Advances and Challenges in Data Centre Cooling Amid Growing AI Demand
Data centres, critical for modern computing needs and especially booming due to AI developments, require efficient cooling systems to maintain high processing speeds and enable overclocking. Liquid cooling technologies are at the forefront of innovation, with methods such as Iceotope’s approach that uses water in a closed loop to chill an oil-based coolant directly contacting components. This can cool multiple parts simultaneously and reduce cooling energy usage by up to 80%.
A notable example of technology reuse is a US hotel chain planning to use heat from its servers to warm guest rooms, laundry, and a pool, illustrating growing interest in sustainable practices. However, concerns about the environmental impact of data centres persist: energy and large water consumption have prompted over 200 environmental groups in the US to call for a moratorium on new data centre construction.
Cooling system failures have tangible operational impacts, as seen when a November incident at CME Group disrupted trading technology. To mitigate future risks, CME has enhanced its external cooling capacity. Meanwhile, technology firms continue to explore novel cooling concepts: Microsoft’s subsea data centre off Orkney achieved a Power Usage Effectiveness (PUE) of 1.07 while using zero water but was ultimately closed due to economic challenges. Lessons learned include the advantage of having fewer human operators to improve reliability, and Microsoft continues to investigate liquid cooling ideas, including microfluidics.
Safety and environmental concerns also extend to the chemicals used in cooling. Two-phase cooling systems can utilize PFAS-containing refrigerants with inherent safety risks, prompting some vendors to shift toward PFAS-free alternatives. Iceotope’s fluids are fossil fuel derived but do not contain PFAS.
In academia, UCSD researchers recently published a July paper proposing a pore-filled membrane-based technology for passive chip cooling, which operates without active pumping and shows potential for commercialization.
Regarding energy consumption beyond hardware, Sasha Luccioni of Hugging Face highlighted the significant energy demands of AI models, particularly large language and reasoning systems, calling for increased transparency from AI companies about their energy usage.
