Oppressive heat forces AI data centers to get smart about temperature control

AI data centers across Korea are scrambling to keep their servers cool as record heat pushes temperatures toward 40 degrees Celsius (104 degrees Fahrenheit).

Naver Cloud, which operates the massive Gak Chuncheon and Gak Sejong AI data centers, is watching server room temperatures especially closely during the heat wave.

The key threshold is 27 degrees.

“We keep indoor temperatures between 18 and 27 degrees Celsius year-round,” an industry source said.

If temperatures inside a data center rise above 30 degrees, GPUs can shut down automatically.

“If a GPU shuts down, services can stop or experience disruptions,” a Naver Cloud representative said. “To prevent that, we operate an automated system that blocks outside air once temperatures exceed an appropriate level and automatically supplies chilled water to control the server room temperature.”

AI data centers generate enormous amounts of heat during large-scale computing. High-performance hardware such as GPUs produces five to six times as much heat as equipment in conventional data centers, which makes adequate cooling essential.

“Fires caused by heat are common at data centers, so maintaining constant temperature and humidity is essential,” said Lee Byung-hoon, a professor of semiconductor engineering at Pohang University of Science and Technology.

Poor heat management has already forced data centers to shut down in several countries. In May, overheating cut power to an Amazon Web Services data center in northern Virginia, disrupting some services. During a heat wave in Britain last month, a problem with the University of Cambridge data center’s cooling system shut down its AI supercomputer, Dawn.

The most common cooling method draws in outside air to cool servers. But when outdoor temperatures are high, air cooling becomes less effective even as power consumption rises.

Liquid cooling, which uses water instead of air to bring down chip temperatures, has recently gained attention. LG Electronics has introduced a direct-to-chip (D2C) system that circulates water directly to chips to lower their temperature. The company became the first in Korea to receive Nvidia certification for high-efficiency AI data center cooling technology last month.

“High-performance Nvidia chips such as Vera Rubin and Blackwell generate large amounts of heat, so they require liquid cooling that can remove heat quickly,” an LG Electronics representative said.

KT Cloud has also applied D2C cooling at its Gasan AI data center in Guro District, western Seoul, which opened last year.

Some companies and researchers are taking the concept further by exploring the idea of data centers underwater or in space. The Chinese company Highlander built the world’s first commercial underwater data center off Hainan, China, in 2023. The California Institute of Technology and other institutions are studying space-based data centers that would release equipment heat as infrared radiation.

The challenge is expected to grow as AI technology advances. More powerful chips can consume dozens of times more electricity, and increasingly frequent extreme weather caused by climate change could make that demand harder to meet.

The World Economic Forum projected that climate disasters could add $81 billion in annual costs for data centers worldwide by 2035, a figure that could rise to $168 billion by 2065.

“If a heat wave and drought occur at the same time, water for cooling facilities could also run short,” said Jeong Su-jong, the director of Seoul National University’s Climate Tech Center. “Operating data centers could then disrupt water and electricity supplies elsewhere. It is now time to consider an energy transition that takes the climate crisis into account.”

BY SUNG JI-WON [kim.hayoon1@joongang.co.kr]