Artificial intelligence is the Gulf’s hottest investment right now in more ways than one: How do you keep a data center cool in one of the world’s hottest regions?
Let’s start with the basics. AI data centers can draw 100 MW of power or more — two to four times what a traditional data center might burn. That produces a lot of heat — heat that needs to be dissipated with air or water.
A decent ballpark figure is that cooling comprises about 20-40% of the energy use and about the same percentage of operating cost of a large data center. Those figures are higher in less efficient data centers and lower in more efficient ones.
Cooling is even more critical in hot regions like the Gulf, where temperatures don’t take a break at night. “In cooler climates, you have what I’d describe as free cooling, especially at night,” Maria Goodpaster (LinkedIn), a partner at the consulting firm McKinsey & Company who advises companies on AI infrastructure, tells EnterpriseAM. “In hot climates like the Gulf, you just don’t have that as much. You have to use energy for most of your cooling.”
From air to water: Traditionally data centers were cooled with air cooling systems, similar to big air conditioners, but those can’t keep up with the Gulf’s cooling needs. There are various methods of water cooling, but the most common in large AI data centers is direct-to-chip cooling. That means a metal plate with channels carved inside of it is placed directly onto the chip, then cold, chemically treated water runs through those channels, keeping the chip cool.
A more experimental method is called immersion cooling. “This is like putting the server components directly into a big tub of special, non-conductive fluid [resembling] mineral oil,” Shaolei Ren (LinkedIn), a professor of electrical and computer engineering at the University of California, Riverside, tells EnterpriseAM. Theoretically, immersion cooling will be more efficient in many cases (if it’s widely adopted), but it also creates problems of its own. For example, imagine trying to repair a malfunctioning server rack right after it emerges from a big tub of goo.
Direct-to-chip cooling is likely to be the industry standard for big AI data centers for at least the next decade, Goodpaster says.
These systems minimize water waste by operating as much as possible on a “closed loop.” That means the same chemically treated water flows across the chips, sucks out the heat, then gets cycled out of the server racks and cooled before it’s cycled back in again.
But: Separate water is also often used in the process of cooling down the chemically treated water before it repeats its journey. And those facilities still consume water for other purposes, such as cleaning and plumbing. Ironically, older, air-cooled systems often waste more water than water-cooled systems because they deliberately evaporate water that’s used to cool the air that goes on to cool the servers.
Best-case scenario: Microsoft has boasted that its new data center design, launched in 2024, “consumes zero water for cooling,” reducing water consumption by 125 mn liters per data center per year, the company said.