Cooling the cloud, conserving Canada’s water at data centres

According to the Canada Energy Regulator, out of over 7,000 data centres worldwide, only 239 are in the country. However, this number is on the rise due to Canada's low energy prices in select regions, its renewable resources, and cooler climate.

Last month, OpenAI CEO Sam Altman tweeted that a ChatGPT query uses 0.000085 gallons of water per query (roughly one-fifteenth of a teaspoon) and eleven times more electricity than a Google search. With the increasing number of data centres storing AI processors and innovative chips, energy use is also rising.

The International Energy Agency's Energy and AI analysis from last month reported that global electricity demand from data centres is set to more than double by 2030. They estimate that energy to be 945 terawatt-hours (TWh), the same total consumption as Japan.

Gregory Lebourg, OVHcloud's Global Environmental Director, reviewed the methodology of this report. He says that much of this energy increase is due to the AI boom that occurred within the last year.

In January 2024, the IEA estimated that the TWh of global data centre energy consumption could reach 1000, but later revised this to 500 by 2030.

Lebourg says the fluctuation of estimated TWh by 2030 is constantly changing because "no one really knows" how far the AI boom will escalate. He says that what can be put into data is the efficiency of a data centre, while innovations will put DCs on track to curb any future increase in energy consumption.

Major cloud players like Apple, AWS, and Microsoft have committed to achieving water-positive status by the next decade to address this increase in energy.

"Racks are becoming more dense in terms of power, we are talking about racks above 40 kilowatts, even 50 kilowatts," says Lebourg. "You simply can't cool them down with air flow. So you can't use the traditional air cooling system that all the data centers have."

According to a report by the World Economic Forum, AI consumes a large amount of energy when training learning models. Adding and processing at increasingly advanced levels require water to cool the systems.

In the report, the WEF states that a one megawatt (MW) data centre can use up to 25.5 million litres of water annually just for cooling. The same amount of water would be used in the daily water consumption of approximately 300,000 people.

"We've seen a lot of these data points come out, and what they're based on is something that we always need to look into," says Alex Setmajer, Program Manager of Water & Nature at Equinix.

She says quantifying data from a consumer consumption perspective doesn't always produce accurate data, as every data centre works with different designs, power usage effectiveness, and cooling systems.

According to Setmajer, closed loops in server rooms prevent evaporation from occurring while cooling computer systems. It is when the server rooms transfer the heat to the building's vent systems that results in evaporative or air cooling to reduce temperatures.

In an effort to cool without depleting water resources, cooling systems are being designed to use non-potable water in tandem with air cooling to offset environmental costs.

At Equinix's Toronto TR1 Data Centres, they utilise a Deep Lake Water Cooling (DLWC) system to reduce water usage by up to 50 per cent. Water is extracted from the depths of Lake Ontario, where the water is consistently four degrees Celsius. 

Toronto-based energy company Enwave, which manages the local project, cools over 100 downtown buildings using its system of intake pipes off Lakeshore Boulevard West. 

The TR2 facility plans to utilise this system in the near future, according to Sanjeevan Srikrishnan, Principal Global Solutions Architect at Equinix.

Srikrishnan says the new GB200 chips are more powerful than the H100, but they are only using 10 per cent of the compute energy to cool, as water cooling solutions continue to innovate.

"As these chips densify they also don't need the same [environmental] footprint that the non AI chips need," says Setmajer. "AI [chips], might be half or less the amount of space, and in older buildings, may not be able to use that actual floor square footage space anymore, because that power and that cooling is all going to that same load just in a smaller footprint."

OVHcloud's data centre outside Montreal is situated along the St. Laurence River, where it generates power from the current of the waterway. Leboug says Canada is a naturally gifted place for data centres, as it is filled with renewable resources at a cost-effective rate compared to other global sites.

Lebourg says OVHcloud's data centres work at 1.26 PUE worldwide (compared to the global standard of 1.56). He says this is manageable through dry coolers that use the natural environment, supplemented with additional water, to maintain a constant temperature. 

"It's fair to say that this PUE is decreasing over time, because the more we grow, the more we deploy new lineups, which are way more efficient. If you look at the number [in 2023] at 1.29, last year 1.26 - I'm super positive that this year again, it's going to be below 1.26."

This month, OVHcloud unveiled its Environmental Impact Tracker, which enables clients to access a comprehensive overview of the carbon emissions linked to their cloud usage. Lebourg says the product gives transparent, reviewed analytics tailored to specific cloud customers. "Its called an Environmental tracker becasue it covers not only carbon, but water as well, and abiotic resources."

As the AI boom continues, estimates will fluctuate, but one thing will stay the same: cooling innovation.

Setmajer says, "There's older technology, there's newer technology, and as we're advancing and then building new facilities we're getting better, and better technology, and that's continued to innovate."