South Korea's 8.4GW AI Data Center Push Exposes Critical Gaps in Power Delivery and Cooling Design

South Korea's ambitious plan to build 8.4 gigawatts of AI data center capacity by 2029 has thrust power procurement and internal infrastructure design into the spotlight. While securing external electricity supply is a major hurdle, experts warn that the real bottleneck may lie deeper—inside the data center itself, where power distribution and cooling systems must be re-engineered to support next-generation AI workloads.

In June, the government unveiled its "Three Mega Projects for South Korea's Great Leap Forward," committing to a first phase of 8.4GW of AI data center capacity, including 5GW from SK, 2.4GW from GS, and 1GW from Naver. The long-term blueprint aims to expand total capacity to 18.4GW, with government support for power and water supply, site diversification, and the demonstration of domestically developed power and cooling technologies.

What distinguishes AI data centers from conventional facilities is the dramatic escalation in power draw per server rack—a bundle of servers housing GPUs, CPUs, memory, and storage. Under existing low-voltage architectures, rising rack power translates into higher current, greater heat generation, and increased power losses. Observers argue that both the external grid connecting power plants to the data center and the internal power delivery path from the facility to the GPU—often described as the "electric highway"—must be validated together for the entire system to function reliably.

Lee Jeong-ho, a professor of mechanical engineering at Ajou University, highlighted the scale of the challenge: "These days Nvidia doesn't sell GPU chips one by one—it sells them as racks. Current AI server racks draw roughly 150kW, the next generation will reach 600kW, and in the long term a single rack could scale to 1MW."

The 48V power architecture widely used in server racks today offers familiarity and safety advantages due to its low voltage. However, transmitting the same amount of power at lower voltage requires higher current. Supplying 100kW to a single rack under a 48V design demands more than 2,000 amps. As rack power climbs toward 600kW and 1MW, the current requirements compound wiring, thermal, and power-loss burdens significantly.

A paper published last month on the preprint server arXiv raised the same concerns. An international research team, including Lee Sang-hwi, a professor at Korea University, analyzed how the growth of AI workloads is increasing data center power demand, instantaneous current swings, and thermal stress. The researchers concluded that existing 48V server rack architectures and low-voltage AC distribution could soon hit their physical limits.

The team proposed alternatives such as raising intra-rack voltage to 400V or 800V, shifting internal distribution to a DC-centric model, and connecting the grid to the data center via medium-voltage solid-state transformers. Solid-state transformers use advanced power electronics to finely control voltage conversion and power flow. Higher voltage enables the same power to be delivered at lower current, while DC distribution reduces conversion stages and cuts losses.

These design shifts, however, remain in early stages and require further research and demonstration. Data centers must operate around the clock without interruption, making high-voltage DC deployment particularly challenging. Engineers must validate fault interruption points, electric shock and fire prevention, grounding strategies, and long-term operational stability. The researchers identified protection and fault response, grounding, standardization, long-term reliability, co-design of thermal and electrical systems, and digital twin-based verification as the key bottlenecks for next-generation power architectures.

Cooling technology must evolve in tandem with power delivery, as a substantial share of GPU power consumption is converted into heat. Lee Jeong-ho explained, "GPUs generally need to operate below roughly 85 degrees Celsius to deliver full performance, and as power consumption rises, so does heat. It's the same principle as a phone getting hot after prolonged use."

Conventional data centers have relied primarily on air cooling, blowing cold air into server rooms to extract heat. But as AI server power draw escalates, the industry is shifting from air to liquid cooling. Liquid cooling circulates water or dielectric coolant close to the chip, transferring heat far more efficiently than air. As next-generation AI chips consume even more power, cooling devices must move closer to the chip—and ultimately inside the chip or semiconductor package.

Kim Sung-jin, a professor of mechanical engineering at KAIST, noted that the technology is still maturing: "Liquid cooling is already in use, but companies like Intel and HP are developing technologies to embed cooling devices inside the chip itself. The technology isn't mature yet, so it will take time. Integrating multiple chips and cooling devices at scale is not easy."

Whether South Korea can actually validate new power delivery and cooling architectures remains uncertain. An AI data center is not a facility where server racks, power supplies, and cooling units can be swapped out independently. Power delivery and cooling hardware are attached according to the rack architecture defined by the server manufacturer.

Lee Jeong-ho pointed to a structural gap: "South Korea says it's building AI data centers, but the core is missing. We can produce the component technologies needed for cooling, but there is no company in South Korea that builds complete systems or server and rack technology." He added, "Even if we actually set up an AI data center, it is nearly impossible right now to configure it the way we want."

The external grid also poses significant challenges. Researchers at the University of Calgary in Canada have analyzed how the rapid proliferation of AI data centers amplifies not just total power demand but also issues of location, timing, and power variability. If data center loads concentrate in specific regions, they could outpace the expansion of clean energy, while grid flexibility, reliability, and carbon emissions concerns grow in tandem.

Lee Yu-soo, a professor in the Department of Energy Policy and Technology Convergence at Soongsil University, cautioned against over-reliance on renewable energy from South Korea's Honam region for AI data centers. "AI training workloads require large-scale GPUs running simultaneously, so power must be supplied stably 24 hours a day," he said. "It won't be easy unless renewable energy is combined with sources like nuclear or liquefied natural gas (LNG)."

Transmission grid construction speed is another wild card. Lee Jeong-ho noted, "South Korea has many long-distance transmission arrangements from power plants, so the pace of transmission grid construction is critical. Transmission projects come with community acceptance issues, costs, and schedule delays." He added, "It's uncertain whether data center commissioning and power supply timelines will align. The whole picture needs to be re-examined."