"The power used to ask a single question to an artificial intelligence (AI) service is about 10 times that of a regular internet search. People developing AI think about performance, but those in charge of infrastructure are now worrying about electricity."
Lee Il-jun, head of technology business at NHN Cloud (director), said the same at SMARTCLOUD SHOW 2026, the country's largest tech conference held at the Westin Josun Hotel in Jung-gu, Seoul, on the 26th, noting, "AI competition ultimately comes down to a race to secure data centers and power."
In a lecture on "Understanding AI infrastructure and data centers, and what AI changes," Lee said, "In the past, data centers typically ranged from 5 to 8 megawatts (MW), but recently 60 to 100 MW is being exceeded, and even gigawatt (GW)-class data centers are being discussed." A 20 MW data center that NHN Cloud recently began using consumes power comparable to Gongju, South Chungcheong, a city with a population of about 100,000.
The spread of AI is also changing the nature of data centers. While conventional data centers handled web, search, and databases services based on central processing units (CPUs), AI data centers are operated around graphics processing units (GPUs) that perform parallel processing with thousands to tens of thousands of cores. Because they take data as raw material to produce "intelligence," data centers are turning into "AI factories."
The problem is that even if you buy the latest GPUs, there are not enough facilities to run them. Existing data centers were designed to supply 4 to 7 kilowatts (kW) per rack, but AI GPU racks require 40 to 130 kW. Nvidia's next-generation Vera Rubin NVL72 is expected to consume about 210 kW per rack.
Lee explained, "If you put a 210 kW rack into a legacy data center designed at 5 kW per rack, one rack hogs the power that about 40 existing racks would use," adding, "Even if there is space, the lack of power and cooling capacity makes it impossible to install the latest AI equipment."
As power density rises, traditional air-based cooling has also hit its limit. Air cooling is effectively capped at 30 to 40 kW per rack. Beyond that, "direct-to-chip" cooling, which brings coolant into direct contact with the chip, is needed, and immersion cooling—submerging servers in a nonconductive liquid—is likely to spread in the future. AI data centers are, in effect, shifting from computer rooms to mechanical facilities equipped with power, piping, and cooling systems.
Site selection criteria are changing, too. Lee said, "In the past, we preferred places close to the Seoul metropolitan area with many users and low network latency, but now the priority is whether large-scale power can be secured," adding, "Because real-time response speed matters relatively less for AI training, there is ample room to build data centers in non-capital regions." The stronger the power constraints in the metro area, the faster AI data centers may disperse regionally.
Global big tech's moves that resemble power generation businesses are for the same reason. Microsoft (MS) decided to secure power by restarting the Three Mile Island nuclear plant in the United States, and Amazon and Google are investing in Small Modular Reactor (SMR) projects. The competitiveness of the data center industry is shifting from buildings or server counts to the ability to procure stable power—toward an "energy business."
Lee said, "In Northern Virginia, the world's largest concentration of data centers, the vacancy rate is only 0.3%, and there are projections that by 2027 there will be a 40% shortfall in the power data centers require," adding, "We are in an era where you cannot build data centers because there is no electricity to draw, even if you have money."
NHN Cloud built the Gwangju National AI Data Center in 2023 and is operating a large-scale AI cluster, "NHN Factory X," in Yangpyeong-dong, Yeongdeungpo-gu, Seoul. Factory X is equipped with 7,656 Nvidia B200 GPUs, and the related business scale is about 1.01 trillion won.
Lee advised corporations to validate business viability with cloud GPUs before investing massive expense to build their own AI infrastructure. He said, "A single AI server costs 1 billion to more than 1.5 billion won, so you must first check whether you can create value commensurate with that," adding, "It is advisable to use the cloud for about three months to a year and then decide whether to invest on your own."
He added, "Investment to secure AI competitiveness will continue, but not every corporation needs to build a data center directly," noting, "The key going forward will not be owning the most expensive GPUs, but how efficiently you can operate AI infrastructure, including power and cooling expense."