Can we afford the environmental cost of AI?

Artificial intelligence has made incredible strides as a tool for biodiversity conservation.

But paradoxically, as AI technology has grown ever more complex and innovative, so too has its environmental impact.

What’s more, generative AI – think chatbots like ChatGPT, or AI image and video generators – is taking a growing toll on the planet and will likely continue to do so.

Tanya Berger-Wolf, a computational ecologist and professor of computer science and engineering at Ohio State University, has spent over a decade creating algorithmic solutions for conservation.

“The majority of AI that is being used for biodiversity and climate change is old-school AI,” she says.

“The most useful type of AI for biodiversity monitoring, for example, is detection algorithms that can detect objects in images or sounds in a recording – and those are super simple and affordable.”

But then came generative AI, with the unveiling of OpenAI’s ChatGPT in 2022 – and it was made available not only to scientists and analysts but to the general public as well.

A deluge of data centers

Data centers are facilities that store computer equipment used to generate, manage and store data. More than 6,000 have sprung up all over the world, and that number will rise to almost 8,400 by 2030, largely driven by the growth of generative AI.

Most generative AI runs on large language models that “only work when you have billions of data points to train them and massive resources,” Berger-Wolf explains.

Those resources include electrical power, which generates a lot of heat, as well as the water required to cool the equipment down.

Tech companies themselves have admitted that their products are placing increasing pressure on natural resources.

In July 2022, Microsoft used 52 million liters of water to train ChatGPT in West Des Moines, Iowa. That amounted to 6 percent of the entire city’s water use that month and caused Microsoft’s water consumption to spike by 34 percent that year.

Not wanting to be left behind, Google raced to bring its Gemini AI model to market in 2023.

In its 2024 Environment Report, it disclosed that its carbon emissions had increased by 48 percent compared to 2019, “primarily due to increased data center energy consumption and supply chain emissions.”

Its data center water use, meanwhile, increased by nearly 88 percent during the same period, “primarily due to water cooling needs at our data centers, which experienced increased electricity consumption year-over-year.”

If AI were a country…

According to the International Energy Agency (IEA), a single ChatGPT query uses almost 10 times as much electricity as a simple Google search.

While it’s still unclear how people will adopt AI for personal use – for instance, videos are far more energy-intensive to generate than text – the IEA predicts that data centers could double their electricity consumption between 2024 and 2026.

Similarly, a 2024 study by Goldman Sachs Research estimates that AI could cause power use from data centers to increase by 160 percent by 2030.

It suggests that data centers in the U.S. could suck up nearly 10 percent of the country’s entire electricity use in the next five years, up from 4 percent in 2023. Power utilities will need to invest about USD 50 billion in new generation capacity to meet that demand.

In Europe, the study projects that power demand could grow by 40 to 50 percent between 2023 and 2033 “thanks to both the expansion of data centers and an acceleration of electrification.”

The power needs of data centers alone, it says, “will match the current total consumption of Portugal, Greece and the Netherlands combined” by 2030.

For some countries, data centers are already putting a strain on resources. In Ireland, they already consume 21 percent of electricity – more than all of the country’s urban homes combined.

Ireland’s electricity consumption grew by nearly a quarter between 2012 and 2022, the second-fastest rate in the EU. This was almost entirely driven by the expansion of data centers, according to a recent report by Friends of the Earth.

Globally, data centers are expected to produce about 2.5 billion tons of carbon dioxide-equivalent emissions between now and 2030 – amounting to roughly 40 percent of the U.S.’s annual emissions.

Drinking the Earth dry

AI isn’t just driving an explosion in energy consumption and carbon emissions. It’s also a drain on water supplies.

Shaolei Ren, an associate professor of computer engineering at the University of California, Riverside, has led
multiple studies on the environmental impacts of AI.

He’s calculated that a large generative AI model uses between half a liter and three liters of water to write a short email of 100 to 250 words, depending on where the server is located and the ambient temperature.

“Water is a regional issue,” he says. “If you look at particular regions – let’s say Arizona, Spain, Taiwan or Singapore – they don’t really have a lot of freshwater resources. We can’t just look at the whole continent.”

Data centers don’t just use water onsite for cooling, he adds, but also further up the supply chain.

“Most companies just focus on reporting their direct water consumption, but the true water cost of data centers include both water at the facility for cooling the servers and also the water used to produce the electricity offsite.”

Both direct and indirect water use need to be taken into account when evaluating the overall impact on water resources, Ren points out.

Likewise, when calculating emissions, it’s important to include both direct discharge from the facility and indirect discharge from the electrical grid supplier.

“We need to have a more comprehensive cost-benefit analysis to make an informed decision, which I think is crucial,” says Ren. “At this point, there is very little information.”

To that end, he has been looking at another AI metric: public health impacts from the burning of fossil fuels.

Is AI a public health risk?

In a paper he co-authored last year, he calculates that data centers could be costing the U.S. upwards of USD 20 billion a year – double the public health impact of the country’s coal-based steelmaking and “comparable to that of on-road emissions of California.”

He also estimates that just training an AI model the size of Meta’s Llama 3.1 “can produce air pollutants equivalent to more than 10,000 round trips by car between Los Angeles and New York City.”

Most major tech companies have set net zero goals, but they’re mainly relying on purchasing carbon credits – often generated in a different region with a cleaner energy grid – rather than reducing their own emissions.

For Sebastián Lehuedé, an assistant professor at the Department of Digital Humanities at King’s College London, that is woefully insufficient.

“There are so many issues around carbon credits from an ecological perspective,” he says.

“If you consume water somewhere to the point where it affects biodiversity in one area, that cannot be offset by having a nice project elsewhere. You’re going to cause irreversible damage if you keep to that logic.”

Big tech companies are aware of these pitfalls, which is why they’re increasingly turning to renewable energy. Microsoft has even signed a deal with Constellation Energy to resurrect a unit at the Three Mile Island nuclear plant in Pennsylvania, five years after it was shuttered.

Yet for Lehuedé, who has been working with communities affected by AI, these investments show that tech companies are becoming de facto policymakers, deciding where projects will be located and what standards they will use.

“That’s very dangerous,” he says. The building of renewable energy facilities “is something that needs to be planned with democratic input. How these systems are going to work shouldn’t be left to the market.”

Can we make AI sustainable?

So, can tech companies use all their computing prowess to make their products more energy-efficient?

That expertise exists, and “people are working on it,” says Berger-Wolf, “developing resource-efficient models and computational approaches to eliminate the need for retraining every time you update your data sets.”

Work is also being done on better cooling systems, and energy-efficient computer chips and data storage – which used to be very cheap, she says.

“But the problem is the way that these models access data. Moving data is expensive. These models access data for training and inference, which is also energy-intensive. So, all of these engineering aspects are being worked on.”

Berger-Wolf remains concerned, nonetheless, about the lack of political will to invest in and use less resource-hungry software and hardware. As long as companies are focused on their bottom line, there is a limit to how much they’re willing to cut.

“If they can pass the cost to the consumer, they will,” she says.

Communities are already paying the cost of data centers – whether it’s through water shortages, emissions, higher energy bills or noise. From the U.S. to Ireland to Mexico, angry protests are becoming increasingly common.

“We are going to reach a tipping point where the increasing cost of data and hence, AI, is not just environmentally expensive but also socially expensive,” says Revathi Kollegala, a digital strategist at CIFOR-ICRAF.

“This will undermine the logic that AI can democratize access to knowledge and reduce inequity. We may have reached that point already or will very soon.”