[Insight & Opinion] Three Perspectives on China's AI 'DeepSeek'

Could Spark Debate Over Excessive Investment in the U.S.
Demonstrates the Potential Success of Open-Source Models
A Case That Highlights the Power of Chinese Developers

President Donald Trump is putting his life on the line to win the 'hegemony war' related to artificial intelligence (AI) with China after taking office. Through an executive order, he declared that the United States will act decisively to maintain global leadership in AI. A plan to support this will be established within 180 days.

Meanwhile, the AI startup 'DeepSeek,' founded by the Chinese quant trading firm High Flyer, has introduced a high-performance AI model without using NVIDIA's advanced semiconductors. The reason Silicon Valley and the world are on edge is that DeepSeek's inference-specialized model 'DeepSeek-R1' slightly outperforms OpenAI's initial inference model 'o1.' We will examine the related implications.

First, DeepSeek-R1, which delivered efficient results with low-spec chips, may spark controversy regarding excessive AI investment in the U.S. The New York Times reported that the development cost of DeepSeek-R1 was about one-tenth of what Meta spent building its AI technology. Attention was drawn to the 'Mixture-of-Experts' technology, which can create very powerful AI models with relatively low computing power.

This technology selectively activates only the 'expert' models suitable for specific tasks, reducing computation during training and inference to enhance performance and efficiency. It is necessary to closely watch whether small app developers switch from OpenAI or Anthropic to DeepSeek. However, it is not advisable to prematurely underestimate the importance of AI hardware chips. Technologies like ChatGPT, robotics, and autonomous driving can further advance with high-spec chips.

Second, the potential success of open-source models rather than closed models in AI development. Advocates of open-source models argue that anyone can modify and redistribute the code, allowing technology to develop rapidly and democratically. Of course, there are also claims that keeping AI model code private is safer. At this point, we need to check whether the open-source community leadership is shifting to China. If China accelerates new technology development through the power of open source, which generates new ideas and builds upon others' work, it poses a significant threat to the U.S.

Meta CEO Mark Zuckerberg has emphasized the importance of open source. Meta has utilized AI across its business, centered on Meta AI used in Facebook, Instagram, and WhatsApp. Meta's Llama 3.1 is an open-source model trained on more than 16,000 NVIDIA H100 GPUs. DeepSeek reportedly used only 2,000 of NVIDIA's lower-spec H800 chips, raising controversy regarding costs compared to the H100.

Third, the power of AI industry personnel, the developers. DeepSeek engineers devoted themselves to research at wages incomparable to those of U.S. developers. Yet, they are free from any restrictions under Chinese government AI regulations. In China, AI is the highest-paying sector, surpassing finance. Still, their monthly salary is less than 3 million KRW, showing the researchers' great passion.

President Trump announced a plan the day after his inauguration to invest $500 billion (about 718 trillion KRW) in AI infrastructure construction through the Stargate project. OpenAI, the developer of ChatGPT, Oracle, a corporate software company, and Japan's SoftBank, which has strong capital, established the 'Stargate' joint venture to build data centers across the U.S. However, skepticism exists about the project's success due to the relatively limited capital strength of the three companies compared to the required investment scale. It is necessary to carefully consider President Trump's remark that DeepSeek should serve as a wake-up call to the U.S. industry.

Jo Won-kyung, Professor at UNIST and Director of the Global Industry-Academia Cooperation Center