Calls grow for stronger AI infrastructure in Korea amid rise of DeepSeek

Chinese chatbot's breakthrough sparks sense of urgency among domestic developers

By Nam Hyun-woo

Calls are growing to improve Korea’s infrastructure for artificial intelligence (AI), as the rise of Chinese AI service DeepSeek presents both concerns and hopes for domestic AI developers.

The Ministry of Science and ICT on Thursday held a meeting with domestic developers in Seoul and collected the private sector's opinions on the current state of the nation's infrastructure for AI.

According to the ministry, participating companies and institutions, including Naver, Kakao, LG AI Research, SK Telecom and KT, shared their views that DeepSeek’s recent achievements present “both a crisis and opportunities” for Korean companies.

They also expressed opinions that the government should make efforts to form an AI ecosystem through infrastructure investments.

DeepSeek’s AI models created a shock in the global tech world by demonstrating performances on par with ChatGPT with up to 95 percent less costs, and without reportedly using the most advanced AI processors or graphics processing units (GPUs). In its technical report, DeepSeek claimed to have pretrained its V3 model on only 2,048 Nvidia H800 GPUs, a scaled-down version of the mainstream H100 or H200, with each chip costing about $2 per hour to run.

DeepSeek’s rise is seen as a model case of overcoming investment gaps through innovation, encouraging domestic AI developers to achieve noteworthy outcomes without the same level of resources as global big tech companies. At the same time, however, this sparked a sense of crisis, with Korean developers feeling they were lagging behind their Chinese rivals.

During the ministry meeting, LG AI Research head Bae Kyung-hoon revealed that LG’s high-performance Exaone 3.5 32B model, made public as an open-source model in December, cost only 7 billion won ($4.83 million) to develop — lower than the known development cost of $6 million for the DeepSeek V3 model.

“Exaone 3.5 also uses the Mixture of Experts technique, which was a key factor in DeepSeek’s successful low-cost development,” Bae said. “If the model gained more global attention and is promoted better, it could have gained more recognition.”

Mixture of Experts is a machine learning technique that divides the AI model into subnetworks.

AI industry officials are raising concerns that adequate government support is crucial to harness the potential of Korean AI developers and help them compete in the global market.

During a recent National Assembly meeting earlier this week, Naver Cloud Head of AI Innovation Ha Jung-woo said DeepSeek V3’s $6 million development cost only represents the successful model, pointing out that countless additional expenses in failures were likely incurred behind the scenes.

Ha stressed the need for national-level infrastructure expansion, citing the costs of securing advanced GPUs. He said purchasing 20,000 H100 GPUs costs approximately 1 trillion won, which is a major challenge for AI developers.

Though many Korean tech companies are now promoting AI as their main business product, the country’s infrastructure remains disappointing. According to the science ministry’s 2023 survey, Korea had only 1,961 H100 processors, while most GPUs that Korean companies had were a simplified version of V100 processors.

On the other hand, global tech firms are quickly expanding their AI infrastructure. Citing market researcher Omdia, the Financial Times reported that Microsoft bought 485,000 Nvidia Hopper chips in 2024, while China’s Tencent and ByteDance ordered approximately 230,000 Hopper GPUs each last year.

To level the AI playing field, the government on Tuesday announced its plan to purchase 30,000 advanced Nvidia GPUs by 2026, securing 15,000 this year.

“If the government and private sectors work together to foster an AI ecosystem, we can achieve our own success amid fierce global competition,” Second Vice Minister of Science and ICT Kang Do-hyun said at Thursday’s meeting.

“The government plans to accelerate the establishment of a national AI computing center and will discuss detailed strategies for Korea’s advancement as an AI powerhouse at the upcoming national AI committee meeting later this month.”