By: Ashley Bergeron
Staff-Writer
With fires in and around Los Angeles displacing millions and destroying miles of the city, as well as snowstorms taking the eastern side of the country by surprise, climate change is as salient a discussion as ever. Within these discussions, the topic of how artificial intelligence has impacted the environment has risen.
Since 2012, according to OpenAI, the amount of computing power required to train generative AI has doubled every 3.4 months. With the increasing popularity of the technology, the usage of AI also increases. In order to support the demand for AI development, Google’s emissions have increased by 48% over the past four years.
In 2023, a study from University of California Riverside found that running 10 to 50 medium-length responses (with input of less than 800 words and output of 100 to 300 words) on GPT-3, a large language model created by OpenAI, requires 500mL of water to run properly – half of a Nalgene water bottle. Why is this?
Like any energy-intensive process, the use of generative AI creates heat. In order to prevent their servers from overheating, AI data centers use water-based cooling systems.
Furthermore, a study published in December 2024 looking at data centers in Africa showed that using GPT-4 to create a 10-page paper report consumes between 700mL to 60,000mL of water (up to 60 Nalgenes). Generating a 100- to 120-word email also resulted in a loss of water ranging from 400mL to 3000 mL.
Carbon emissions related to AI have also risen over the years. According to the Washington Post, by 2030, data centers in Virginia will require the same amount of energy it takes to power 6 million homes. The decommissioning of some coal plants has also been delayed to keep up with AI energy demands.
Companies are also planning to reopen nuclear plants to support AI usage. Microsoft plans to restart a nuclear power plant in 2028 with Constellation, and other companies like Amazon and OpenAI are planning to collaborate with X-Energy to make new, smaller nuclear power plants. X-Energy’s design still needs approval from the Nuclear Regulatory Committee but expects to see its first plant to open by 2030.
Some people are skeptical about the push towards nuclear energy, as it takes a few years to launch a power plant.
Unfortunately, since AI is such a new technology, it is difficult to confidently predict the environmental impact it will have. Therefore, many estimates will either underestimate or overestimate. On top of that, companies like OpenAI don’t openly share their carbon emissions, making it more difficult to feel confident with the estimates.
Despite these issues, AI can also be helpful with tracking and combating climate change. Researchers in China discovered that AI can help with clearing up carbon emissions, as well as to track methane emissions and calculate environmental footprints.
If you’re curious about the emissions of AI, there is a calculator called EgoLogits that calculates the impact of each AI model. You can visit it at huggingface.co/spaces/genai-impact/ecologits-calculator.
Links to the studys (just in case the citation needs to be better or if we want to do hyperlinks on the web)
https://arxiv.org/pdf/2304.03271
