South Korea’s 80MW Samsong Data Center highlights a broader shift in AI infrastructure as next-generation GPU racks push conventional air cooling toward its limits
As artificial intelligence infrastructure becomes denser and more power-hungry, the race to build AI factories is increasingly moving beyond access to GPUs. Power delivery and cooling are emerging as equally important constraints.
LG CNS is positioning itself for that shift by deploying direct-to-chip (DTC) liquid cooling at the 80-megawatt Samsong Data Center in Goyang, South Korea, in partnership with Naver Cloud.
The infrastructure is being designed to support NVIDIA’s next-generation Vera Rubin platform, with LG CNS integrating liquid cooling, GPU operating environments, power systems and the infrastructure required to run large-scale AI services.
The project provides a glimpse into how the physical architecture of AI data centers is changing as rack power densities rise sharply.
Air cooling has traditionally been sufficient for many conventional data centers. LG CNS says existing air-cooling systems typically handle around 20 to 30 kilowatts per rack. High-performance GPU server racks, however, are moving beyond 200kW.
At those densities, cooling is no longer simply a facilities issue. It becomes a fundamental part of computing infrastructure.
Why 200kW changes the data center equation
Direct-to-chip cooling addresses the problem by placing coolant-carrying cold plates directly on heat-generating components such as GPUs. Heat is transferred into the liquid rather than relying primarily on chilled air circulating through server rooms.
The approach allows substantially greater amounts of computing equipment to operate within a rack while managing the heat generated by increasingly powerful processors.
That becomes particularly important as NVIDIA moves toward rack-scale AI systems.
NVIDIA describes its Vera infrastructure as designed for high-density AI factories and has developed liquid-cooled rack systems to operate alongside Vera Rubin NVL72. The company’s broader Vera Rubin architecture reflects an industry transition away from treating individual GPUs as isolated components and toward designing compute, networking, power and cooling as an integrated system.
For data center operators, that shift changes the infrastructure requirements considerably.
Facilities originally designed around relatively modest rack densities must now accommodate AI systems consuming several times more power within the same physical footprint.
Samsong becomes a test case for Korea’s AI factory ambitions
The Samsong Data Center could become an important example of that transition in South Korea.
The facility has a total incoming power capacity of 80MW and is being developed as a large-scale AI computing hub. LG CNS already operates the data center under an outsourcing arrangement and has a colocation agreement with Naver Cloud.
Under the new project, Naver Cloud will operate NVIDIA Vera Rubin GPU resources owned by South Korea’s National IT Industry Promotion Agency, an organization affiliated with the Ministry of Science and ICT.
The infrastructure is intended to support AI services for Korean public institutions and enterprises.
LG CNS and Naver Cloud plan to complete construction and verification of the liquid-cooling infrastructure by 2027.
The significance of the project extends beyond the installation of a new cooling system.
LG CNS plans to integrate DTC cooling with GPU server operations, power supply infrastructure and the operating foundation for AI platforms. That effectively moves the company further into the physical infrastructure layer required to operate AI at scale.
The AI infrastructure battle is moving below the GPU
The development illustrates a broader change taking place across the global AI industry.
During the first phase of the generative AI boom, much of the infrastructure competition centered on securing advanced GPUs. But owning accelerators is only part of the equation.
Thousands of high-performance GPUs must be supplied with enormous amounts of electricity, interconnected through high-speed networks and kept within safe operating temperatures around the clock.
As rack density rises, these supporting systems increasingly determine how much AI compute a data center can actually deploy.
NVIDIA itself is moving toward this integrated model. Its latest AI factory architecture combines rack-scale computing with high-speed networking and liquid-cooled infrastructure rather than treating cooling as an external facility component.
That transition creates opportunities for companies whose capabilities extend beyond conventional data center construction.
LG CNS has built its data center business around design, build and operate, or DBO, capabilities. Adding liquid cooling allows it to combine those capabilities with increasingly specialized AI infrastructure.
The result is a potential expansion of its role — from building and operating data centers to engineering the environments required for next-generation AI computing.
Cooling could become a strategic AI infrastructure market
The economics surrounding the technology are also becoming increasingly significant.
According to MarketsandMarkets estimates cited by LG CNS, the global data center liquid-cooling market is expected to expand from $4.07 billion in 2026 to $27.65 billion by 2033.
That growth is being driven by a straightforward physical constraint: processors can continue becoming more powerful, but the heat they generate still has to be removed.
For South Korea, the Samsong project also demonstrates that competition in AI infrastructure is broadening beyond semiconductor manufacturing and AI models.
The country already occupies an important position in the global semiconductor supply chain through companies such as Samsung Electronics and SK hynix. Building the infrastructure capable of operating increasingly dense AI systems adds another layer to that ecosystem.
LG CNS’s liquid-cooling project therefore represents more than a cooling upgrade at one data center.
It points toward the next phase of AI infrastructure competition, where the winners may be determined not only by who can secure the most advanced GPUs, but by who can supply enough power, remove enough heat and operate those systems efficiently at scale.
