South Korea's GPU sovereign computing power project extended with positive reviews: next year's budget of $300 million to buy GPUs, but space is almost full.

The South Korean government plans to extend its AI core GPU support program. Next year’s budget could reach as high as 4 trillion won (about NT$8.6 billion), but available installation space for devices is about to become saturated, which could become a hidden concern.
(Background: Unfazed by the big chip-stock shakeout—Goldman Sachs urges South Korea’s Kospi index to hit 12,000 points, revealing 3 major bullish catalysts.)
(Background add-on: Samsung surges on memory—up 20%! UBS raises its price targets: DRAM up 32% quarter over quarter in Q3, and NAND up 30%!)

Table of Contents

Toggle

• Program mechanism: The government buys GPUs, cloud providers set up cloud infrastructure, shared use across industry, academia, and research
• The meaning of 4 trillion won: roughly US$320 million
• Hidden concern: installation space is nearing saturation
• Taiwan can take cues: where is our GPU shortfall?

The South Korean government plans to extend its AI core GPU support program, with next year’s budget potentially reaching as high as 4 trillion won (about NT$8.6 billion). The program invested 1 trillion won last year and increased to 2 trillion won this year, already securing more than 20,000 GPUs. But as demand for advanced resources such as NVIDIA’s next-generation GPU Vera Rubin is surging, the government is investigating the expansion capacity of major cloud service providers. ETNews reported on July 7 that South Korea’s Ministry of Science and ICT has filed an application for “a quota-limited additional budget” with the Budget Planning Division.

Program mechanism: The government buys GPUs, cloud providers set up cloud infrastructure, shared use across industry, academia, and research

The core model of the program is that the government funds the purchase of GPUs, deploys them in cloud providers’ data centers, and then makes them available for industry, academia, and research use. This “government procurement + cloud deployment + shared use” model is similar to the operating logic of Taiwan’s National Supercomputing Center, but on a larger scale and with faster speed.

According to industry information, major cloud service providers’ additional construction capacity is currently being assessed. The government not only wants more GPUs, but also needs to ensure that these GPUs are used effectively rather than left idle.

The meaning of 4 trillion won: roughly US$320 million

Based on the current exchange rate (1 US dollar is about 1,310 Korean won), 4 trillion won is about 305 million US dollars. Using an estimated unit price of around US$12,000 for an NVIDIA H200 GPU, in theory, about 25,000 units could be purchased. If Vera Rubin (GB200-class) has a higher unit price, the actual number of GPUs that can be procured may fall between 15,000 and 20,000 units.

For comparison, Taiwan’s AI-related budget for 2025 is about 100 billion New Taiwan dollars (about 300 million US dollars). South Korea’s single-year GPU procurement budget is already on par with Taiwan’s overall AI budget.

Hidden concern: installation space is nearing saturation

From an industry perspective, compared with the GPU supply itself, “device installation space” is the next challenge. Running thousands of high-performance GPUs at the same time requires a large amount of power supply and advanced cooling systems. The device installation space held by companies in South Korea is already close to saturation.

This matches a common pain point across global AI infrastructure—last year, NVIDIA CEO Jensen Huang also said that the second half of the AI race is “power and cooling,” not simply the number of chips.

Taiwan can take cues: where is our GPU shortfall?

South Korea’s GPU support model offers lessons for Taiwan. Taiwan’s AI chip supply chain is strong in manufacturing (TSMC foundry) and memory (competition between Samsung and SK Hynix), but the government’s GPU computing power sharing initiative started relatively later. At present, Taiwan’s AI computing power is mainly concentrated in major cloud companies (GCS, Amazon) and a small number of startups, while GPU accessibility for small and medium-sized enterprises and academic institutions remains limited.

If South Korea can truly put up a GPU budget of 4 trillion won next year, the competitive landscape for AI computing power in Asia may further shift toward the Korean side.