South Korea commits 260,000 GPUs to build domestic foundation models
South Korea's government will double down on support for domestic artificial intelligence development, with Science Minister Bae Kyung-hoon announcing on May 29 that the country aims to become a "full stack" supplier of AI technology. The government secured more than 260,000 GPU units from NVIDIA last year and is building additional data centers to support the effort.
The announcement signals a national-level resource commitment to foundation-model development. Bae said South Korea should develop a frontier foundation model comparable to those from the United States and China, citing Stanford University's recognition of South Korea as the third-largest AI power last year.
Four teams selected for first round
A government-led selection process for homegrown foundation models advanced four teams from the first round: LG AI Research, SK Telecom, Upstage, and Motif Technologies. A second-round announcement is expected in early August.
The ministry is also emphasizing use of locally produced neural processing units (NPUs) alongside GPU infrastructure. Bae noted that government AI budget allocation remains modest compared to spending by individual U.S. technology companies.
What practitioners should track
Foundation-model development at national scale typically encounters constraints around GPU availability, data center power and cooling, and integration of alternative accelerators. These constraints often influence model size, training frequency, and cost structure.
Observers should monitor the selection process timeline, actual deployment schedules for the announced GPUs, and any technical disclosures from the selected teams about model scale, training frameworks, and NPU integration.
For government officials overseeing AI policy and procurement, this initiative reflects the infrastructure and sustained investment required for large-model development. Learn more about AI for Government and Generative AI and LLM strategies.
Your membership also unlocks: