AI datacenters create heat islands around them, paper finds

Researchers at the University of Cambridge have found that AI datacenters create significant heat islands that raise surrounding temperatures by several degrees at distances up to 10 kilometers (over 6 miles), potentially impacting nearby communities. The findings, published in a paper that has yet to be peer-reviewed, reveal that land surface temperatures around datacenters were higher by between 0.3°C and 9.1°C (0.54°F and 16.38°F) after each facility began operations, with an average increase between 1.5°C and 2.4°C (2.7°F and 4.3°F).

The heat effect can still be measured up to 10 km away, although the intensity drops by approximately 30 percent at about 7 km (over 4 miles) from the facility. An average monthly land surface temperature increase of 1°C (1.8°F) can be measured up to 4.5 km (about 3 miles) from a typical AI datacenter, which the researchers say is comparable to urban heat island effects.

This research comes at a critical time as global server farm capacity grows rapidly. The industry is expected to become one of the most power-hungry sectors in the next decade, with power consumption for data processing projected to exceed manufacturing requirements within three to five years. Hyperscale operators have tripled their infrastructure spending over the past three years in response to AI demand, with quarterly capacity additions increasing by 170 percent by the end of last year.

As a result, global electricity use by massive server farms is set to more than double by 2030, with AI identified as the primary driver. The researchers note that AI datacenters are, in the vast majority of cases, reliant on fossil fuels for electricity generation. This means the steep growth in AI training and use for various applications would translate into a significant increase in emissions.

In the United States, gas-fired power plant projects in development nearly tripled last year, and the demand for electricity for AI datacenters is leading to a resurgence in coal-fired power plants. The paper estimates that up to 343 million people worldwide could be affected by data heat island effects, potentially leading to impacts on welfare, healthcare, and energy systems.

However, the authors suggest that technological advances could mitigate these effects. More energy-efficient electronics or computational methods that make AI training more efficient could reduce power consumption and associated heat generation.

Dr. Andrea Marinoni, lead author of the paper titled "The data heat island effect: quantifying the impact of AI datacenters in a warming world," told New Scientist that the results were surprising and could become a significant problem in the future. "The message I would like to convey is to be careful about designing and developing datacenters," he said.

Vlad Galabov, Omdia Senior Research Director for Enterprise Infrastructure, offered a more cautious interpretation. He noted that the study is a single, early analysis that has not yet been independently replicated or vetted through peer review, and its claims should be treated with caution.

Galabov pointed out that the study examines changes in land surface temperature (how hot roofs, tarmac, and ground surfaces get in the sun), not the near-surface air temperature that people actually experience. "The signal they see is therefore best understood as another example of the urban-heat-island effect from new large buildings, paved areas and associated infrastructure, rather than clear evidence that datacenter waste heat is dramatically warming whole neighborhoods," he explained.

He added that "simple physics suggests that even very large datacenters contribute only a small additional heat flux when spread over kilometres, so most of the observed effect is likely driven by land-use and surface-cover changes, not by AI compute itself."

The research highlights the complex environmental challenges posed by the rapid expansion of AI infrastructure, suggesting that the industry's growth may have broader community impacts than previously recognized.