'We cool 7,656 GPUs with water': Inside NHN Cloud's Seoul AI data center

"We're cooling most of the heat generated by the GPUs and CPUs with water," said Lee Il-jun, head of NHN Cloud's technology business division.

NHN Cloud is pushing into large-scale AI infrastructure with NHN Factory X Seoul, an AI data center in Yeongdeungpo-gu equipped with liquid cooling technology. As AI adoption has made heat management in high-performance GPUs a central challenge, the company is betting its proven operational capabilities will help it win market share.

On Thursday, this reporter visited NHN Factory X Seoul in Yeongdeungpo-gu. After clearing security and stepping into the sixth-floor data hall, rows of black server racks — each towering well above head height — stretched from row A through row L.

A steady hum from cooling fans on hundreds of servers filled the air, though the noise was mild enough for easy conversation. That is because the facility uses direct liquid cooling to draw heat away from GPUs and CPUs with water rather than fans. NHN Factory X Seoul is the first facility to build an entire large-scale cluster of the latest Nvidia B200 GPUs using liquid cooling throughout.

"A standard air-cooled GPU server can exceed 110 decibels — the fans spin so hard that conversation is sometimes impossible," Lee said. "Here, because we remove most of the heat from the CPUs and GPUs with water, you can hold a normal conversation even in a densely packed server environment."

Lee explained the mechanics: "We press a metal cold plate directly onto each CPU and GPU chip and run coolant through it to absorb heat at the source — that is the direct liquid cooling, or DLC, method. The water handles the heat from the CPUs and GPUs directly, while the remaining heat from memory chips, network cards and other components is still dissipated by air."

The liquid cooling keeps GPU temperatures stable. In a typical air-cooled environment, GPU temperatures can climb to around 50–70 degrees Celsius, but at this facility they are held to roughly 48–50 degrees, allowing the chips to run at full capacity. "Once a GPU exceeds about 70 degrees, it automatically throttles its own compute load to prevent damage, which degrades performance," Lee said. "Maintaining a consistently low temperature is the key to getting the most out of a GPU."

Power consumption also falls. "A single Nvidia B200 GPU server draws about 14.3 kilowatts in an air-cooled environment, but with liquid cooling that drops to about 12.3 kilowatts," Lee added.

NHN Cloud stressed that all B200 GPUs installed at the center are state-owned assets. The company operates services for major public institutions, drawing on its experience building and running large-scale GPU and cloud infrastructure. It is leading the cloud migration of Government 24, the government's integrated online service portal, with about 70 percent of the platform already running on NHN Cloud and the remainder being migrated in stages.

"The government does not separately pay for data center construction costs, electricity bills or staffing expenses over the next five years," Lee said. "Instead, NHN Cloud is allowed to use 20 percent of the total state-invested assets for its own research and development or sell capacity to private customers, and we cover the data center's operating costs from that revenue."

NHN Cloud plans to keep refining its operational capabilities and expand its AI infrastructure business into the private corporate market. Shinhan Financial Group has already set up a private cloud region inside the data center. More recently, the company built a dedicated cluster of 1,000 Nvidia B300 GPUs for Krafton at its Pangyo data center.

"In future data centers, power supply and cooling capacity will likely matter more than floor space," Lee said. "We will position ourselves as a trusted cloud provider not only for public institutions but for corporate customers as well."

NHN Factory X Seoul was built through the Ministry of Science and ICT and the National IT Industry Promotion Agency's AI Infrastructure Reinforcement project. The facility began full commercial service in April and currently houses 7,656 Nvidia B200 GPUs with a combined compute capacity of approximately 574 petaflops — about 5.7 times the scale of the Gwangju National AI Data Center.