Chinese artificial intelligence (AI) startup DeepSeek is shaking up the global AI market by implementing a model with performance comparable to ChatGPT-4 at an ultra-low cost, even amidst U.S. export controls and technology blockades.
According to Bloomberg Businessweek on the 14th (local time), the founder of DeepSeek, Liang Wenfeng, known as the 'Tech Madman,' is directly at the helm of AI research and development. Under his leadership, DeepSeek has grown rapidly, cracking the Silicon Valley-centric order and posing fundamental questions regarding the overall AI industry strategy in the United States.
DeepSeek surpassed major models in the U.S. and Europe in standard tests with its large language model (LLM) 'R1' launched in January this year. According to DeepSeek, the model was completed for just 5% of the expense compared to ChatGPT-4, and immediately after this announcement, global cloud corporations rushed to collaborate with DeepSeek. Companies such as Amazon AWS, Microsoft Azure, Meta, and Mistral are reportedly considering cloud deployment of the DeepSeek model.
Founder Liang Wenfeng, a Zhejiang University alumnus, made billions of dollars in the financial quant field before investing in AI research and growing the company. After establishing DeepSeek in 2023, he absorbed numerous high-level talents with a closed but efficiency-centered development culture. Some corporations and political circles in the United States have raised concerns that DeepSeek may have pilfered OpenAI training data or smuggled restricted Nvidia chips, deeming it a national security threat, but the Chinese government and DeepSeek have categorically denied these allegations.
DeepSeek's strategy is the complete opposite of the American closed model. The R1 model has been fully released as open source, with a structure that allows adjustments to ethical standards or sensitive issues with the push of a button. Tailored to the relatively limited GPU environment in China, DeepSeek early adopted 'Sparsity' technology to implement a learning and inference system utilizing a sparse expert model. While this technology is partly applied by Google, Mistral, and others, DeepSeek has applied it on a large scale to secure both performance and expense.
DeepSeek's rise has sparked criticism that the U.S. government's technology blockade strategy is inadvertently prompting China's self-strengthening. In fact, Jensen Huang, CEO of Nvidia, noted in a recent interview that 'export controls have stimulated China's innovation will, and missing one-third of the global market could be catastrophic for the U.S. tech industry.'
In China, a group of AI startups centered around DeepSeek, known as the 'Six Little Dragons,' is fiercely competing for technological leadership. Giants like Alibaba, Huawei, and ByteDance are pouring billions of dollars into developing large-scale models. The Chinese government has classified the AI and semiconductor industries as national strategies and is providing comprehensive support.
AI corporations in the U.S. remain skeptical about DeepSeek's true intentions. However, an increasing number of U.S. startups are adopting or modifying and using DeepSeek's models. The AI search engine perplexity is training a model based on R1 separately, and internally, it has named it '1776,' symbolizing the number of U.S. independence.
In China, DeepSeek has emerged as a symbol of technological pride. Students take pride solely in being from Zhejiang University, and it is not uncommon for tourists and social media influencers to visit DeepSeek's headquarters. Liang Wenfeng continues to avoid media interviews and has replaced some meetings with hologram videos, maintaining a reclusive stance.
In the AI industry, evaluations are emerging that 'China is no longer a follower but has become a threat' through the rise of DeepSeek. Arnaud Barthelemy, a partner at AI investor Alpha Intelligence Capital, stated, 'DeepSeek was an unannounced competitor,' adding, 'However, the real lesson is that Chinese technologists have devised ways to achieve the best results with fewer resources.'