Korea’s Trillion-Dollar Sovereign AI Investment: Nvidia Wins, Hynix Loses

What it claims

SemiAnalysis argues that a near-frontier open model can be pretrained from scratch for far less than most assume. Its example is Motif 3 from Motif Technologies, a Korean startup of under 30 people. Motif has raised $17M and has access to 768 B200s (under 2MW). The authors put the total compute cost of Motif 3, including experimentation, at about $15M at current prices. On the Artificial Analysis Intelligence Index, Motif 3 leads Upstage, the next-best Korean model, by 10 points. It is also ahead of Inkling and Nemotron 3 Ultra, which the authors call the two best American open-source models. SemiAnalysis calls Motif 3 the best non-Chinese open-source model.

How it works

Korea's "Independent AI Foundation Model" project, announced in June 2025, is a tournament rather than a single pick. Teams get subsidized compute, data and researchers. They are evaluated about every six months, and eliminated teams' resources go to the survivors.

• Field. Fifteen consortiums applied. Five were picked in August 2025: Naver Cloud, LG AI Research, SK Telecom, NC AI and Upstage.
• Round 1 subsidies. About 3,000 H100 equivalents were rented from SKT and Naver for the other three teams. About $45M went to shared Korean data such as books, news, broadcasts and government records. Each team got about $2M for its own data and about $1.4M for overseas hiring, which only Upstage used.
• Budget and timeline. The total government budget is about $350M. Two winners are due at the end of 2026 and get scale-up support through 2027.
• Scoring. 40% benchmarks, 35% expert review, 25% user testing.

Round 1 eliminated NC and also Naver. Naver was disqualified for using Qwen vision and audio encoders. The government ruled only after submissions that foreign architectures are allowed, but every component must be trained from initialized weights. Motif, spun out of the AI infrastructure company Moreh in February 2025, joined through a supplemental competition.

Evidence

All four remaining teams open-sourced models in July 2026. Results were announced on August 18. Motif ranked first on benchmarks but last on expert review and user testing, and was eliminated.

The authors call those two categories opaque:

• User testing was not blind, and its tasks and criteria are not published.
• Ten experts scored "development strategy", "future plans" and "ecosystem impact".
• SemiAnalysis reports hearing that reviewers worried Motif's technology could move abroad through foreign investors. It notes this is odd, since the competition requires all technical components to be open-sourced.

Infrastructure and vendors

In July 2026 Korea announced a $919B plan. Phase 1 assigns 5GW to SK Group, 2.4GW to GS Group and 1GW to Naver. SK Group also takes the remaining 10GW of phase 2. SemiAnalysis has identified three active phase 1 sites totaling 4.4GW. The authors expect much of this capacity to be sold to Anthropic or OpenAI.

Announced Nvidia deals include:

• 2GW of Rubin within SK Group's 5GW.
• 200MW for Naver.
• A Samsung AI factory with more than 50,000 GPUs.
• SK Telecom's 2GW DSX factory on Vera Rubin with SK Hynix HBM4.

SemiAnalysis estimates, without confirmation, that Nvidia secured a long-term SOCAMM agreement with SK Hynix. It also estimates favorable HBM pricing and large 2027 volume commitments for Nvidia. This is the basis for the "Hynix loses" headline.

Limits

The per-model sizes and the benchmark comparison table were images and are not in the fetched text. Motif 3's parameter count, licence and download location are therefore not given here. The authors say benchmarks can be gamed and concede that Motif 3 may be "benchmaxxed". The $15M training cost and the Hynix deal terms are SemiAnalysis estimates. Full site details are behind the paid Datacenter Model.