Artificial Intelligence, commonly known as AI, is becoming a huge part of culture. It’s been around for a while—starting as simple suggestions for what to search, and growing into advanced models like ChatGPT and Sora.

Commonly, when AI is brought up, it’s implied to be generative AI. Generative AI models are used to answer questions, make photos, and even form videos. Every time a prompt or query is put into an AI model, it goes through billions of parameters to come up with an accurate answer.

Creating an AI model takes up a lot of energy. According to research out of Penn State University, AI models need to be trained with billions of parameters and tested repeatedly, which uses massive amounts of energy.

Haven has an AI committee, consisting of ten teachers. English teacher Mr. Matthew Morris, who is on the AI committee, explains the hardships of doing research about AI models.

“The most popular companies use closed models, and they don’t share the information about how much energy they’re using per prompt,” Morris said. “Some researchers have looked into that, extrapolating from open models, but it’s pretty hard to tell how the most recent models’ energy uses on a per prompt basis.”

Despite this drawback, researchers at MIT used Meta-Llama, an open-sourced AI model, to estimate the energy usage of more popular, closed models.

This research is discussed in an article in the MIT Technology Review, and it shows how the number of parameters in a model affects the amount of energy it requires. In the tests, different prompts were inputted to models with various amounts of parameters, and the energy used was recorded.

The results were that a small model with eight billion parameters only needed 114 joules to respond to a prompt, while a bigger model with 405 billion parameters needed 6,706 joules for every response. 

For context, watching an hour of Netflix takes approximately 2.8 million joules.

Relatively speaking, this isn’t a lot of energy, but it grows as the number of queries does.

However, companies such as DeepSeek have created models with many more parameters. The energy might not seem like a lot with the smaller models, but when thinking about the number of queries popular generative AI models answer each day, as well as the amount of parameters they must have, the energy used to respond to the questions increases rapidly.

While MIT’s test is a good comparison of the energy AI models could consume, it isn’t accurate for the popular AI companies, as the researchers could only guess as to the number of parameters and the needs of the big AI models.

The Penn State research dives deeper into the total percentage of energy used on AI. The article states that in 2023, 4.4% of the electricity in the United States went to AI data centers. It is also shown there that the number would increase significantly as time passes. (It could use more than 10% of the United States’ electricity as soon as 2028.)

Clearly, there is a lot of energy being used for AI models, but is there anything being done to make it more energy efficient? 

Morris explains that major AI companies are more focused on business and production, instead of conserving energy.

“Rather than make things more efficient as a goal, [companies are] trying to ramp up energy production by doing things like reopening old nuclear power plants,” Morris said. 

Not only is this taking up a lot of energy, but the type of energy being used has a huge impact on the environment, which could potentially bring harm to it.

“There’s a bunch of data centers in Virginia that are using natural gas to power them, and in some places in the country where the regulations aren’t as strict, they’re using more coal, fire power, plant power,” Morris said.

The Penn State research mirrors this sentiment, elaborating that it contributes to greenhouse gas emissions, as most of the AI models use electricity made from fossil fuels. The article also explains that huge quantities of water are needed for the AI model cooling systems, leading to water scarcity.

AI consumes large amounts of energy, leading to various impacts on the environment. Primarily, the companies are not decreasing the energy used, keeping them in an upward cycle. 

Morris explains how the advancements of AI aren’t changing the way energy is used.

“Companies are using more and more powerful processors to run those processes and run those queries, and that more powerful hardware also demands more power usage,” Morris said.