DeepSeek is definitely one of the hottest topic at the moment. The Chinese company, with less than 6 million US dollars in training cost, trained an AI model which is almost as good as OpenAI’s ChatGPT.
The development of DeepSeek is not to my surprise at all. DeepSeek is a subsidiary of High-Flyer, one of the largest quantitative trading firm in China. As the founder and CEO of another quant firm myself, I fully understand why the breakthrough happened in a quant firm.
In Alpha Star Research, our trading decisions are solely determined by our own AI model. With 20% improvement in our AI model, we can see 20% increase in revenue. That’s actual US dollars on our balance sheet. The feedback loop is short and direct: the better our AI models are, the more money we can make. The training cost is paid by our own company treasury. Therefore any saving in training cost would also be actual saving on our company’s balance sheet. Like DeepSeek’s parent High-Flyer, Alpha Star Research is by itself very profitable and the reason we invest in AI, we actually care about the cost and the result of the AI model. We care about the technology.
In OpenAI, it’s a different story. OpenAI itself is not profitable. OpenAI’s valuation (and Sam Altman’s personal wealth) is determined by its investors. The improvement of the performance of ChatGPT is related to that valuation, but the feedback loop is too long for Sam to care. When we look at the training cost, OpenAI could care less about saving money in training. The more OpenAI spend in GPU chips, the more Sam can convince investors about the cost required to train an AI model, and then boost the valuation of OpenAI even more. This creates an interesting situation where the more OpenAI waste in training cost, the more OpenAI’s valuation is.
So if we look deeply, we can discover the interesting contrast: DeepSeek is a finance company, but its profit comes from technology, and it cares about technology and technology only. OpenAI is a tech company, but its money come from valuation (finance) and it cares most about finance. If we look at the background of the founders, DeepSeek’s founder and CEO Wenfeng Liang is a cofounder and software engineer from its parent High-Flyer, and OpenAI’s founder and CEO Sam Altman is a venture capitalist from Y-Combinator.
There are other reasons why quant firms excel most in AI research. In financial markets, we were forced to handle with very limited and polluted data. In our company, we need to train an AI model to predict the market, with only gigabytes of data, polluted by malicious market manipulators. We are forced to spend all our time and energy improving the structures of our AI models instead of brainlessly buying Nvidia H100 and pumping petabytes of data into the system.
Another reason is that when building trading infrastructure, in order to trade before other competitors, we are forced to make our systems as fast as light, literally. We are building microwave transmission system between Tokyo, Hong Kong, Chicago and New York City. And we are making sure the computation time of all components in our trading system less than 1 millisecond, if not microsecond. Quantitative trading firms have technology power and culture to create the fastest and most efficient systems. It was reported that DeepSeek bypassed CUDA and use low level languages to save cost and speed up the training. That’s exactly what I expect a quantitative trading firm would do on AI.
Before starting Alpha Star Research, I worked as a research engineer in Citadel Securities, a leading quantitative trading firm in America. Before that, I work as a software engineer in Amazon and Meta. When I worked at tech companies, when I talk to everyday American about my job, I received a huge amount of praise and admiration, and often heard people thank me for helping the advancement of technology. However when I worked for Citadel, every time I talked about my employer, I got awkward silence and occasionally criticism for being an “evil hedge funder”. I believe this is short-sighted. Just like computers and the Internet were developed by US military for military purposes, but benefit people around the globe eventually, I would expect quantitative firms like DeepSeek/High-Flyer contribute to the development of AI technology as much as Software and Internet companies. And Alpha Star Research and I will also work relentlessly for the advancement of AI technology.
About the Author: Sinya Lee is the founder and CEO of Alpha Star Research, an AI research and quantitative trading firm based in New York City. Like DeepSeek’s founder and CEO Wenfeng Liang, Sinya is also from Guangdong Province originally.
