South Korea’s AI push tests sovereign compute strategy

South Korea is moving ahead with one of the most ambitious sovereign AI programs outside the US and China, seeking models that Korean organizations can train, modify, and operate without relying on foreign AI labs. The “Independent AI Foundation Model” project uses a tournament format, with teams receiving compute, data, and researcher subsidies and being reevaluated every 6 months as weaker contenders lose support.

The first round included rented ~3000 H100 equivalents from SKT and Naver for the other 3 contestants, ~$45M USD for shared data, ~$2M per company for additional data purchases, and ~$1.4M for overseas recruiting, which only Upstage accepted. Naver was later disqualified after using vision and audio encoders from Alibaba’s Qwen model family, while Motif Technologies emerged as a standout startup with a model that benchmarked ahead of leading American open source models despite limited funding and compute. Motif was then eliminated after scoring last in expert review and user testing, a result that could undercut Korea’s effort to keep promising domestic AI talent anchored at home.

The infrastructure plan is much larger than the model tournament. South Korea announced a $919B investment to build 8.4GWs by 2029 and 18.4GWs by 2035, with SK Group, GS Group, and Naver leading the first phase. The buildout gives Nvidia a major sovereign AI customer as SK Group commits to 2 GWs of Rubin and Naver signs on for another 200MW, while deeper ties with Samsung and SK Hynix could improve Korea’s access to advanced compute but may also give Nvidia leverage over HBM pricing and future memory supply agreements.