'DeepSeek Shock' Sends Ripples Through the AI Market
Examining Its Impact on the Domestic Industry
"Good Opportunity for Startups" Raises Expectations
As Chinese AI startup DeepSeek unveiled a model that reduces costs while increasing efficiency, expectations are rising in the domestic market that "a new path in AI competition has opened." Domestic companies, which had to engage in a "money war" against American big tech firms like OpenAI, now see a contest worth competing in.
According to the IT industry on the 31st, DeepSeek is known to have used Nvidia's low-spec AI accelerator 'H800' in the development process of its proprietary AI inference model 'R1.' The H800 is a product released by Nvidia to circumvent U.S. export restrictions to China and is evaluated to have about half the performance of the current flagship product 'H100,' which is used by major AI companies.
Until now, the industry has regarded securing high-performance AI accelerators as the top priority for developing generative AI. This is because the speed of AI development processes such as machine learning depends on the performance of AI accelerators. For this reason, the domestic AI industry has also prioritized securing Nvidia's high-performance chips, and the U.S. has controlled the export of AI semiconductors to adversarial countries like China.
Experts have evaluated that China's counterattack began around the time of the Trump administration's inauguration, when China was the perennial second-ranked AI country. Choi Kyung-jin, president of the Korean Artificial Intelligence Law Society and professor at Gachon University, said, "China's AI counterattack has begun in earnest," adding, "It shows the possibility that the second and third-ranked countries can catch up with the U.S.-led AI models." Kim Deok-jung, professor at Sookmyung Women's University and director of the Pub AI Research Institute, stated, "The emergence of DeepSeek will be a turning point for China to overcome U.S. protectionism," and added, "It is a matter of national pride between the two countries over who will have the upper hand in the future AI market and can serve as a measure to evaluate the effectiveness of U.S. policies."
So, what message does the 'DeepSeek shock' send to our AI ecosystem? The fact that DeepSeek developed R1 despite the absence of high-performance AI accelerators suggests that AI development capabilities may be considered more important than hardware acquisition. If the domestic industry benchmarks DeepSeek's case to enhance AI development capabilities, it can maintain competitiveness even by reducing semiconductor investment costs to one-tenth. Additionally, DeepSeek has the advantage of not using big tech companies' cloud infrastructure but allowing direct download for use within a company's intranet. This hints at the potential expansion of the on-device AI market, where AI functions operate within smart devices without internet connectivity. Given these significant advantages in terms of cost and stability, various portfolio compositions become possible from the perspective of domestic AI services.
President Choi said, "The expanded choice of open-source AI models with excellent performance for cost reduction can be a good opportunity for domestic startups," and added, "We need to benchmark the secret behind developing without cutting-edge GPUs to cultivate efficient large language model (LLM) development capabilities." However, he also noted, "Legal risks including bias in state-supervised or led open-source models, copyright, and personal information issues, as well as national security considerations, need to be addressed," recommending, "Appropriate measures to handle latent risks should be implemented simultaneously."
Professor Kim predicted the business direction of domestic AI companies, saying, "Not all companies need to build sovereign AI, so focusing on applied technologies is reasonable," and forecasted, "AI services specialized in one domain will eventually dominate that market."
© The Asia Business Daily(www.asiae.co.kr). All rights reserved.
