Environmental Impacts of Generative Artificial Intelligence (genAI) Systems
The goals of this notebook is to inform you on
- The systems (data centers) that power language models such as ChatGPT and Microsoft Gemini
- Locations of notable data centers and their effects on communities
- Accuracy vs Emission Tradeoff
- How to handle language models in your work
Click here to access the notebook
Thirsty and Hungry
GenAI systems require a lot of water and electricity to operate them. The systems which answer our every request and millions of others are powered by data centers. Data centers are large buildings housing servers which do the work of answering prompts. These buildings get very hot from the servers, requiring cooling systems to regulate temperature Link. Data centers vary in size, with the average around 20,000-100,000 square feet Link, and the largest eclipsing 10 million square feet Link.
US Data Centers
Over half of data centers worldwide are in the US (roughly 5,000+), with high concentration in Virginia (N), Arizona (PHX), Georgia (ATL), Texas (DAL), and Oregon Link. With the abundance of data centers, there have arose issues with residents and the impact of these computer buildings. Phoenix-Arizona, The Dalles-Oregon, Memphis-Tennessee, Milwaukee-Wisconsin, and Omaha-Nebraska hold key stories of the rise of genAI systems colliding with people and the environment. Specifically, the Memphis, Tennessee battle with xAI is notable for the advancement of technologies by Elon Musk. They encroach on the livelihood of a predominantly Black neighborhood with illegal practices.
Global Data Centers
- The implications of genAI systems on communities are also felt globally.
- Latin American countries such as Uruguay and Chile have had to ward off AI expansion efforts. Link
- Dublin, Ireland halted new data center development over fears of blackouts in the city. Link
- The small vilage of Mekaguda, India recently filed a petition against Microsoft for dumping data center waste on the nearby land, negatively affecting livestock and crops. Link
- Singapore, the second-largest data center market in Southeast Asia, implemented a new “green strategy” to enforce cleaner resource use in the country. Link
Accuracy-Emission Tradeoff
Context-Specific Tasks
When completing tasks with generative models, it is important to consider model accuracy and model emissions. We will view 3 graphs of notable models answering questions from a study done in May 2025. Utilizing different models can have a big impact on emissions and energy use.
2023 Intergovernmental Panel on Climate Change (IPCC) report
- The IPCC is the United Nations governing body responsible for assessing science pertaining to climate change.
- Every 5-7 years, the IPCC publishes a report on the latest climate research.
- This study gives the various models questions from the report, such as “What are some examples of slow-onset events caused by climate change, and how do they affect the fire season length?”
- The 9 models performance is noted in the table below, as well as a graph of the respective accuracy and energy of their answers.
| Model | #Params | Duration(s) | Accuracy | Estimated Energy (Wh) |
|---|---|---|---|---|
| Qwen3-235B-A22B | 235B | 429.44 | 0.867 | 286 |
| phi-4 | 14.7B | 130.53 | 0.8 | 12.69 |
| Qwen2.5-72B-Instruct | 72B | 147.89 | 0.767 | 65.77 |
| Qwen3-32B | 32B | 167.97 | 0.733 | 65.32 |
| DeepSeek-R1-Distill-Qwen-32B | 32B | 441.23 | 0.733 | 35.30 |
| Llama-3.3-70B-Instruct | 70B | 336.71 | 0.567 | 149.64 |
| Phi-3-mini-4k-instruct | 3.82B | 103.79 | 0.533 | 2 |
| c4ai-command-r-plus-08-2024 | 104B | 482.24 | 0.533 | 428.44 |
| Llama-3.1-8B-Instruct | 8B | 279.95 | 0.52 | 5.6 |
Results
- Marginal decrease in accuracy but with large decrease in estimated energy.
- Bigger models (higher # of parameters) typically used more energy.
- Phi-4 (14.7B parameters) performed well on all three tasks, despite being one of the smallest.
- These models are all openly-sourced, so more studies have yet to be run to understand the tradeoff more.
How to proceed with language models through a critical lens
The study of environmental impacts of genAI systems is very new, but growing. Some models have the full open code and pre-trained data, but many do not. Many of the “best” models for doing homework or revising a resume have unknown side effects on the world. Consider the cost of such systems, and critique them further to shine more light on a relatively unknown, emerging technology.
Tools below for support
Environmental Transparency Visualization
TRY OUT A CHATBOT ENERGY CALCULATOR
Click here to access the notebook
The next section will allow you to tinker with a basic chatbot. We’ll be asking key questions that will initiate critical thinking when it comes to GenAI systems, the perception of these systems, and the implications of said systems.
Once you are done with this section and the notebook navigate to Prompting to begin!