As the government said it will build 8.4 gigawatts (GW) of artificial intelligence (AI) data centers by 2029, securing power has emerged as a key task. Experts see not only where and how to draw electricity as an issue, but also how to distribute and cool the electricity inside the data center as a bottleneck.
In June, the government announced the three major "Korea great leap" megaprojects and said it would build AI data centers totaling 8.4 GW in the first phase, including 5 GW by SK, 2.4 GW by GS, and 1 GW by Naver. In the long term, it plans to expand the total to 18.4 GW and support power and water supply, site dispersion, and demonstrations of domestic power and cooling solutions.
What sets AI data centers apart from conventional ones is that the power used by a single server rack increases sharply. A server rack is a bundle of servers that includes a graphics processing unit (GPU), central processing unit (CPU), memory, and storage. In the existing low-voltage power architecture, as rack power increases, so do current, heat, and power loss burdens. That is why some say the entire "electric path" needs to be vetted, from the external grid that runs from the power plant to the data center to the internal power delivery structure that runs from inside the data center to the GPU.
◇ 1 MW per rack power… Limits to the existing "48V electric path"
Lee Jeong-ho, a professor in the department of mechanical engineering at Ajou University, said, "These days Nvidia does not sell GPU chips one by one but sells by the rack," and added, "Currently, AI server racks use about 150 kW, the next generation will be 600 kW, and in the long term a single rack could grow to the 1 MW class."
The widely used 48V power architecture in today's server racks has the advantage of being familiar in design and highly safe due to its low voltage. However, at the same power, a lower voltage means higher current. To supply 100 kW to a rack, a current of more than 2,000 A is required in a 48V architecture. As rack power grows to 600 kW and 1 MW, even larger currents are needed, increasing the burden of wiring, heat generation, and power loss.
A paper posted last month on the online preprint server arXiv raised the same issue. An international research team including Professor Lee Sang-hwi of Korea University analyzed that as AI workloads increase, data center power demand, transient current changes, and heat burdens are growing, and that the existing 48V server rack power architecture and low-voltage AC distribution may hit limits.
As alternatives, the team proposed raising the rack internal voltage to the 400V or 800V class, shifting in-data-center distribution to direct current, and linking the grid and data center with medium-voltage solid-state transformers. Solid-state transformers use more power electronics than conventional transformers to finely control voltage conversion and power flow. Higher voltage allows the same power to be delivered with smaller current, and DC distribution can cut losses by reducing power conversion stages.
However, such architectural shifts still need more research and demonstrations. Because data centers must run around the clock without stopping, adopting high-voltage DC distribution requires verifying where to cut power during a fault, how to prevent electric shock and fire, how to ground, and whether long-term operation is stable. The team also cited protection and fault response, grounding, standardization, long-term reliability, co-design of thermal and electrical, and digital twin–based validation as bottlenecks in next-generation power architectures.
◇ As per-rack power grows, cooling must be redesigned too… Demonstration base is also a task
If the electric path of AI data centers changes, thermal management changes with it, because a significant portion of the power used by GPUs turns into heat. Professor Lee said, "GPUs generally need to operate at around or below about 85 degrees to deliver full performance, but as power consumption rises, heat increases too," adding, "It is the same principle as a smartphone getting hot after long use."
Conventional data centers mainly used air cooling that blows cold air into server rooms to remove heat, but as the power requirements of AI servers rise, the center of gravity in cooling is shifting from air cooling to liquid cooling. Liquid cooling draws heat away by flowing water or an insulating coolant close to the chip, which can transfer heat more efficiently than air.
As next-generation AI chips require even more power, technologies that place cooling devices closer to the chip—and further, inside the chip or within the semiconductor package—are becoming necessary. Kim Seong-jin, a professor in the department of mechanical engineering at KAIST, said, "We are already using liquid cooling, but several companies, including Intel and HP, are developing technologies to embed cooling devices inside chips," adding, "The technology is not yet mature, so it will take time. It is not easy to integrate many chips and cooling devices over a large area."
Whether new power delivery and cooling architectures can be verified in practice in Korea is also a variable. AI data centers are not facilities where server racks, power supplies, and cooling devices can be swapped independently. Power and cooling devices attach according to the rack architecture set by the server manufacturer.
Professor Lee Jeong-ho said, "Korea says it will do AI data centers, but the essence is missing," adding, "We can build the component technologies needed for cooling, but there are no domestic companies that build whole systems or server and rack technologies." He said, "Even if we practically build AI data centers, configuring them the way we want is almost impossible at this point."
◇ External electric paths are also variables… The speed of renewable energy and transmission grids is key
Issues remain with the grid outside AI data centers. A research team at the University of Calgary in Canada analyzed that the rapid increase in AI data centers amplifies not only total electricity demand but also issues of where and when power is used, as well as power variability. They noted that if data center loads concentrate in specific regions, they could outpace the expansion of clean energy, and challenges with grid flexibility and reliability and carbon emissions could grow together.
Lee Yu-su, a professor in the department of energy policy and technology convergence at Soongsil University, also said the idea of using Honam's renewable energy for AI data centers should be viewed with caution. Lee said, "AI training requires large numbers of GPUs to run simultaneously, so power must be supplied reliably 24 hours a day," adding, "Without a composite mix of power sources such as nuclear power or liquefied natural gas (LNG) with renewable energy, it will not be easy."
The pace of transmission grid construction is also a variable. Lee said, "In Korea, there are many cases of sending power long distances from power plants, so the speed of building transmission grids is important," adding, "Transmission projects come with resident acceptance, expense, and schedule delay issues." Lee continued, "It is uncertain whether the timing of data center construction will match the timing of power supply," adding, "There is a need to review everything again."
References
arXiv (2026), DOI: https://doi.org/10.48550/arXiv.2606.25095
arXiv (2026), DOI: https://doi.org/10.48550/arXiv.2606.21064