Hello. I am reporter Ahn Hye-min, who handles and analyzes data. On June 28, the government announced a massive project. Named the "Three Mega Projects for South Korea's Great Leap Forward," it unveiled large-scale investment plans across three sectors: semiconductors, physical AI, and AI data centers. In today's OhGraph, we will break down this mega project. We will look at what the project entails, the background behind this decision, and the key points that must not be missed for this project to succeed, using various data and graphs.


Overwhelming Semiconductors... Widening the Gap with a Speed Battle
The very first area among the three major projects is semiconductors. As semiconductors are used everywhere in the AI era, OhGraph has mentioned multiple times that the world is suffering from supply shortages. At the Computex Taipei held in June, SK Chairman Chey Tae-won predicted that the memory shortage would continue until 2030. SK Hynix and Samsung Electronics, the overwhelming leaders in the memory sector, already had plans to expand their fab factories in Yongin to meet the supply shortage.

In the Yongin National Industrial Complex, six Samsung Electronics fabs are scheduled to be built, while four SK Hynix fabs are planned for the general industrial complex. Originally, Samsung Electronics aimed for completion by 2047 and SK Hynix by 2045. However, looking at how things are unfolding, the decision was made that they cannot afford to drag their feet.

For reference, SK Hynix's plan for the Yongin semiconductor cluster was first introduced in February 2019, but ground was broken only six years later, in February last year. Foundation work began in April this year. Samsung Electronics has not even broken ground yet. Therefore, the government's strategy for the semiconductor project is a "speed battle."

The completion timeline for the Yongin semiconductor cluster has been pulled forward significantly: Samsung Electronics shortened its timeline by 7 years to 2040, and SK Hynix by 12 years to 2033. In addition, there are plans to build a new semiconductor industrial complex in the southwestern region.

The plan is to build it on the site of the Gwangju military airport, which is three times the size of Yeouido, and this project is also being approached as a speed battle. First of all, since it is state-owned land, it has the advantage of skipping land compensation procedures. Additionally, because it is an airport site that has already been leveled, the plan is to complete it within the current administration's term.

The reason South Korea is speeding up despite already leading in the memory semiconductor sector is that competitor nations are all running fast. As AI transformation takes place across various industries and semiconductors emerge as core infrastructure, demand has surged, leading to bottlenecks and congestion everywhere. Consequently, many semiconductor companies are building factories extremely aggressively. This has evolved beyond mere investment competition between companies into an all-out national war.

Representative examples include the United States and China. Let's look at the investment situation of Micron Technology in the U.S. through the first graph.

Micron is currently pursuing investment plans worth 200 billion USD in the U.S. alone. It plans to build two fabs in Idaho and a mega-fab worth 100 billion USD in New York State. Micron's factory expansion is not limited to the U.S. In January this year, it acquired a fab from Taiwanese foundry PSMC for 1.8 billion USD. Furthermore, it plans to invest 1.5 trillion JPY in Hiroshima, Japan, to build a new HBM production facility.

The situation in China, as well as the U.S., is also unusual. ChangXin Memory Technologies (CXMT), China's largest DRAM maker, is making its presence felt with its first-quarter revenue this year surging by more than 700% year-on-year. While it previously produced DRAM at its Hefei and Beijing plants, it has recently been building a new DRAM fab in Shanghai. CXMT plans to list on the Shanghai Stock Exchange to secure the funds needed to expand its production lines. Not only CXMT, but other Chinese semiconductor companies, such as YMTC and Baidu's AI semiconductor subsidiary Kunlunxin, are also preparing for IPOs to secure funds as they join the expansion race.


Capturing the Physical AI Market with the Advantage of a Manufacturing Powerhouse
Next is physical AI, which refers to AI that understands and makes decisions in the physical environments of the real world.

Among the various cards South Korea holds, what else can we boast besides semiconductors? It is manufacturing. South Korea possesses a high level of manufacturing capability in various fields, including semiconductors, automobiles, shipbuilding, batteries, and defense. In this announcement, the government stated that it would accelerate the AI transformation of the manufacturing sector to maximize synergy between South Korea's flagship manufacturing industries and the robot industry.

The key to determining the performance of an AI model is data. The same goes for physical AI. However, compared to the Large Language Models (LLMs) we commonly use, data for physical AI is severely lacking. This is because while the internet is full of text data that companies can easily scrape, manufacturing data does not exist on the internet. This data is on the factory floor.

The government plans to drive AI transformation in the manufacturing sector by utilizing crucial data held by South Korean manufacturers, such as the temperature and pressure under which products are manufactured on each production line. Furthermore, since South Korea is already a robotics powerhouse where robots are deployed throughout various industries, it has the advantage that AI transformation can directly lead to increased productivity.

According to data released by the International Federation of Robotics (IFR), South Korea's robot density stands at 1,220 robots per 10,000 employees, ranking an overwhelming first. This is more than nine times the global average of 132.

However, the problem is that South Korea only uses robots. While other countries are racing into the physical AI era and producing robots like humanoids, South Korea is merely buying and using them. In the meantime, South Korea has focused only on utilizing robots well and has produced very little. Meanwhile, the humanoid market is dominated by China and the United States.

In the robot market report published by global market research firm Omdia, South Korean companies are nowhere to be found. From the top-ranked Agibot and second-placed Unitree to the sixth in market share, all are Chinese companies. China is overwhelming, accounting for 86.9% of the total. Following them, the seventh, eighth, and ninth-ranked companies are from the U.S., while the remaining companies account for about 9.8%.

Through this mega project, the government plans to develop industry-specific AI robots and supply more than 1,000 units to work sites every year. It also plans to secure technological competitiveness in the components used to build robots and the AI that will serve as the robot's brain.

The key company in the physical AI sector is undoubtedly Hyundai Motor Group. The processes of welding and assembling car bodies and the processes of manufacturing robots are highly similar, overlapping by 70% to 80%.

Hyundai Motor Group's robot foundry will be built in Saemangeum, and support will be provided to help automotive and home appliance component companies in the Daegu-Gyeongbuk region transition to robot components, thereby expanding the robot production base.


Building Domestic AI Data Centers and Even Exporting Them
If one had to choose the most necessary core infrastructure in the great AI transition, it would undoubtedly be AI data centers. This is because they are required to train models and perform inference. Looking at large-scale national projects, most of them are related to AI data centers. For instance, the U.S. Stargate project, which involves an investment of 500 billion USD, is about building AI data centers. It is also reported that China plans to invest 2 trillion CNY over the next five years to build AI data centers.

The government plans to build AI data centers across the country so that they can also contribute to regional revitalization. First, in Phase 1, it plans to build 8.4 GW of AI data centers in cooperation with SK, GS, and Naver.

In Ulsan, SK will build a 1 GW-class data center; GS will build a 2.4 GW-class one in Donghae; and Naver will build a 1 GW-class AI data center in Sejong. SK plans to establish 1 GW-class infrastructure not only in Ulsan but also in four other regions.

The number 1 GW is mentioned so often that it might be hard to grasp, but it is actually an enormous amount. A 1 GW-class AI data center requires the equivalent of one standard nuclear power plant. Currently, xAI and Amazon are the only ones operating gigawatt-class AI data centers globally.

This shows the power capacity of AI data centers currently operating worldwide. With xAI's Colossus 2 starting full-scale operations this year, it is serving as the world's first 1 GW-class AIDC (AI Data Center). Amazon's data center for Anthropic, "Project Rainier," has also recently seen its power capacity increase to the 1 GW level.

For reference, SK plans to go further and expand its AI data center capacity to 15 GW by 2035. If this happens, a total of 18.4 GW of AI data centers will be built in South Korea by 2035. Furthermore, the government plans to export these AI data centers to other countries. One might wonder what we can sell with AI data centers that haven't even been built yet. The key lies in "inference."

As mentioned earlier, AI data centers are needed to train models and use them in the inference process. In reality, there is little room for South Korea to break into the training stage. Although the country is developing models through projects like "Dokpamo" (Independent AI Foundation Model project), realistically, state-of-the-art AI models are tightly controlled by U.S. big tech companies that possess massive quantities of Nvidia's cutting-edge GPUs. However, the AI inference stage, where actual services are delivered after training is complete, could be different.

The inference stage refers to the process of inputting new data into an already trained model to obtain results. Generally, the entire process of us using AI corresponds to the inference stage. During inference, real-time interaction with the user must occur, making it difficult to use a distant AIDC. Especially in physical AI like robots, which goes beyond the level of chatbots, latency is highly sensitive, so using an AIDC located in a far-off country can be problematic. Additionally, the transmission of sensitive data back and forth to servers in other countries could also pose a problem.

In other words, the AI data centers where inference computations take place must be located near the users, and for security reasons, they need to be processed within domestic infrastructure. Since every country will likely need its own AI inference infrastructure in the future, South Korea aims to target this niche.

The plan is to first build AI data centers in South Korea, utilize domestic NPUs and servers to operate network technologies, power, and cooling solutions, and then package this operational technology to export it to other countries—much like how we export nuclear power plant technology abroad.


The Hurdles the Mega Project Must Overcome Are Also 'Mega' Sized
No one would object to the direction of actively participating in the AI transition, which has already become a reality, while simultaneously adhering to the principle of balanced regional development. However, because it is literally a mega-scale project involving a massive amount of capital, quite a few people are concerned about whether it can run smoothly without issues.

Among these, the most concerning point is the power issue. It remains a question how we will generate electricity for AI, which is a power-hungry beast. The power required for the semiconductor cluster to be built on the Gwangju military airport site is about 6.3 GW, and the newly constructed AI data centers across the country will require 18.4 GW by 2035 according to the plan. Combined, an additional power demand of 24.7 GW is expected. To put this into perspective, South Korea's peak power demand in August last year was 96 GW. This means we will need power equivalent to a quarter of that peak in the future.

For reference, the government forecasts power demand and establishes response policies every two years. It began establishing the 12th Basic Plan for Electricity Supply and Demand late last year and disclosed its forecast in April this year. The problem is that none of the aforementioned AI mega projects were reflected in this plan. Let's look at the graph to see how big the difference is.

According to the forecasts disclosed in the 12th Basic Plan for Electricity Supply and Demand, South Korea's peak power demand in 2050 is projected to be around 131.8 to 138.2 GW. Within this estimate, the consumption by data centers is a mere 4 GW. However, according to the plans of this mega project, data centers alone will require more than four times that amount of power.

Therefore, experts say that a major revision of the 12th Basic Plan for Electricity Supply and Demand is necessary. They predict that the forecasts could be raised significantly, and the existing plan centered on renewable energy could also be revised. Nuclear power, which can stably supply large-scale power demand without emitting carbon, is emerging as an alternative. Indeed, the Korean Nuclear Society is arguing that new nuclear power plants and Small Modular Reactors (SMRs) must be additionally reflected.

In addition to electricity, there is also the water issue. Data centers do not just consume a lot of electricity; they also consume an immense amount of water. The water used for server cooling must be extremely clean, requiring comprehensive consideration of water utilization plans. Furthermore, not only the water directly used in data centers but also the issue of indirect water usage has recently emerged.

Most big tech companies only disclose how much water they use for server cooling in their environmental reports. However, according to a recent report by a national laboratory under the U.S. Department of Energy, indirect water costs are substantial. An analysis by the Lawrence Berkeley National Laboratory revealed that the amount of water consumed by power plants to generate the electricity used by data centers is, on average, 12 times higher. In other words, running AI infrastructure hides a much larger consumption of water than what is visible.

For now, it appears the government plans to overhaul the national water management system. Currently, water management plans are established every 10 years, but there are limits to handling AI investments of this scale. It is expected to be revised to break the plans down into two-year units, similar to the power plans examined earlier, to secure more continuity.

Lastly, a noteworthy point is that this project is heavily concentrated on infrastructure, namely hardware. There is nothing wrong with infrastructure being important. However, the criticism is that it is too heavily skewed. For reference, a report published by Gartner in May advises against recklessly predicting or overestimating the long-term demand for AI infrastructure. This is because the leaders of the AI market are changing literally day by day.

How long will the AI model architectures that everyone is currently using remain on top? No one can guarantee how long the currently popular optimization technologies will remain effective. In a situation where even five-year forecasts carry high risks, we need to consider the inherent danger of rushing forward based on long-term demand predictions.

For instance, designing various scenarios to respond flexibly and building plans based on them could be one approach. Alternatively, a strategy could be established to secure diverse options even when adding infrastructure, such as adding infrastructure in a modular fashion and investing in stages.

Meanwhile, critics also point out that compared to the hardware into which massive amounts of money are being poured, the software strategy to run on top of it is relatively lacking.

The mega project has taken its first step. While they say a good start is half the battle, the future seems even more critical. Countless countries around the world are jumping into the AI infrastructure race, and the market landscape is changing day by day. Sincerely hoping that this project can flexibly adapt to such changes, I will wrap up today's OhGraph here. Thank you very much for reading this long article to the end.

References
- Report from the National Report Meeting on the Three Mega Projects for South Korea's Great Leap Forward
- Presentation from the Public Hearing on the 12th Basic Plan for Electricity Supply and Demand
- Micron global expansion: 2026 and Beyond | trendforce
- Global Robot Density in Factories Doubled in Seven Years | IFR
- General-Purpose Embodied Robot Market Radar | Omdia
- Google Environmental Report 2026 | Google
- 2026 Environmental Sustainability Report | Microsoft
- 2024 United States Data Center Energy Usage Report | Lawrence Berkeley National Laboratory
- Top 10 AI Data Center Myths Debunked | Gartner

Reported by Ahn Hye-min | Designed by Ahn Jun-seok | Intern: Shin Yeon-sung