"The biggest difference between an artificial intelligence (AI) dedicated data center and a general data center is power density. High-performance graphics processing units (GPUs) consume a lot of power and generate significant heat, so we applied a liquid cooling method that cools the heat with cold water."Lee Il-jun, NHN Cloud head of technology business division (executive director)
NHN Factory X Seoul in Yangpyeong-dong, Yeongdeungpo District, Seoul, visited on the 18th, is an artificial intelligence (AI) data center that operates 7,656 of Nvidia's latest B200 graphics processing units (GPUs). As one of the key hubs of the AI infrastructure business that NHN Cloud has put forward as a future growth engine, it provides high-performance computing resources needed by domestic corporations, government agencies and academia.
When entering the sixth-floor data hall of NHN Factory X Seoul that day, six black GPU server racks, which looked like cabinets for storing items, lined up on both sides came into view. The rows of server racks filled dozens of aisles at regular intervals. Each rack housed six B200 GPU servers and a cooling distribution unit (CDU), and thick liquid-cooling pipes above the racks supplied cold coolant and returned heated water. A "whirr" like dozens of fans and small motors spinning resonated through the space.
Liquid cooling dissipates heat through cold plates that touch GPUs and Neural Processing Units (NPUs), which account for 90% of server heat. While traditional air cooling lowers equipment heat by using airflow, liquid cooling uses a liquid with high thermal transfer efficiency to deliver the stable cooling performance required for AI data centers with high GPU density.
Although the data center where NHN Factory X Seoul is housed was designed for air cooling, NHN Cloud separately built direct liquid cooling (DLC) facilities to operate large-scale B200 GPUs. Lee Il-jun, NHN Cloud head of technology business division (executive director), said, "Cold water moves through separate piping to the server racks to cool the equipment, and the heated water exits in a refrigerant circulation system."
NHN Cloud pointed to the fact that it is the first in Korea to apply liquid cooling to large-scale GPU resources as a differentiator that day. Each server rack built here holds 48 B200 GPUs, and rack power density is 75 kW, about twice that of a typical data center. As the power density of high-performance GPU servers increases, heat generation is also intensifying, raising the importance of cooling technology. The company said liquid cooling is essential to efficiently dissipate the heat generated.
Executive director Lee said, "Because the next-generation GPUs to be introduced by Nvidia will have increasingly higher power density per rack, AI data centers to be built going forward will have no choice but to use liquid cooling," adding, "Even Blackwell GPUs alone require around 230 kW per rack, so air cooling tuned for 30–40 kW per rack cannot control the heat."
NHN Factory X Seoul is AI infrastructure created with a government budget of 1 trillion won and is a state-owned asset. NHN Cloud will be responsible for building and operating the infrastructure for the next five years. To that end, it leased the third through sixth floors—half of a private, eight-story data center—in Aug. last year and began operations in Apr. this year.
Of the space occupied by NHN Factory X Seoul, the third through fifth floors operate government AI computing resources, and the sixth floor is used for NHN Cloud's own AI cloud service operations. Of the state-owned asset B200 GPU infrastructure, 80% is provided to corporations, universities and research institutes designated by the government, and the remaining 20% is used by NHN Cloud.
Lee said, "It costs about 50 million won to rent one server for a month, which is an expense that most institutions and corporations find hard to bear," adding, "Industries designated by the government can use AI computing resources at about one-tenth of the market price, and schools and research institutes are receiving them free of charge." He added, "Because AI infrastructure is so expensive, the government project focused on invigorating the AI ecosystem by providing the AI computing resources needed across domestic industry at lower cost."
The total IT load of NHN Factory X Seoul is 13 megawatts (MW), exceeding the power used by a small to mid-sized city of 100,000 people, the company said. External power is supplied with a dual-feed system to ensure supply stability, and the facility is designed to maintain service continuity in the event of a failure through UPS and emergency generators.
NHN Cloud, NHN's cloud subsidiary, operates data centers in five locations—Seoul, Pangyo, Pyeongchon, Gwangju and Daegu. AI data centers and GPU infrastructure, now essential backbone facilities for the AI industry, have established themselves as NHN's core growth axis.
With NHN Factory X Seoul in Yangpyeong beginning operations in Apr. this year, NHN posted record second-quarter results. Revenue rose to 757.9 billion won and operating profit to 57.9 billion won, up 25.3% and 164.1%, respectively, from a year earlier.
In particular, as growth in NHN Cloud's AI GPU business gathered momentum, NHN's technology institutional sector revenue rose 52.9% to 159.8 billion won. Looking only at NHN Cloud, second-quarter revenue jumped 85.3% year over year as stable revenue began to flow from the Yangpyeong data center in Seoul, where supply of GPU as a service based on Nvidia B200s has ramped up. NHN Cloud is currently discussing new long-term contracts for the remaining portion of volumes it operates on its own in the Yangpyeong region.