DeepSeek has achieved viral status.
This week, the Chinese AI lab DeepSeek captured mainstream attention following its chatbot app’s ascent to the top of both the Apple App Store and Google Play Store charts. The AI models developed by DeepSeek, which employ compute-efficient techniques, have prompted Wall Street analysts and technologists alike to question whether the United States can maintain its lead in the AI domain and whether there will be sustained demand for AI chips.
The rapid rise of DeepSeek on the international stage raises questions about its origins and the journey to its current prominence.
Origins of DeepSeek
DeepSeek is supported by High-Flyer Capital Management, a Chinese quantitative hedge fund leveraging AI for trading decisions. Liang Wenfeng, an AI enthusiast and co-founder of High-Flyer in 2015, reportedly explored trading while studying at Zhejiang University. In 2019, Wenfeng launched High-Flyer Capital Management as a hedge fund with a focus on AI algorithm development and deployment.
By 2023, High-Flyer initiated DeepSeek as a lab to explore AI tools independently from its financial business. Funded by High-Flyer, the lab transitioned into an autonomous company named DeepSeek.
DeepSeek, from its inception, constructed its own data center clusters for training AI models. However, like many AI enterprises in China, DeepSeek has been impacted by U.S. export restrictions on hardware, compelling the company to utilize Nvidia H800 chips for training one of its latest models, in contrast to the more advanced H100 chips accessible to U.S. entities.
The technical team at DeepSeek is characterized by its youthfulness and the company reportedly recruits AI researchers with doctorates from premier Chinese universities. According to The New York Times, DeepSeek also employs individuals without computer science backgrounds to enhance their technology’s comprehension of diverse subjects.
DeepSeek’s Strong Models
DeepSeek introduced its initial set of models—DeepSeek Coder, DeepSeek LLM, and DeepSeek Chat—in November 2023. However, it was not until the subsequent spring that the release of the next-gen DeepSeek-V2 models attracted substantial attention within the AI industry.
The DeepSeek-V2, known for being a comprehensive text- and image-analyzing system, excelled in various AI benchmarks and was significantly more cost-effective than similar models available at the time. This performance prompted domestic competitors such as ByteDance and Alibaba to reduce the usage prices for some of their models and offer others for free.
Launched in December 2024, DeepSeek-V3 further elevated DeepSeek’s renown. Internal benchmark testing revealed that DeepSeek V3 outperformed many downloadable, openly available models such as Meta’s Llama and "closed" models like OpenAI’s GPT-4o.
Equally notable is DeepSeek’s R1 "reasoning" model, released in January. DeepSeek asserts that R1 matches the performance of OpenAI’s o1 model on key benchmarks. As a reasoning model, R1 performs self-fact-checking, which helps avoid typical model pitfalls. Reasoning models generally require more time, from seconds to minutes, to reach solutions compared to non-reasoning models but tend to be more dependable in fields such as physics, science, and math.
Nevertheless, as Chinese-developed AI, models like R1 and DeepSeek V3 are subject to evaluation by China’s internet regulator to ensure their responses align with core socialist values. For instance, DeepSeek’s chatbot app will not address queries regarding Tiananmen Square or Taiwan’s autonomy.
A Disruptive Approach
DeepSeek’s business model remains somewhat ambiguous, as the company provides its products and services at prices below market value and offers others for free. According to DeepSeek, breakthroughs in efficiency have facilitated the maintenance of extreme cost competitiveness, though some experts challenge the figures provided by the company.
Developers have shown an affinity for DeepSeek’s models, which, while not open source by conventional definitions, are available under permissive licenses permitting commercial use. Clem Delangue, CEO of Hugging Face—a platform hosting DeepSeek’s models—tweeted that developers on Hugging Face have created over 500 "derivative" models of R1, collectively amassing 2.5 million downloads.
DeepSeek’s success, deemed "upending AI" by some and "over-hyped" by others, has had significant impacts, such as a decrease in Nvidia’s stock price by 18% and eliciting reactions from OpenAI CEO Sam Altman.
Microsoft has announced the availability of DeepSeek on its Azure AI Foundry service, a comprehensive AI service platform for enterprises. During Meta’s first-quarter earnings call, CEO Mark Zuckerberg indicated that AI infrastructure investments will remain a "strategic advantage" for Meta, despite inquiries about DeepSeek’s influence on Meta’s AI spending.
Concurrently, several companies, countries, and governments, including South Korea, have banned DeepSeek due to concerns about data risks associated with China. Furthermore, New York state has prohibited the use of DeepSeek on government devices.
The future trajectory of DeepSeek remains uncertain. While enhanced models are anticipated, the U.S. government appears increasingly cautious about perceived harmful foreign influence.
This article was first published on January 28, 2025, and continues to be updated with additional information.
