Energy, water use and pollution of AI and data centres rival most countries

David Gerhard, head of computer science at U of M, discusses a new poll that delves into how Canadians feel about AI data centres.

WASHINGTON — The environmental footprint of data centres already rivals some of the world’s largest countries, according to a United Nations University report, which also predicts their water and energy use and pollution will double in just four years as use of artificial intelligence grows.

Last year, global data centres used 448 trillion watt-hours of electricity, more than all but 10 countries of the world, said the report issued Wednesday. That electricity use produced about 208 million tons (189 million metric tons) of carbon dioxide, about the same amount as Argentina, and producing that much energy consumed about 1.2 trillion gallons (4.5 trillion liters) of water, according to the report on the environmental consequences of AI’s energy use.

By 2030, data centres will account for nearly three per cent of the world’s projected electricity use, with 935 trillion watt-hours. If data centres were a country, the country would be projected to rank sixth-highest in power use in 2030. That would produce nearly 440 million tons (399 million metric tons) of carbon dioxide, the report said. The study focused on energy use and didn’t examine the massive amount of water used to cool data centers.

“If you look at these numbers, we’re seeing scales comparable to nations,” said study co-author Kaveh Madani, a water scientist and director of the United Nations University Institute for Water, Environment and Health in Canada. “The demand is enormous.”

Much of the growth of data centres is being driven by AI. About 20% of data centres’ energy is currently due to AI, but that should grow to 40% by 2030, the report said.

First global look at ecological impact

The report is significant because of the credibility and authority of the U.N., not just because of any one set of eye-popping numbers, said Fengqi You, a Cornell University energy engineering professor who directs the college’s AI sustainability issues.

“Its value is that a U.N. institution is putting carbon, water, land, life-cycle impacts and environmental justice into one frame” for an issue that is often shrouded in secrecy and partial disclosures, said You, who was not part of the report.

• Latest updates on artificial intelligence news here

“The general public should be concerned, but not panicked,” he added.

Jean Su, director of the Energy Justice Program at the Center for Biological Diversity, said the report is important because it is the first U.N., or even global, report “that shines a light on the environmental harms of AI.”

National Artificial Intelligence Association President Caleb Max emphasized how his industry is becoming more efficient and how it benefits the public: “AI is rapidly becoming part of our everyday lives and adding benefits that improve safety, live longer, work more efficiently, enhance food production, and reduce poverty. The evidence is growing daily that the energy return on investment of AI development is transformative for our world and therefore more than worth it.”

Josh Levi, president the Data Center Coalition, said the industry takes its environmental impact seriously.

“We remain committed to working with policymakers, local communities, and industry partners to ensure that as data centres grow, they do so responsibly, transparently, and in ways that reflect the best available practices,” he said in a statement.

How much energy your query uses and how to trim it

Madani, also the winner of the most recent of the Stockholm Water Prize, said the numbers show the environmental cost of AI, which may seem cleaner at first glance than other mechanical devices, such as cars and furnaces, that have visible pollution.

“AI is not just a virtual thing. We’re talking about something that has physics, something that has real impacts. There is infrastructure there. There is energy that is being used,” Madani said. “A lot of hardware is behind all these operations that to us seem very, very clean because we don’t see smoke out of our devices. On our cellphone, there is no visible smoke or out of our computer or something. But somewhere else someone is suffering.”

People can reduce AI’s massive energy appetite by being less polite and more concise in their queries, Madani said. The report found that cutting word use in requests by 30% can reduce energy used by AI by 25%. That would save about the same amount of electricity as what about 700,000 people in Africa use in a year, the report said.

“If you’re too polite, then that extra ‘please’ you put there can make a huge difference,” Madani said. “You’ve got to be very precise and be short.”

A typical ChatGPT-style query is about 200 times more energy-intensive than the type of basic text classification used in an email spam filter, for example. AI-generated images or video require much more energy.

• Latest technology news on BNNBloomberg.ca

And the more complicated the AI, the more energy it takes to train or learn. The report said GPT-3 used about 1.3 billion watt-hours to train, but the next version used 50 to 70 billion watt-hours.

But it’s not training that really feasts on power, said study co-author Miriam Aczel, a United National University environmental policy researcher. About 90% of the power use of AI comes from operational requests, she said. GPT alone accounts for 2.5 billion prompts a day, she said.

Efficiency still means more power use

Even though tech advocates can argue that their machines are becoming more efficient, there’s a common paradox that finds when things get more efficient, they are used more often and total energy use soars even if individual uses are more efficient, Madani said. While some companies tout the use of renewable energy for data centres, Madani said that means the supply of clean electricity is depleted and thus dirtier energy is used elsewhere.

One of the problems in conducting this study is that many companies and places are not transparent about what data centres and AI are consuming or even where and how big they are, Aczel and Madani said.

“We cannot manage what companies do not disclose,” Cornell’s You said.

___

Seth Borenstein, The Associated Press