‘Thirsty’ ChatGPT uses four times more water than previously thought

Shaolei Ren, professor of electrical and computer engineering at Riverside, said that the original water footprint calculation was based on a figure from OpenAI, which developed ChatGPT, in 2020. New figures released in September in a paper by Microsoft showed: “The energy consumption [of GPT-3] will be at least four times as much as the number that we used. This also means the water footprint should be increased four times.”

Data centres have always needed cooling because computers are essentially like radiators, with most of their energy converted into heat. However, the servers that are built for AI are processing huge amounts of data and have more power density and greater cooling demands.

The latest AI server racks unveiled by Nvidia are generating 120kW of power. “It’s an awful lot of heat. It’s human-meltingly hot,” said David Craig, chief executive of Iceotope, a British company that helps to cool data centres.

He said that 1kW was the equivalent of an old three-bar heater. “It’s 100 to 120 of your granny’s fire. In a meter square. It’s two meters tall, one meter square. It’s that hot. And there are thousands of them [in a data centre],” Craig added.

Many data centres use water-based systems to cool the plant with towers evaporating the heat, like a huge perspiration system, which means that the water is lost. It also has to be drinking quality because impurities can damage the servers.

The United States is the world’s biggest location for data centres, with more than 5,000, compared with about 600 in the UK.

New data centres built in Britain over the next five years could require the same amount of water as a city the size of Liverpool, according to estimates by the water industry.

The latest Big Tech sustainability reports show double-digit increases in water consumption by Google (17 per cent), Microsoft (22.5 per cent) and Meta (17 per cent).

“The expansion of AI products and services is leading to an increase in data centre workloads and the associated water footprint required to cool them efficiently,” Google said.

Amazon has not release its total water consumption figures, believing it to be the wrong metric. “Amazon — I think it probably consumes more water than Microsoft,” said Ren, adding that the company was “very secretive” compared with other tech firms.

All of these companies have schemes to put water back into nature using projects that help river flow, capture rainwater, recharge aquifers and modify dams. They have all pledged to become “water positive” by 2030: returning more than they consume.

But where are they putting back the water? Quite often it’s not in the same place it was taken out, which can be in areas of “water stress”. Microsoft said that 41 per cent of the 7,844ML it consumed — the equivalent of 3,100 Olympic swimming pools — were in areas of water stress.

As American tech companies have started to expand data centres internationally, there has been a backlash. Santiago in Chile has become the focus of a clash between the government’s desire to use the centres to drive economic growth and locals’ concerns about the environmental impact.

Over the past 12 years, 16 data centres have been approved in the city, making it one of the biggest hubs in Latin America. However, with the country experiencing a drought that is expected to last until 2040, activists have mobilised against the plans.

In September, they forced Google to scrap its design for a $200 million (£152 million) data centre in the capital and start again, this time using an air-based cooling system. The success has emboldened their plans to oppose Microsoft and Amazon centres.

Tania Rodríguez, a leading Chilean activist, was named by Time magazine as one of the top 100 most influential people in AI. The industry’s appetite to build shows no sign of abating with the current AI gold rush.

BlackRock, Microsoft and a United Arab Emirates-backed investor announced last week a new AI fund that could deploy up to $100 billion to build data centres in the US.

Google and Oracle have announced that they are building centres that consume 1GW of power, similar to a nuclear or coal-fired power plant. In Ireland, they consumed 21 per cent of the country’s electricity last year, up from 5 per cent in 2015.

Craig said: “One of the horror stories of our planet is the sheer utter wastefulness of our power generation processes. We are 60 per cent wasteful in power generation. Every watt of energy we get has consumed three watts of resources to get it out. That’s where a vast amount of water consumption comes from in the first place. Huge amounts of water are lost in that process.”

Water UK has calculated that new data centres built in Britain in the next five years could need the same amount of water as 500,000 people. The sector is concerned that many of the new facilities will be clustered in the water-stressed southeast of England, which is drier than Istanbul and Dallas.

There is no requirement for data centres to report how much water they consume. However, the Environment Agency is doing a survey of the industry. There are ways that centre operators can curb their water use, such as using an air-based cooling system, as Google plans for a new facility in Waltham Cross, Hertfordshire.

Ren urges AI companies to optimise their models and algorithms and distribute workloads to locations with better water efficiency. Google DeepMind has developed an AI system that helped the company reduce its energy consumption for cooling centres by 40 per cent, in part by routing traffic more efficiently.

Data centres are also being built in northern Europe’s cooler countries but it’s not practical to do this at scale with AI because of the demands for speed of response.

One company, Digital Realty, is already ensuring 43 per cent of the water it uses comes from non-drinking sources, such as rainwater harvesting.

It operates about 300 centres globally that consume the same amount of water as about 17 golf courses each year. The firm prefers air-cooled technology to cut water use, as well as “closed loop” systems, such as one that draws water at London Docklands but circulates the water back into the dock waters after it’s used for cooling.

Aaron Binkley, vice-president of sustainability at Digital Realty, said: “Energy usage is the primary metric most people associate with data centre environmental impacts, but behind that is water. Water is a critical component.”

Iceotope uses “precision liquid cooling” to capture nearly 100 per cent of the heat generated by servers. It runs a special electronic-friendly liquid through the hottest part of the stack, which is then piped out and air-cooled, which means no water is needed.

Craig, who retires as chief executive next week, said that Iceotope works with two of the big technology companies and that, despite their voracious appetite for water, they are more efficient than smaller companies.

“The vast majority of data centres are completely consumptive. The big boys, the ‘hyperscalers’, are doing the best job of trying to be efficient and are probably, today, 60 per cent of the market. That means 40 per cent [are not]. And it’s crap.”

While energy-intensive AI algorithms have prompted concerns about growing energy and water use, Binkley said AI also offered the promise of using less water. His company is using AI tools on at least five of its data centres to find points in buildings where it can use less water. “It’s about flipping that narrative about AI on the head,” he said.

An OpenAI spokesman said: “AI can be energy-intensive and that’s why we are constantly working to improve efficiency. We carefully consider the best use of our computing power and support our partners’ efforts to achieve their sustainability goals.”