Cooling AI chips while using less water and energy

Researchers working in Switzerland are developing technologies to cool data centres for artificial intelligence with lower consumption. The work involves the Swiss Federal Institute of Technology in Lausanne and the start-up Corintis.

The scale of the problem is already measurable. According to the International Energy Agency, data centres currently take around 1.5% of world electricity consumption, and that share is expected to approach 3% by 2030. A United Nations study forecasts that the water footprint of these facilities will double over the same period.

Systems tailored to each processor

Corintis has designed cooling systems matched to the thermal profile of each chip. The arrangement can cut the electricity required by the chillers by around 20%.

The company was founded in Lausanne in 2022 and aims to produce one million cooling plates a year. The first significant installations are expected from 2027. Microsoft is the only publicly confirmed partner.

EPFL's «Heating Bits» project takes a different route. It recovers waste heat at high temperatures, between 70 and 75 degrees, for reuse in district heating networks or for reconversion into electricity using Organic Rankine Cycle technology.

The researcher Jonathan Koomey warns that cooling the chips represents only half the equation. Efficiency gains risk being cancelled out by rapidly growing demand linked to artificial intelligence.

Water use also varies enormously from one facility to another, with differences reaching up to ten thousandfold depending on design and operating conditions. That range makes any sector average of little use.

Professor Mario Paolone poses the underlying question, namely whether current rates of growth in artificial intelligence are compatible with the goals of the energy transition.