South Korea's decision to elevate artificial intelligence to a national strategic priority has captured global attention, with the scale of its investment drawing considerable market scrutiny. Yet, beneath the surface of this national campaign, the question of who will truly benefit is far more intricate than it initially appears.

In July of this year, the South Korean government unveiled an ambitious plan to invest $919 billion, with the objective of constructing 8.4 gigawatts of data center computing capacity by 2029 and ultimately expanding the total scale to 18.4 gigawatts by 2035.

This initiative is poised to open a vast sovereign AI market for NVIDIA, with the SK Group having already committed to procuring 2 gigawatts of Rubin systems, while Naver has also signed a 200-megawatt order.

However, this development does not represent an unqualified positive for Samsung and SK hynix. South Korea currently lacks a domestic supplier capable of independently producing cutting-edge AI accelerators. According to SemiAnalysis, NVIDIA may have already secured a long-term agreement with SK hynix regarding SOCAMM, along with preferential pricing for HBM that has been locked in for 2027.

Meanwhile, the government-led competition to develop an "independent AI foundation model" is advancing, yet it has produced a bewildering decision at the most critical elimination stage—excluding Motif Technologies, a small startup that topped the benchmark rankings.

This outcome illuminates an inherent contradiction within the sovereign AI race: a clear gap frequently exists between the government's stated goal of technological independence and the actual logic driving resource allocation.

For investors, Korea's trillion-dollar commitment signals a significant expansion of NVIDIA's customer base. However, for shareholders of Samsung and SK hynix, this massive construction plan may not yield proportionate returns.

Why NVIDIA champions open-source and sovereign AI

NVIDIA CEO Jensen Huang recently published an open letter articulating the importance of open-source AI, a statement that subsequently received endorsements from nearly all major AI companies, with the notable exception of Anthropic. Behind this public stance lies a clear commercial rationale.

Selling tokens for frontier models at API prices represents the application scenario with the highest return per unit of computing power. This implies that Anthropic and OpenAI will account for an ever-increasing proportion of net new computing demand in the future, exposing NVIDIA to the risk of excessive customer concentration among a handful of ultra-large-scale purchasers.

Even when hyperscale cloud providers are included in the statistics, the number of truly core customers amounts to no more than seven, and all but one of these seven—with the exception of SpaceX—are actively developing their own in-house chips to reduce their reliance on NVIDIA.

Sovereign AI represents a pivotal avenue for NVIDIA to diversify its customer base. The classic framework of "commoditize your complements" applies equally here: AI models and applications serve as complements to NVIDIA's GPUs, meaning the proliferation of open-source models translates into broader demand for computing power.

NVIDIA's recent announcement of a $500 billion memorandum of understanding with the world's largest capital allocator represents the financial extension of this strategy—by establishing "NVIDIA AI factory computing power" as an "investable asset class," the company enables pension funds, insurance companies, and private credit capital to flow into large-scale computing procurement.

Data center construction boom: Benefits for NVIDIA, unresolved questions for HBM landscape

South Korea has confirmed three data center construction sites that are currently underway, collectively totaling 4.4 gigawatts. Regarding construction entities, SK Group, GS Group, and Naver are responsible for the initial phases of 5 gigawatts, 2.4 gigawatts, and 1 gigawatt respectively, with SK Group also undertaking the remaining 10 gigawatts in the second phase. President Lee Jae-myung has explicitly positioned AI infrastructure as a historic undertaking comparable to the broadband expansion of the 1990s.

Nevertheless, this large-scale construction plan does not bring an unmitigated boon to South Korea's domestic semiconductor enterprises.

The country still lacks a local supplier capable of autonomously manufacturing frontier-level AI accelerators. Rebellions and FuriosaAI have made progress, but they remain significantly behind NVIDIA in terms of deployment scale, software maturity, and commercial readiness. Consequently, regardless of the capital invested, South Korea cannot rely solely on its domestic manufacturing capabilities to establish a competitive accelerator ecosystem.

This is precisely where the strategic value of deepening ties between NVIDIA and both Samsung and SK hynix lies. Samsung is reportedly planning to construct an AI factory equipped with more than 50,000 NVIDIA GPUs, while SK Telecom's 2-gigawatt DSX AI factory is expected to deploy Vera Rubin systems featuring SK hynix's HBM4.

According to SemiAnalysis, NVIDIA may have already entered into a long-term SOCAMM agreement with SK hynix, securing preferential HBM pricing and substantial purchase commitments for 2027, although final negotiation outcomes could still differ from current estimates.

If this assessment holds true, then this round of infrastructure expansion in Korea, while strengthening NVIDIA's market position, presents a considerably more nuanced picture of actual returns for shareholders of Samsung and SK hynix than the headline figures might suggest.

Korea's 'Squid Game': The sovereign AI championship explained

In June 2025, the South Korean government formally launched the "Independent AI Foundation Model" project, with the goal of developing a large language model that Korean institutions can autonomously train, modify, and operate without reliance on foreign AI laboratories.

The project's most distinctive feature lies in its tournament-style design. Rather than pre-selecting a "national team," the government adopted an elimination-format competition: all participants receive subsidies in computing power, data, and research personnel, undergo evaluations every six months, and the resources of those eliminated are reallocated to the winners.

A total of 15 consortia registered for the competition, with 10 advancing past the initial document screening. In August 2025, five teams were ultimately confirmed: Naver Cloud, LG AI Research, SK Telecom, NC AI, and Upstage.

In the first round, the government provided each team with approximately 3,000 H100-equivalent GPUs, allocated around $45 million for data procurement, and offered each company roughly $1.4 million for overseas talent recruitment.

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

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

Top-performing open-source model faces unexpected elimination, raising questions

The four teams publicly released their latest models in July 2026, with evaluation results unveiled on August 18. The data revealed that the two startups, Motif and Upstage, significantly outperformed their conglomerate-backed counterparts, a gap made even more pronounced by the fact that their models used less than half the parameter count of the larger teams' models.

According to SemiAnalysis data, on the Artificial Analysis intelligence index, Motif 3 surpassed Upstage by 10 points, establishing itself as the best Korean model and exceeding the performance of the two best American open-source models, Inkling and Nemotron 3 Ultra.

Motif is a startup with fewer than 30 employees, having raised only $17 million in funding. Its accessible computing capacity is below 768 B200 GPUs, and the total cost of training Motif 3, including experimental phases, amounts to approximately $15 million.

It is precisely for this reason that Motif's elimination comes as a considerable surprise. Under the competition's scoring rules, benchmarks account for 40%, expert evaluation 35%, and user testing 25%. Motif ranked first in benchmarks yet finished last in the other two categories, ultimately leading to its disqualification.

A notable lack of transparency surrounds the evaluation methodology. In the user testing phase, neither the test tasks nor the judging criteria were disclosed, and the tests were not conducted as blind tests, potentially introducing bias in favor of higher-profile larger companies.

It has also been reported that the expert evaluation raised concerns about the potential outflow of Motif's technology through foreign investors—a stance that appears to directly contradict the requirement that all technical components be open-source.

Furthermore, certain evaluation dimensions, such as "ecosystem influence" and "future plans," structurally favor large enterprises like LG and SKT.

Now, Motif may be compelled to relocate outside of South Korea to secure the capital and computing power necessary to continue its research—a conclusion that constitutes perhaps the most ironic outcome for the most promising dark horse in this national AI competition.