Chinese ChatGPT Rival DeepSeek Is Making Waves In Silicon Valley Despite U.S. Curbs

DeepSeek, the ChatGPT rival developed by Chinese hedge fund mogul Liang Wenfeng, has topped download charts and started a debate over how a relatively obscure company managed to make advanced AI products—despite Washington’s export controls on related technologies.

Powered by a proprietary AI model, DeepSeek-AI Assistant is now the most popular free app in Apple’s App Store in both China and the U.S., according to Sensor Tower, a San Francisco-based market intelligence platform. Similar to ChatGPT, the chatbot can perform tasks such as answering questions, drafting content and helping users gather information.

The app’s surge in popularity comes as DeepSeek, the Hangzhou-based company behind it, has transfixed Silicon Valley with its technological breakthroughs. It appears to have developed advanced AI at a fraction of the cost spent by the San Francisco-based OpenAI while using less powerful chips, since U.S. officials imposed restrictions on the sale of semiconductors to China.

Released earlier this month, the DeepSeek R1 model is now the world’s fourth most popular AI model by user vote, according to Chatbot Arena, an evaluation platform developed by tech researchers including from the University of California, Berkeley. The product ranks ahead of the OpenAI o1 model released last December, according to Chatbot Arena.

It is capable of complex reasoning and solving sophisticated mathematical problems, with a performance matching OpenAI o1, DeepSeek claimed when the product was launched earlier this month. Liang met Chinese Premier Li Qiang last week as part of a government meeting to discuss AI.

The company didn’t respond to requests for further comment. Founded in 2023 and with no known outside funding, DeepSeek’s early success is impressing Silicon Valley deeply. “DeepSeek R1 is one of the most amazing and impressive breakthroughs I’ve ever seen—and as open source, a profound gift to the world,” billionaire investor Marc Andreessen wrote in a Jan. 24 post on the X social media platform.

Paul Triolo, a Washington, D.C.-based partner at advisory firm DGA-Albright Stonebridge Group, says by email that DeepSeek’s models are considered among the best, especially when benchmarking its performance against OpenAI.

The company’s rapid rise is driven by Liang Wenfeng, a reclusive hedge fund mogul fixated on advanced technologies. Liang was born in 1985 in Guangdong, where his father was an elementary school teacher. Liang started his career as an investor, according to local media reports. After receiving a master’s degree related to computer vision from the Hangzhou-based Zhejiang University in 2008, Liang began researching how to use AI in stock trading.

Liang cofounded hedge fund High-Flyer Quant in 2015 and a year later was deploying his algorithms to assess investment opportunities. By 2017, almost all of High-Flyer Quant’s investments were made by AI, the company says on its website.

By 2019, High-Flyer Quant's assets under management (AUM) had soared to 10 billion yuan ($1.4 billion). Some of its top-performing funds have returns of over 200% as of 2024, according to data from China’s Hithink Flush Information cited by local media.

Yet Liang’s true interest appears to be exploring frontier technologies. In a rare 2023 interview, the entrepreneur told local media 36Kr that he wants to research and innovate. While running High-Flyer Quant, Liang started to stockpile Nvidia chips before Washington’s export controls kicked in. He also invested returns from his hedge fund into the training of AI models.

In 2023, DeepSeek was founded. According to Qichacha, a business information service, the startup remains self-financed with no outside investor. The company launched its first AI model in November 2023.

Xu Huazhe, an assistant professor of information sciences at China’s prestigious Tsinghua University, says via WeChat that some of DeepSeek’s innovations appear to have come from a technology called mixture of experts (MoE). It allows researchers to train many smaller AI models at the same time, and combine selected outputs to respond to users.

Such a technique may not require that many advanced chips or gathering as much data, meaning DeepSeek can potentially save on costs, according to Xu. The company has claimed that some of its AI models only took $5.6 million to train, which is reportedly up to 95% less than the amounts spent by OpenAI.

DeepSeek, in turn, charges users less. The AI assistant chatbot is free to use, but application developers pay to access the underlying model, which they can tweak to build other products based on it. DeepSeek R1 charges 14 cents per million tokens, which in AI refers to the amount of data processed, while OpenAI charges $7.50.

Shen Meng, a Beijing-based managing director at boutique investment bank Chanson& Co., says via WeChat that DeepSeek has carved out its own technological path while most Chinese models are copycats. He says this has raised hope for China’s chip developers as well.

“The training of DeepSeek doesn’t rely on as much computing power, which means it doesn’t need that many Nvidia’s graphic processing units (GPUs, a type of AI chip),” says Shen. “This means that chips from China now have more room for growth.”