On October 13, at the Advanced Thermal Management Laboratory in the Department of Mechanical Engineering at Ajou University in Suwon, Gyeonggi Province, researchers were studying thermal management technologies required to operate AI data centers. When the power output of the machinery was increased, tiny bubbles began to rise in a line on the surface of a metal plate. Below the plate lay a "mock chip" designed to mimic a semiconductor chip. As the chip heated up, the coolant boiled vigorously, absorbing heat and turning into gas.
“When liquid turns into gas, ‘latent heat’ is generated. It’s a principle of extracting hidden heat all at once. Through this process, overheated AI chips can be rapidly cooled,” explained Lee Jeong-ho, a professor in the Department of Mechanical Engineering at Ajou University. This is a glimpse into the development of so-called “two-phase cooling” technology, considered next-generation cooling.
About 300 meters away, in another laboratory, researchers were studying “immersion cooling,” which submerges entire servers in liquid to cool them. Lee noted, “This is where the world’s first single-phase immersion cooling device without a chiller or pump is being researched.”
As AI competition intensifies, cooling technologies to manage heat from AI data centers and cool overheating AI chips are becoming increasingly critical. The government has announced plans to build AI data centers with a capacity of 8.4 gigawatts (GW) by 2029. However, experts argue that without research supporting not only how to secure electricity but also how to distribute it within data centers and cool AI chipsets, achieving this goal will be challenging.
AI data centers cannot operate without cooling technology
AI chips are getting hotter every year. In 2020, the maximum heat output (TDP) of NVIDIA’s A100 (Ampere) GPU was 400 W. In 2022, the H100 (Hopper) reached 700 W, the 2024 B200 (Blackwell) hit 1,000 W, and last year’s B300 (Blackwell Ultra) rose to 1,400 W. The “Rubin,” expected in the second half of this year, is estimated to reach 1,800 W, while the “Feynman” after 2028 is projected to hit 2,800 W. These levels exceed the heat output of typical electric heaters.
Lee stated, “The latest AI chips consume so much power and generate so much heat that operating a single AI data center with 1,000 one-megawatt (MW) racks would require the output of an entire nuclear power plant.”
The massive power consumption of AI data centers means they generate enormous amounts of heat. Without efficient thermal management technology to cool overheated AI chips, data centers cannot function properly for even a second.
The problem is that over 98% of domestic data centers still rely on air cooling (convective cooling), which blows cold air to dissipate heat. While this method works for older servers with lower heat output, it is unsuitable for cooling the latest AI chips. Lee explained, “With the advent of NVIDIA’s Blackwell chips, the paradigm shifted from air cooling to liquid cooling, where cold water flows to cool AI chips. Recently, even this has become insufficient, and there is an urgent need to adopt phase-change cooling, where liquid turns into gas to dissipate heat.”
Source technology in the U.S., supply chain in Taiwan: “Korea must take the lead now”
Currently, the U.S. holds the source technology for AI data center hardware, while Taiwan dominates the component supply chain. The U.S. government has already invested tens of millions of dollars in next-generation cooling research to dominate the hardware market. Taiwanese companies like AVC are estimated to supply 40–50% of the cooling plates (cold plates) for NVIDIA’s latest AI servers. Approximately 90% of the world’s AI servers are also assembled by Taiwanese companies. In contrast, South Korea imports nearly all its servers and cooling equipment. According to the Ministry of Science and ICT, the localization rate of domestic data center servers was only 3.3% as of 2024.
It is still not too late to catch up. While the U.S. and Taiwan dominate the “single-phase cooling” market, including air and liquid cooling, domestic conglomerates could still enter the market with slight technological differentiation. The key area of competition is “two-phase cooling.” Market research firm Deloitte predicts that two-phase cooling will begin to be widely adopted from 2026–2027. U.S. cooling startup Axeleus has already released a product that cools server equipment generating heat equivalent to 100 electric heaters using phase-change technology, marking the early stages of this market.
The government is also viewing cooling as an export industry. The Ministry of Trade, Industry and Resources forecasts that the data center cooling market will more than double to 17.2 billion dollars by 2030, with liquid cooling growing by 15% annually. Lee emphasized, “Since the two-phase market is still in its infancy, if academia and thermal management companies like LG and Samsung focus their efforts, South Korea could lead the AI data center industry after 2030.”
Original article: Chosun Ilbo
English brief by Korea Wind Intelligence, koreawindintel.com
