Experts say grid capacity — not GPUs — could become biggest constraint on Korea's AI expansion
South Korea's bid to stay competitive in artificial intelligence is running into a constraint that cannot be solved by chips alone: electricity.
As the US and China pour hundreds of billions of dollars into AI infrastructure, experts say Korea's biggest challenge is no longer securing Nvidia GPUs but ensuring it has enough power, grid capacity and suitable sites to run the data centers those chips require.
US, China raise the stakes
The pressure is mounting as the world's largest economies accelerate investment in AI infrastructure.
In the US, the Stargate Project, backed by OpenAI, SoftBank, Oracle and MGX, was unveiled in January 2025 with a pledge to invest $500 billion over four years in AI infrastructure. The buildout is largely private-sector-led, with government support expected in areas such as permits, land and power.
China is pursuing a more state-led model. Bloomberg News reported that Beijing is planning to spend around 2 trillion yuan ($295 billion) over five years to build a nationwide network of AI data centers. State-owned firms such as China Mobile and China Telecom are expected to operate much of the infrastructure, while domestic suppliers including Huawei provide key AI technologies.
Korea's response is smaller but increasingly ambitious. The Ministry of Science and ICT aims to stimulate 65 trillion won ($42.8 billion) in private AI investment between 2024 and 2027. Separately, the government plans to raise its AI budget to 10 trillion won in 2026 and mobilize up to 30 trillion won in public-private AI investment.
From buying GPUs to powering them A key pillar of Korea's strategy has been securing AI chips. Nvidia last year announced plans to supply more than 260,000 GPUs to the Korean government and major companies, including Samsung Electronics, SK Group, Hyundai Motor Group and Naver Cloud.
But experts say the bottleneck is shifting from buying chips to operating them.
AI data centers consume far more electricity than conventional facilities. Industry estimates show a server equipped with eight Nvidia B200 GPUs can require about 20 kilowatts of power, including cooling. A rack of four such servers can consume roughly 80 kilowatts, compared with the 3 to 5 kilowatts typically required in traditional data centers.
That gap has pushed electricity infrastructure to the center of Korea's AI strategy.
Park Jong-bae, a professor of electrical and electronics engineering at Konkuk University, warned that further concentration of AI data centers in the Seoul metropolitan area could strain the country's power system.
"If AIDCs continue to cluster in areas like the capital region, where electricity demand is already high, it could lead to grid overload, transmission bottlenecks and power supply stability issues," Park said.
More than 70 percent of Korea's data centers are already concentrated in the capital region, which Park identified as a growing vulnerability. He argued that new AI data centers should be built outside the Seoul area and that rules governing direct power purchases should be eased.
The government has introduced measures including the Distributed Energy Act and grid impact assessments to encourage AI data center development outside the capital region. Industry officials, however, say permitting support alone will not be enough unless power supply, cooling capacity, land availability and environmental concerns are addressed together.
A different path for Korea?
Korea's challenge extends beyond electricity.
While Samsung Electronics and SK hynix remain global leaders in memory chips, the country still relies heavily on Nvidia for GPUs and much of the broader AI accelerator ecosystem. That dependence has raised questions about whether Korea should try to compete directly with the US and China in hyperscale AI infrastructure.
For Korea, China's state-led buildout highlights a strategic dilemma: whether to join the race toward ever-larger GPU-centered campuses or pursue a more distributed model built around its industrial strengths.
Yoo Hoi-jun, a professor at KAIST's School of Electrical Engineering, said Korea should avoid competing on the same terms as the world's two largest economies.
"The battle centered on GPUs and hyperscale data centers is already an area where the US and China have secured a clear advantage," Yoo said. "Rather than competing in the same way, Korea needs to find niche markets where it can generate real economic value."
Yoo said Korea should build on its strengths in memory, processing-in-memory technologies and domestic AI accelerators while looking beyond hyperscale facilities.
"For large-scale data center infrastructure, the framework is already largely set, so it will not be easy for new technologies to be adopted quickly," he said. "On-premises systems, small data centers and edge servers may offer more room for Korea to compete."
Such an approach would align Korea's AI infrastructure strategy more closely with its strengths in semiconductors, manufacturing, mobility and cloud services rather than forcing the country into a scale-driven contest with the US and China.
yeeun@heraldcorp.com