South Korea's trillion-dollar sovereign AI investment: Nvidia wins, SK Hynix under pressure

South Korea is elevating artificial intelligence to a national strategy, betting on it on a scale that is drawing market attention. But in this nationwide race, who will truly benefit is a question far more complex than it appears.

In July of this year, the South Korean government announced plans to invest $919 billion to build 8.4 gigawatts of data center computing power by 2029, and eventually expand the total scale to 18.4 gigawatts by 2035.

This plan will open up a huge sovereign AI market for Nvidia. SK Group has committed to purchasing 2 gigawatts of Rubin systems, and Naver has also signed a 200-megawatt order.

However, this isn't entirely positive news for Samsung and SK Hynix . South Korea lacks domestic suppliers capable of independently producing cutting-edge AI accelerators. According to SemiAnalysis, Nvidia may have reached a long-term agreement with SK Hynix on SOCAMM and secured preferential pricing for HBM in 2027.

Meanwhile, the South Korean government-led "Independent AI Foundation Model" competition is underway, but it has made a perplexing decision in the most crucial elimination round—excluding Motif Technologies, a small startup that ranked first in the benchmark test.

This result reveals the inherent contradiction in the sovereign AI race: there is often a significant gap between the government's initial intention to pursue technological independence and the actual logic of resource allocation.

For investors, South Korea’s trillion-dollar commitment means a significant expansion of Nvidia’s customer base, but for shareholders of Samsung and SK Hynix, this massive construction plan may not bring commensurate returns.

Why Nvidia strongly supports open source and sovereign AI

Nvidia CEO Jensen Huang recently published an article outlining the importance of open-source AI, which was subsequently co-signed by almost all major AI companies (except Anthropic). Behind this statement lies a clear business logic.

Selling cutting-edge model tokens at API prices represents the application scenario with the highest return on unit computing power. This means that Anthropic and OpenAI will account for an increasingly larger proportion of the net new computing power demand in the future, and Nvidia faces the risk of its customers becoming overly concentrated in a few ultra-large purchasers.

Even if we include ultra-large-scale cloud vendors in the statistics, there are only seven truly core customers. And apart from SpaceX, all of these seven are actively developing their own chips to reduce their dependence on Nvidia.

Sovereign AI is a key breakthrough for NVIDIA to diversify its customer base. The classic framework of technology competition strategy, "commoditize your complements," also applies here: AI models and applications are complements to NVIDIA GPUs, and the flourishing of open-source models means a broader demand for computing power.

Nvidia's recent announcement of a $500 billion memorandum of understanding with the world's largest capital allocator is a financial extension of this strategy—by making "Nvidia AI Factory computing power" an "investable asset class," allowing pension funds, insurance companies, and private credit funds to flow into large-scale computing power purchases.

Data center construction boom: Nvidia benefits, HBM's position remains uncertain.

South Korea has confirmed three data center sites under construction, totaling 4.4 gigawatts. Regarding the construction entities, SK Group, GS Group, and Naver are responsible for the initial construction of 5 gigawatts, 2.4 gigawatts, and 1 gigawatt, respectively. SK Group will also undertake the remaining 10 gigawatts in the second phase. South Korean President Lee Jae-myung has explicitly positioned AI infrastructure as a historic project comparable to the broadband rollout of the 1990s.

However, this massive construction project does not bring simple benefits to South Korean semiconductor companies.

South Korea currently lacks domestic suppliers capable of independently producing cutting-edge AI accelerators. While Rebellions and FuriosaAI have made progress, they still lag significantly behind Nvidia in terms of deployment scale, software maturity, and commercial availability. Therefore, no matter how much investment is made, South Korea cannot build a competitive accelerator ecosystem solely relying on its domestic manufacturing capabilities.

This is precisely the strategic value of Nvidia's deepening relationships with Samsung and SK Hynix. Samsung is reportedly planning to build an AI factory equipped with over 50,000 Nvidia GPUs; SK Telecom's 2-gigawatt DSX AI factory is expected to deploy Vera Rubin systems powered by SK Hynix's HBM4.

According to SemiAnalysis, Nvidia may have reached a long-term agreement with SK Hynix on SOCAMM and secured preferential pricing and large-scale procurement commitments for HBM in 2027, although the final negotiation results may still differ from current estimates.

If this assessment is true, South Korea's massive infrastructure construction, while strengthening Nvidia's market position, may have far more complex actual benefits for Samsung and SK Hynix shareholders than the figures on paper suggest.

South Korea's "Squid Game": The Story Behind the Sovereign AI Championship

In June 2025, the South Korean government officially launched the "Independent AI Foundation Model" project, with the goal of developing a large-scale language model that South Korean institutions can train, modify, and operate independently without relying on foreign AI labs.

The most distinctive feature of this project is its competition format. Instead of pre-selecting a "national team," the government adopts an elimination tournament model: all participants receive subsidies in three areas—computing power, data, and research personnel—and are reviewed every six months, with the resources of the losers being redistributed to the winners.

A total of 15 consortia registered for the competition. After an initial review of their documents, 10 were selected, and in August 2025, the final five participating teams were determined: Naver Cloud, LG AI Research, SK Telecom, NC AI, and Upstage.

In the first round, the South Korean government provided each team with approximately 3,000 H100 equivalent computing power and invested approximately US$45 million in purchasing data, while also providing each company with approximately US$1.4 million in overseas talent recruitment funds.

However, an unexpected turn of events occurred at the end of the first round. NC AI, which ranked last, was eliminated, but Naver Cloud was also disqualified.

The South Korean government subsequently added Motif Technologies as the fourth participant. The company was spun off from Moreh, a South Korean AI infrastructure software company, in February 2025.

The best open-source model was surprisingly excluded, sparking controversy.

Four teams publicly released their latest models in July 2026, and the judging results were announced on August 18. Data shows that the two startups, Motif and Upstage, significantly outperformed the team backed by a large conglomerate, and their models had less than half the number of parameters as the latter, making the gap even more pronounced.

According to SemiAnalysis data, Motif 3 leads Upstage by 10 points in the Artificial Analysis Intelligence Index, making it the best Korean model and surpassing the two best open-source models from the United States, Inkling and Nemotron 3 Ultra.

Motif is a startup with fewer than 30 employees, only $17 million in funding, and less than 768 B200 GPUs available for computing power. The total computing cost for training Motif 3 (including the experimental phase) is approximately $15 million.

Therefore, Motif's elimination was quite unexpected. According to the competition's scoring rules, benchmark testing accounted for 40%, expert review for 35%, and user testing for 25%. Motif ranked first in the benchmark testing but last in the latter two, ultimately leading to its elimination.

The specific methodology used in the review process was noticeably opaque. During user testing, neither the testing tasks nor the evaluation criteria were disclosed, and the testing was not blind, or showed bias towards larger, more well-known companies.

It is understood that the expert judges also expressed concern that Motif's technology might be leaked to foreign investors—but this clearly contradicts the competition's requirement that all technical components be open source.

In addition, some scoring dimensions, such as "ecosystem impact" and "future planning," structurally favor large companies such as LG and SKT.

Now, Motif may be forced to relocate outside of South Korea to obtain the capital and computing power needed to continue its research—a truly ironic ending for the dark horse that has been the brightest in this national AI race.

Risk Warning and DisclaimerInvesting involves risk; please exercise caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Any investment decisions made based on this information are at your own risk.