on its official WeChat account that the company would set up a new research organization to start a new journey to explore the essence of AGI. In order to recruit the required talents, the company's poster also used the advice written by French director and screenwriter Truffaut to young directors: "Be sure to embrace ambition madly and be madly sincere at the same time."

According to Tianyancha, the main body of DeepSeek, Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd. (hereinafter referred to as DeepSeek), was established on July 17, 2023. The actual controller is Liang Wenfeng, who holds 84.2945% of the shares. The other three partners are Liang Wenfeng’s classmates at Zhejiang University, Chen Zhe, Li Huan, and Zheng Dawei.

The picture shows the beneficial shareholders of Hangzhou Deep Quest Artificial Intelligence Basic Technology Research Co., Ltd. Data source: Tianyancha

Since 2025, the impact of DeepSeek's popularity on AI startups, Internet giants and the entire VC industry is self-evident.

Overseas, OpenAI CEO Sam Altman commented on the social platform X that the DeepSeek-R1 model is "impressive." In China, investor Zhu Xiaohu, who once publicly stated that he was "not optimistic about startups building large models," publicly stated that "DeepSeek is almost making me believe in AGI."

In fact, before DeepSeek-R1 quickly broke the circle, its model had already gone through several versions.

In May 2024, DeepSeek released the open source model DeepSeek-V2. As its inference cost was reduced to only 1 yuan per million tokens, DeepSeek-V2 was nicknamed "Pinduoduo in the AI ​​world". In January 2025, DeepSeek-R1 was released and achieved 100 million user growth in 7 days, becoming a global phenomenon.

Back to the question at the beginning, why was it DeepSeek, a company whose CEO has no AI entrepreneurial aura, has never raised funds, and was incubated by a quantitative investment institution, that triggered OpenAI?

Perhaps, similar to his belief in the ability of mathematical models and algorithms to capture certain patterns in seemingly chaotic data in the field of quantitative investment, Liang Wenfeng's belief in AGI has created infinite possibilities for the continuously iterating DeepSeek.

"What we want to do is general artificial intelligence, or AGI. The large language model may be the only way to AGI, and it has the initial characteristics of AGI, so we will start here, and vision and other features will also be included later," said Liang Wenfeng.