Even with the desire to build AI data centers (AIDC) using the latest graphics processing units (GPUs) like Nvidia’s Blackwell, overheating remains a critical issue. Experts point out that conventional air-cooling systems struggle to manage the elevated heat levels of Blackwell GPUs.
Immersion cooling—where servers are directly submerged in non-conductive coolant—is being proposed as a solution.
According to industry sources on Tuesday, the Blackwell-based AI accelerator GB200 boasts five times the inference performance and eighteen times the data processing capability compared to its predecessor, the Hopper-based H100.
However, its power consumption is four times higher, reaching 2700W, making heat dissipation a major challenge for large-scale operations. This has recently led tech giants like Microsoft (MS), Google, and Meta to cancel orders for GB200-based racks.
Immersion cooling offers higher thermal conductivity and reduces power consumption by 30% compared to traditional air-cooling methods, making it a promising solution. It also minimizes issues related to noise and dust.
In South Korea, oil companies and cloud service providers are already making proactive investments in this technology.
SK Innovation’s lubricant subsidiary, SK Enmove, invested in the U.S.-based immersion cooling company GRC in 2022 and completed a demonstration with SK Telecom in a data center.
KT Cloud has also confirmed positive results through a technology proof of concept (PoC), reporting over a 58% reduction in utility power consumption in server rooms, more than a 15% decrease in overall power usage, a reduction of over 70% in server room area, and extended server lifespan due to improved heat exchange.
Samsung SDS is preparing for future customer demand by establishing dedicated facilities at its AIDC in Dongtan, Gyeonggi Province.
However, the technology has not yet proven its reliability. Since servers are directly exposed to the cooling fluid, measures for waterproofing and corrosion resistance of the circuit boards are essential. Additionally, it is crucial to prevent the generation and accumulation of contaminants, as well as the evaporation and leakage of the coolant.
KT Cloud has indicated in a technical report that the design must also consider monitoring systems for coolant temperature, flow, and pressure, as well as ease of replacing consumables and scalability to accommodate increasing server capacity.
Large-scale technology demonstrations have not yet occurred in South Korea. The current demand is insufficient as the latest GPU supply mainly consists of Hopper series models like the H100. At this level, immersion cooling is not essential.
An industry insider stated that they have invested in immersion cooling facilities and left it as an option for clients, but there is still no clear demand. Another source added that few companies in South Korea have tested liquid cooling to the extent that it significantly affects the durability of AI server equipment.
Nevertheless, due to the shortening lifecycle of GPUs and concerns about depreciation, many in the industry predict it will be better to shift toward the latest GPUs based on immersion cooling rather than sticking with the aging Hopper series. Nvidia, the supplier, is also reportedly seeking immersion cooling partners to help drive demand.