AI drives up water consumption at Microsoft data centers

"In 2020, Microsoft President Brad Smith announced that the company would reduce its water consumption and even become "water positive" by 2030. This means that Microsoft would replenish more water than it consumes. At the time, this goal seemed achievable, but the rapid rise of generative AI has completely changed the playing field. The construction of new data centers has accelerated, and with it the need for water for cooling."

"According to internal estimates made last year, Microsoft's global water consumption could reach 28 billion liters per year by 2030, more than three times as much as in 2020. After questions from journalists, the company revised those figures. Thanks to more efficient designs and new cooling techniques, Microsoft now expects to use around 18 billion liters. That is still an increase of about 150 percent in ten years."

"It is striking that a large part of the water consumption takes place in areas that already suffer from structural drought or water shortages. In the Jakarta region, where land subsidence and depleted groundwater aquifers are a major problem, Microsoft initially predicted a fourfold increase in water consumption. That estimate was later lowered, without the company explaining exactly why. In the Phoenix area, which has been experiencing drought for decades, Microsoft remains one of the largest industrial water users."

Microsoft's water consumption is projected to rise sharply by 2030 due to accelerated data center construction driven by generative AI. Internal estimates initially forecast 28 billion liters per year by 2030, later revised to about 18 billion liters after design and cooling improvements, still roughly a 150 percent increase from 2020. Recent tens-of-billions investments in new capacity were excluded from those estimates. Much of the increased water use is concentrated in regions facing structural drought and shortages, including Jakarta and Phoenix, where Microsoft remains a major industrial water user despite operational adjustments.

Read at Techzine Global