The open-source DeepSeek-R1 and DeepSeek-V3 language models, created in China, have disrupted the US-dominated industry at a time when the US government is planning hundreds of billions of dollars in AI investment. The new language model, which performs at the same level as OpenAI’s ChatGPT-4o and o1 models but is 90% cheaper, caused a shock in U.S. technology stocks, with an immediate 18% fall in Nvidia stock on January 27.
DeepSeek-V3 and DeepSeek-R1 are their most advanced models yet; the main difference between the two is that DeepSeek-R1 is a “reasoning model”, which focuses on producing a chain of complex reasoning visible to the user. OpenAI’s flagship reasoning model is called o1.
DeepSeek’s models, unlike those of competitors like OpenAI and Anthropic, are open-source; its source code is publicly available and can be used with a non-restrictive license. The research group has emphasized the open approach to language models as more conducive to scientific and technological innovation. (While Meta’s Llama models are also open-source, they do not perform as well as DeepSeek’s models.) Indeed, researchers all over the world are starting to take advantage of DeepSeek’s technology.
Due to concerns of data leaks (such as a vulnerability recently found in DeepSeek’s website, though not the models themselves) and Chinese government access to user data, South Korea, Australia, Taiwan, and the state of New York have already blocked access to DeepSeek for government employees. A House bill has been introduced in the US proposing to ban the service from government devices here as well. OpenAI has accused DeepSeek of using its training data inappropriately. DeepSeek does admit to using OpenAI’s training data.
DeepSeek’s creator, Liang Wenfeng, is also the founder of a quantitative hedge fund called High-Flyer, which uses AI to make trading decisions. In 2023, however, the hedge fund chose to “concentrate its resources and strength, wholly devote itself to serve AI technology that benefits all of humanity, create a new independent research group, and explore the essence of AGI [artificial general intelligence].” An AGI is a hypothetical system that surpasses humans in most relevant tasks. High-Flyer’s research and development budget, along with donations, fund the DeepSeek research group.
DeepSeek has reported that the V3 model cost about $5.576 million to train, which would make it far cheaper than any other model. However, this number excludes the investment into infrastructure and research required before the final training run. Some AI experts estimate that the actual cost, including the cost of 10,000 Nvidia A100 chips bought in 2021, may be over $1 billion.
Liang keeps a low profile, but in the interviews he has given to Chinese sources, he emphasizes his personal interest in developing artificial general intelligence and Chinese domestic technological expertise. He believes DeepSeek’s success in AI, which eclipses even that of other Chinese technology firms, comes from recruiting young, talented, and idealistic individuals from China itself, rather than from overseas.
Sources:
- Al Jazeera Staff. “Which countries have banned DeepSeek and why?” Al Jazeera, 06 02 2025, https://www.aljazeera.com/news/2025/2/6/which-countries-have-banned-deepseek-and-why. Accessed 10 02 2025.
- Baptista, Eduardo. “High-Flyer, the AI quant fund behind China’s DeepSeek.” Reuters, 29 01 2025, https://www.reuters.com/technology/artificial-intelligence/high-flyer-ai-quant-fund-behind-chinas-deepseek-2025-01-29/. Accessed 10 02 2025.
- Mok, Charles. “Taking Stock of the DeepSeek Shock.” Stanford Cyber Policy Center, 05 02 2025, https://cyber.fsi.stanford.edu/publication/taking-stock-deepseek-shock. Accessed 10 02 2025.
- Rich, Brandon. “DeepSeek Explained: What Is It and Is It Safe To Use?” AI@ND, 31 01 2025, https://ai.nd.edu/news/deepseek-explained-what-is-it-and-is-it-safe-to-use/. Accessed 10 02 2025.
- Sina Corporation. “DeepSeek创始人专访:中国的AI不可能永远跟随,需要有人站到技术的前沿.” Sina Finance, 26 01 2025, https://finance.sina.com.cn/tech/2025-01-26/doc-inehhksk9178057.shtml. Accessed 10 02 2025.