Surpassing the World's Top Big Tech... How Did Chinese Hedge Fund Create DeepSeek? [TechTalk]

Quant Developers: The Hidden Talent Pool of the AI Industry
Experts in Networks, Chips, Mathematics, and Programming
Now Emerging as Standouts in Deep Learning

Chinese company 'Deepseek,' which has introduced a 'cost-effective AI' with performance comparable to cutting-edge generative AI at less than one-tenth the cost of competitors from the US and other countries, has recently caused a huge stir in the industry.

DeepSeek, which released a free chatbot emphasizing low development costs. Photo by AFP Yonhap News

The origin of this company's AI model is extraordinary. In fact, Deepseek's parent company is the Chinese hedge fund operator 'Huanfang Quant (Highflyer),' and the AI model research reportedly started merely as a 'side project.' How did a hedge fund, seemingly unrelated to AI, manage to surpass even the world's top big tech companies?

Outpacing Big Tech at Less Than One-Tenth the Cost

Ryang Won-pung (right), who led the development of DeepSeek. He is a researcher at the Chinese hedge fund 'Huanfang Quant'. CGTV Yonhap News

Deepseek attracted industry attention when it claimed that its inference AI model, Deepseek-R1, was trained last December at a cost of only 8 billion KRW (approximately 6.7 million USD). This model currently demonstrates performance nearly on par with GPT-4o developed by OpenAI. Due to US export restrictions, China cannot import NVIDIA's high-performance GPU H100, so this model was trained using 2,048 H800 GPUs (semiconductors with less than half the interconnect bandwidth of the H100). Compared to big tech companies that utilize massive data centers and AI supercomputers, this is a mere drop in the bucket.

Deepseek adopted pure reinforcement learning instead of the commonly used 'fine-tuning' in deep learning training. Reinforcement learning is a machine learning method where AI is trained by rewarding it when achieving specific goals. In fact, reinforcement learning is more cost-effective than general deep learning in terms of training expenses, but it has significant side effects such as reward hacking (errors where the AI bypasses tasks or ignores rules to gain rewards), which have prevented it from being widely adopted. Therefore, various complementary methods like 'Reinforcement Learning with Human Feedback (RLHF)' have been proposed.

After the U.S. Department of Commerce imposed export restrictions on GPUs to China, Nvidia designed and is selling a lower-performance, budget GPU called the 'H800.' The photo shows a server farm with eight H800 80-gigabyte (GB) GPUs connected, which is currently rented out. Small and medium-sized AI companies in China build such equipment themselves and use it as research infrastructure. Screenshot from the Made in China website. Photo by XXX

Thus, Deepseek did not bring an entirely new technology to outpace all big tech companies, but it boldly delved into an area that other companies had not focused on and created its own breakthrough. Considering the extremely limited resources and personnel they had, this is a remarkable achievement.

Hedge Fund Armed with Mathematics and Programming

Above all, it is significant that Deepseek was merely a side project of its parent company, Huanfang Quant. The 2,048 H800 GPUs used by Deepseek were reportedly just shared from those Huanfang Quant used for algorithmic trading. Huanfang Quant is known as a hedge fund managing assets under management (AUM) of 8 billion USD (approximately 11.5 trillion KRW).

However, hedge funds have long been deeply connected to the AI industry. Many core developers of today's leading AI companies such as DeepMind, OpenAI, and Stability AI have worked at algorithmic trading firms like hedge funds and market makers (liquidity providers).

David Ha (left), former CTO of the UK AI company Stability AI and currently CEO of Sakana AI, Japan's first generative AI company, is a former Goldman Sachs quant trader. Today, many AI researchers are known to come from quant backgrounds such as hedge funds and market makers. Screenshot from NVIDIA website

Algorithmic trading, or quant (Quant), aims to profit by quickly capturing price fluctuations across various asset classes such as stocks, bonds, and commodities. Therefore, those who develop algorithms or models used in quant trading are all proficient in mathematics and programming, overlapping with the expertise required for AI development.

AI Competition May Not Be Exclusive to Big Tech

The hedge fund industry has only recently begun directly researching neural network AI and releasing generative AI. Although quants have automated financial trading using machine intelligence, the machine intelligence in this field was quite different from generative AI.

Algorithmic trading has mainly used computer programs called HFT (high-frequency trading). These programs process millions of trades in less than a second and are known in the industry as 'algobots.' Algobots exploit subtle price fluctuations at trading speeds beyond human cognitive ability to generate profits.

Asia Economy DB

Consequently, program development in hedge funds has traditionally been a pure 'speed competition.' The focus was on purchasing custom high-speed routers and computer chips to build faster servers and physically relocating these devices as close as possible to actual exchanges to maximize transaction throughput. They also researched various mathematical algorithms to improve program response speed.

However, recently some quant firms have started to develop superior trading strategies by leveraging AI's predictive capabilities rather than speed. Specifically, strategies have emerged that use deep learning AI trained on vast amounts of data about asset price fluctuations to predict prices 3 to 10 minutes ahead and trade for profit.

In fact, late last year, a market maker that ousted algobots and posted record profits with such a 'prediction bot' was reported by the Financial Times (FT). The company, called 'XTX Markets,' became the UK's number one and among the top three quant firms worldwide within 10 years of its founding. It was established by Alex Gerko, a Russian-born mathematician, who ranked 12th on the UK rich list last year with a net worth of 15 billion USD (approximately 21 trillion KRW).

As it has been proven that deep learning AI can outperform ultra-high-speed trading programs, algorithmic trading firms such as hedge funds are gradually entering the AI industry. The photo shows the UK-based XTX Market, which quickly rose to become one of the world's top three quant firms by leveraging deep learning AI. XTX Market homepage

This company owns 25,000 state-of-the-art NVIDIA GPUs and employs about 250 specialized AI researchers with exclusive datasets for deep learning AI training. This infrastructure scale rivals that of many big tech companies.

Huanfang Quant, which delivered the 'Deepseek shock' to the tech industry, is also presumed to be a company with trading strategies based on deep learning AI like XTX Markets. It is still unclear whether Deepseek will completely overturn AI development approaches or remain just one of many training strategies. However, this incident suggests that AI innovation can arise not only from big tech but also from other industries. Hedge funds, equipped with sharply honed mathematicians trained in financial engineering, vast datasets, and now GPUs, may be the most suitable new players in this field.