5 Stark Stats Showing AI’s Unsustainable Impact

The energy required for the surge in artificial intelligence and large language models like ChatGPT is creating a new set of sustainability challenges for brands and agencies.

Generative AI is rapidly changing how marketers do their jobs, improving systems and streamlining tedious processes.

As the deadlines for corporate net zero targets inch nearer, the use of the technology is creating an increase in energy demand that skyrockets past the estimates that companies set years ago.

In an event last week on AI and sustainability hosted by the 4A’s, Jack Morton senior vice president and global head of sustainability Julien Le Bas outlined the impact of AI and what strategies marketers and advertisers can use to reduce impact.

Here are five eye-popping stats on AI’s sustainability impact.

1. One data center uses the annual equivalent of 7 million laptops running eight hours per day

The energy needs currently associated with AI are hard to visualize, but estimates show that data centers—which hold the hardware that makes AI work—could account for 10% of global energy use by 2025, or 25% of U.S. energy use by the end of the decade.

2. Each year, training AI models will require 10 times more energy

That’s according to OpenAI’s Sam Altman, and it is a main reason why he has pushed for investment in new energy sources like nuclear fusion. A breakthrough is necessary to power AI, he has said.

The problem? Research on nuclear fusion has been ongoing for decades. And in the meantime, the energy sources we do have are accelerating climate change. As tech companies build more data centers across the country to support LLMs, power grids from Washington to Georgia to Mississippi to Nebraska to Utah are feeling the strain—and shelving plans to close coal plants.

“We have a little bit of magical thinking that renewable energy is going to scale up to meet all of AI’s anticipated demands in the next decade,” Alison Pepper, 4A’s executive vp of government relations and sustainability, said. “I’m not sure that’s entirely feasible.”

3. For every five to 50 questions asked, ChatGPT requires 500 milliliters of water

It’s not just power that AI requires—these data centers also need fresh water to operate the cooling systems necessary to keep facilities safe.

“Think about your computer, when it starts to ventilate and gets very hot,” Le Bas said. “[Then] multiply that by thousands. We need massive cooling systems.”

By 2027, the water needed for those cooling systems will amount to one-half of what the entire U.K. is currently using, according to one estimate from Cornell University.

4. ChatGPT-3 was trained on 45 terabytes of text data, translated for AI using human labor

The social cost of training AI isn’t currently well understood, Le Bas noted.

But reporting from Time showed that some contractors in Kenya were paid less than $2 per day to sift through disturbing content generated by early versions of ChatGPT, highlighting the labor-related risks associated with outsourcing some aspects of training LLMs.

5. General models can use 30 times more energy than specialized models

One way marketers can curb the impact of their AI use is by using specialized models when applicable. A 2023 study by Carnegie Mellon and Hugging Face showed that specialized models are much more energy-efficient that general models like ChatGPT.

Another way companies can cut the carbon cost of AI is by adopting “conscious computing” practices, which aim to encourage lower-impact tools when appropriate, Le Bas said. Rather than putting every query through ChatGPT, conscious computing principles would urge people to consider whether there’s a more specialized or simple option for the task. Browser plug-ins like Carbon Scaledown can track GPT-related emissions.