Add preferred source
South Korean AI infrastructure companies and the government warned that the country must break free from its dependence on an Nvidia-centered GPU supply chain. Moreh CEO Cho Kang-won noted that if a single vendor halted GPU supplies for a year, South Korea could not build data centers at all, while MangoBoost CEO Kim Jang-woo identified AI cost efficiency as the core challenge. South Korea’s Ministry of Science and ICT said it will build a large-scale demonstration reference architecture next year linking AMD CPUs and GPUs with South Korean-made NPUs, and will support NPU-based infrastructure from a total cost of ownership perspective as the inference market takes off.
Key Elements
South Korean artificial intelligence (AI) infrastructure companies and the government have simultaneously issued warnings that the country must move away from a graphics processing unit (GPU) supply chain dominated by Nvidia. The message: break the structure in which a single supply decision by one semiconductor company determines whether South Korea can build data centers at all, and quickly secure alternatives that combine South Korean-made neural processing units (NPUs) with open software ecosystems.
Cho Kang-won, CEO of Moreh, said at a panel discussion at the “AMD Advancing AI World Tour Seoul 2026” held at the Westin Seoul Parnas hotel in Gangnam District, Seoul, on November 4: “If a certain company decides not to supply GPUs to South Korea for a year, we currently have no way to build a data center.” The discussion, moderated by Jang Jung-hee, CEO of AMD Korea, also featured Kim Jang-woo, CEO of MangoBoost, and Lee Do-gyu, director-general of the Information and Communications Policy Office at South Korea’s Ministry of Science and ICT.
Cho stressed that while data centers were once
Cho stressed that while data centers were once designed around whichever GPU operators wanted to use, the industry now needs the capability to run infrastructure on whatever semiconductors are available. GPU demand is surging while supply remains limited, he argued, making it difficult to secure infrastructure on time through reliance on specific products. “The idea that a model runs well only on Nvidia GPUs and won’t work on others is not true,” he said, while adding that “running a model and achieving high performance are different matters.”
The South Korean government has set open AI infrastructure as the axis of its support. Lee said: “It matters where sensitive data — in defense, security, and medical care — is trained and stored,” adding that South Korea is working to build an environment where critical sensitive information can be processed domestically. “That doesn’t necessarily mean it can only be done with South Korean technology,” he said. “What matters is securing choices while using our technology within the global open ecosystem.”
The government is shifting its support framework from chip-level assistance to system-level demonstrations. Lee said the market will flip by 2030, when 60% of AI workloads will go to inference and 40% to training. “To solve the enormous power and cost problems that arise when inference surges, securing efficiency through NPUs is essential,” he said. “Until now, demonstration projects have stayed at the chip-support level, which has limited market expansion. Next year we will build a large-scale demonstration reference architecture linking AMD central processing units (CPUs) and GPUs with South Korean-made NPUs, so that public institutions, private companies, and even global buyers can confidently choose South Korean NPUs.”
Cooperation between AMD and South Korean semiconductor companies
Cooperation between AMD and South Korean semiconductor companies is also taking shape. Lee described the memorandum of understanding (MOU) signed between the Ministry of Science and ICT and AMD as unusual in explicitly including cooperation between AMD’s CPUs and GPUs and South Korean NPUs, and said cooperation is already under way in the field. With the MOU signed in July covering cooperation with South Korean NPUs, next year’s demonstration project is expected to be the first test of results.
Cost efficiency has also emerged as a core agenda item. Kim Jang-woo, CEO of MangoBoost, said: “Data sovereignty matters, but the more fundamental problem is that AI is too expensive,” adding that “some South Korean conglomerates have brought in large numbers of GPUs but fail to deliver proper performance.” He said his company has improved performance per dollar by more than twofold by raising customers’ GPU efficiency and supplying its own hardware. “Just as you have to build a complete car and enter it in a race to finish first, we must build a complete computer system spanning semiconductors, software, and storage to prove cost-effectiveness,” he said.
The panelists framed the coming year as a golden window for infrastructure transition. Cho urged companies to build their own data centers and test alternative semiconductors, while Kim called for completing a single system that proves service cost-effectiveness. Lee said: “AI is advancing so fast that making the right decision at the optimal time is critical. When the inference-centered market fully opens, the most economical choice from a total cost of ownership (TCO) perspective will be South Korean NPUs combined with an open software ecosystem.”
The discussion once again highlights the structural vulnerabilities of South Korea’s AI infrastructure market. As AI demand shifts from training to inference, computing resource requirements will grow, meaning reliance on a single vendor translates directly into risks to service costs and supply schedules. Whether South Korean NPUs can demonstrate advantages in performance per watt and total cost of ownership in large-scale demonstrations is expected to be a watershed for the direction of South Korea’s data center investment and the reshaping of its semiconductor ecosystem.
