Why South Korea's Sovereign AI Push May Benefit NVIDIA First, Not SanDisk

· South Korea plans to establish a strategic account of at least 20 trillion won under KIC by 2027, investing in strategic industries such as artificial intelligence, data centers, and semiconductors.
· SemiAnalysis believes NVIDIA is more likely to benefit first, with the controversy focusing on whether South Korea can break free from GPU and CUDA ecosystem lock-in.
· Related tickers: NVIDIA (NVDA), SK Hynix, Samsung Electronics, and the South Korean semiconductor and data center supply chain.

The South Korean government plans to set up a new strategic account under the Korea Investment Corporation (KIC) with a size of at least 20 trillion won, investing in long-term areas such as artificial intelligence, data centers, semiconductors, robotics, and defense, with a target launch in 2027, pending parliamentary approval of the relevant legislation.

This initially appears to be an industrial policy news. South Korea is redirecting the fiscal space generated by the semiconductor boom into the next round of AI infrastructure. However, SemiAnalysis offers a sharper judgment: in this round of South Korea's "sovereign AI" investment, NVIDIA wins, and SK Hynix loses.

This statement is counterintuitive. South Korea is one of the world's strongest memory chip nations, and SK Hynix is a core supplier of HBM (high-bandwidth memory). The divergence lies in whether South Korea is acquiring stronger domestic AI capabilities or becoming more deeply embedded in NVIDIA's GPU and CUDA (NVIDIA's software ecosystem) system.

For investors, this determines whether national-level AI spending ultimately flows to platform companies or remains more concentrated in the local hardware supply chain.

Turning Chip Boom into Long-Term AI Capital

The core of South Korea's policy this time is not just increasing government spending, but converting the tax revenue and assets generated by the semiconductor cycle into more patient long-term capital.

The 20 trillion won strategic account will be established within KIC. KIC originally handled South Korea's overseas investment functions, but this time it has been given a domestic strategic investment role, indicating that the government hopes it will first provide a capital base and then attract enterprises and external funds to follow.

AI data centers, advanced semiconductor clusters, physical AI, and domestic model research and development are all difficult to recover stably within a short cycle. If government funds only focus on short-term returns, it will be hard to sustain the construction period of such projects.

On September 1, South Korea announced the Future Response Fund related to the 2027 budget, with a size of 162.3 trillion won, funded by above-trend tax revenues from the semiconductor boom and other sources. Not all of it is directed to AI and semiconductors; the growth engine account is approximately 14.2 trillion won, used to support AI, semiconductors, and other directions.

The "three major projects" proposed in June also set data center targets of 8.4GW by 2029 and 18.4GW by 2035. This pace is very aggressive, and whether it can be realized depends on electricity, land, grid connection, and equipment delivery.

South Korea is not just subsidizing a single company; it is using fiscal and chaebol investment to rebuild AI infrastructure. SemiAnalysis's counterpoint is that infrastructure localization does not equal technology stack localization. Data centers may be built in South Korea, but the racks may still run NVIDIA GPUs.

Data Center Expansion Flows First to NVIDIA Stack

The most expensive and hardest-to-replace part of AI data centers is the accelerated computing platform. Today, most large model training and inference software is optimized around NVIDIA GPUs and the CUDA ecosystem.

This is platform lock-in. GPU is hardware, CUDA is the software interface and development ecosystem. Developers, model frameworks, operations tools, and performance optimization experience are all accumulated within this system.

Switching to other hardware is not just replacing a chip. Models need to be retuned, systems need to be retested, and operations teams need to re-accumulate experience. National-level projects that want to build usable computing power within a few years usually prioritize purchasing the most mature solutions currently available.

This logic already has real-world clues. NVIDIA and SK Group announced expanded cooperation in July, covering AI factories, AI clouds, and next-generation memory, with a total scale exceeding $500 billion. Among them, SK Telecom plans to build a 2GW NVIDIA Vera Rubin DSX AI factory and AI cloud, using SK Hynix HBM4, with the first AI factory planned to go online in 2027.

This does not mean South Korea has no independent path. South Korea can still accumulate capabilities in packaging, memory, data center operations, domestic applications, and model training. But under current technological realities, the faster data centers are built, the more likely they are to first strengthen NVIDIA's demand visibility.

SK Hynix Wins Demand, Pricing Power Remains with Platform

"SK Hynix loses" should not be understood as business damage. If South Korea's data center construction materializes, SK Hynix's HBM demand will be directly supported.

HBM is the key supporting memory for AI GPUs. Without sufficiently high-speed memory, GPU computing power cannot be fully unleashed. SK Hynix is in a leading position in HBM supply, and the larger the NVIDIA platform scale, the stronger the demand for high-end HBM.

The issue lies in the different positions of value capture. HBM suppliers earn component value, while NVIDIA earns platform value. The former benefits from volume, price, and supply tightness, while the latter controls hardware architecture, software ecosystem, developer interfaces, and system solutions.

SemiAnalysis's judgment can be understood this way: SK Hynix can win revenue, but in the narrative of whether South Korea masters the AI sovereignty stack, its voice may be weaker than NVIDIA's. The dominance in computing power procurement, software migration costs, and developer ecosystem does not lie with HBM suppliers.

Samsung Electronics is in a similar situation. It has memory, foundry, and system semiconductor ambitions. But to weaken NVIDIA's dominance in the AI full stack, it needs not only capital expenditure but also alternative software ecosystems, customer validation, and long-term stability.

Order Share Will Bring the Debate Back to Reality

This controversy is not yet settled. The 20 trillion won strategic account does not mean cash will be immediately invested in GPU procurement; the funds still need to go through legislative approval, asset arrangements, and specific project screening.

Data center targets also face physical constraints. 8.4GW by 2029 and 18.4GW by 2035 are tests for South Korea's electricity, land, cooling, grid connection, and local operational capabilities. The larger the target, the higher the execution risk.

The most direct verification variable is the procurement share of platforms like NVIDIA Rubin in future projects, and whether South Korea's domestic models and open technology routes can reduce dependence on CUDA in actual scenarios. If orders continue to concentrate on the NVIDIA stack, SemiAnalysis's judgment will be closer to reality.

What can be confirmed at this stage is that South Korea is converting the semiconductor boom into national-level AI infrastructure capital. Under the current technology stack, this capital first enhances the demand visibility of the NVIDIA platform. SK Hynix remains one of the key winners, but it wins in the high-speed memory segment, not the pricing power of the entire AI computing stack.