LLMs' Environmental Footprint: A Comparative Analysis

The relentless pursuit of artificial intelligence (AI) innovation has led to the development of increasingly sophisticated large language models (LLMs). While these models showcase impressive capabilities in various domains, the environmental consequences associated with their training and deployment have largely remained obscured. Companies developing AI models readily share performance data on benchmarks, but tend to avoid the environmental impact. Recent research sheds light on the often-overlooked energy, water, and carbon costs associated with these powerful AI tools.

A New Benchmark for Assessing Environmental Impact

In the quest to quantify the environmental impact of AI, a team of researchers from the University of Rhode Island, Providence College, and the University of Tunis has introduced an infrastructure-aware benchmark for AI inference. This research, available on Cornell University’s preprint server arXiv, offers a more precise evaluation of AI’s ecological effects. The benchmark combines public API latency data with information on the underlying GPUs and regional power grid compositions to calculate the environmental footprint per prompt for 30 mainstream AI models. This comprehensive approach considers energy consumption, water usage, and carbon emissions, culminating in an “eco-efficiency” score.

Abdeltawab Hendawi, assistant professor at the University of Rhode Island, explains the motivation behind the study: "We started to think about comparing these models in terms of environmental resources, water, energy, and carbon footprint." The findings reveal significant disparities in the environmental impact of different AI models.

Disparities in Energy Consumption: OpenAI, DeepSeek, and Anthropic

The study highlights substantial differences in energy consumption among leading AI models. OpenAI’s o3 model and DeepSeek’s primary reasoning model consume more than 33 watt-hours (Wh) for a single extended response. This contrasts sharply with OpenAI’s smaller GPT-4.1 nano, which requires over 70 times less energy. Anthropic’s Claude-3.7 Sonnet emerges as the most eco-efficient model in the study.

The researchers emphasize the crucial role of hardware in determining the environmental impact of AI models. For example, the GPT-4o mini, which uses older A100 GPUs, consumes more energy per query than the larger GPT-4o, which operates on more advanced H100 chips. This underscores the importance of leveraging cutting-edge hardware to minimize the environmental footprint of AI.

The Environmental Toll of Query Length

The study reveals a direct correlation between query length and environmental impact. Longer queries invariably lead to greater resource consumption. Even seemingly insignificant, short prompts contribute to the overall environmental burden. A single brief GPT-4o prompt consumes approximately 0.43 Wh of energy. Researchers estimate that at OpenAI’s projected 700 million GPT-4o calls per day, the total annual energy consumption could range from 392 to 463 gigawatt-hours (GWh). To put this into perspective, that’s enough energy to power between 35,000 American homes annually.

The Cumulative Impact of AI Adoption

The study emphasizes that individual users’ adoption of AI can quickly escalate into substantial environmental costs. Nidhal Jegham, a researcher at the University of Rhode Island and the study’s lead author, explains that "Using ChatGPT-4o annually consumes as much water as the drinking needs of 1.2 million people annually." Jegham cautions that while the environmental impact of a single message or prompt seems negligible, "once you scale it up, especially how much AI is expanding across indices, it’s reallybecoming a rising issue."

Delving Deeper into the Environmental Impact Metrics

To fully appreciate the implications of the study’s findings, a more detailed examination of the environmental metrics used to evaluate the AI models is essential. The following sections provide a breakdown of the key metrics:

Energy Consumption

Energy consumption is a fundamental measure of the electrical power required to operate AI models. The study quantifies energy consumption in watt-hours (Wh) per query, allowing for a direct comparison of the energy efficiency of different models. Minimizing energy consumption is critical for reducing the carbon footprint and overall environmental impact of AI.

Factors Influencing Energy Consumption:

• Model Size and Complexity: Larger and more complex models typically require more energy to operate than smaller, simpler models. This is due to the increased computational resources needed for processing and generating responses. The number of parameters, the depth of the neural network, and the complexity of the algorithms all contribute to energy demands. For instance, models with billions of parameters require significantly more energy to run than models with a few million.
• Hardware Efficiency: The GPUs and other hardware components used to run AI models play a significant role in energy consumption. More advanced and energy-efficient hardware can substantially reduce the energy footprint of AI. Modern GPUs are designed to perform complex calculations more efficiently, using techniques such as reduced precision arithmetic and optimized memory access. The choice of hardware can make a substantial difference in the energy required to run a given AI model.
• Query Length and Complexity: Longer and more complex queries generally require more computational resources and thus consume more energy. The model needs to process, understand, and generate responses for longer inputs, which demands more processing power and memory. Complex queries may involve multiple steps, requiring the model to perform more calculations and access more data, further increasing energy consumption.
• Optimization Techniques: Various optimization techniques, such as model compression and quantization, can reduce the energy consumption of AI models without sacrificing accuracy. Model compression involves reducing the size of the model by removing redundant parameters or using more efficient data representations. Quantization reduces the precision of the model’s parameters, allowing for faster computation and lower memory requirements. Other optimization techniques include knowledge distillation and pruning.

Water Usage

Water usage is an often-overlooked aspect of the environmental impact of AI. Data centers, which house the servers that run AI models, require substantial amounts of water for cooling. The study estimates water usage based on the energy consumption of the data centers and the water intensity of the regional power grids that supply electricity to those data centers.

Factors Influencing Water Usage:

• Cooling Requirements: Data centers generate significant heat and require cooling systems to maintain optimal operating temperatures. Water is often used as a coolant, either directly or indirectly through cooling towers. The heat generated by servers and other equipment needs to be dissipated to prevent overheating and ensure reliable operation. Traditional cooling systems rely on water to absorb and transfer heat, leading to significant water consumption.
• Power Grid Water Intensity: The water intensity of the power grid refers to the amount of water required to generate a unit of electricity. Power grids that rely heavily on thermoelectric power plants, which use water for cooling, have higher water intensities. Thermoelectric power plants, such as coal, natural gas, and nuclear plants, use water to condense steam and cool equipment. The amount of water required varies depending on the technology and the efficiency of the plant.
• Data Center Location: Data centers located in arid regions or regions with water scarcity issues can exacerbate the environmental impact of AI. Operating a data center in a water-stressed area puts additional strain on local water resources and can lead to conflicts with other water users. It is important to consider the availability of water resources when selecting a location for a data center.

Carbon Emissions

Carbon emissions are a primary driver of climate change. The study calculates carbon emissions based on the energy consumption of the AI models and the carbon intensity of the regional power grids. Carbon intensity refers to the amount of carbon dioxide emitted per unit of electricity generated.

Factors Influencing Carbon Emissions:

• Energy Source: The type of energy used to power data centers has a significant impact on carbon emissions. Renewable energy sources, such as solar and wind power, have much lower carbon intensities than fossil fuels like coal and natural gas. Using renewable energy sources can drastically reduce the carbon footprint of AI.
• Power Grid Carbon Intensity: The carbon intensity of the power grid varies depending on the mix of energy sources used to generate electricity. Regions with a higher proportion of renewable energy sources have lower carbon intensities. Some regions rely heavily on coal-fired power plants, resulting in high carbon emissions per unit of electricity. Other regions have invested heavily in renewable energy, leading to lower carbon emissions.
• Energy Efficiency: Reducing energy consumption is the most effective way to lower carbon emissions. The less energy an AI model consumes, the fewer carbon emissions are associated with its operation. Improving energy efficiency can be achieved through a variety of techniques, including model optimization, hardware upgrades, and efficient cooling systems.

Implications and Recommendations

The study’s findings have significant implications for AI developers, policymakers, and end-users. The environmental impact of AI is not negligible and needs to be carefully considered as AI technology continues to advance and proliferate.

Recommendations for AI Developers:

• Prioritize Energy Efficiency: AI developers should prioritize energy efficiency when designing and training AI models. This includes using smaller models, optimizing code, and leveraging efficient hardware. Explore techniques like knowledge distillation, pruning, and quantization to reduce model size and computational complexity without significant loss of accuracy. Efficient coding practices and optimized algorithms can also minimize energy consumption. Using specialized hardware, such as TPUs (Tensor Processing Units), that are designed for AI workloads can also contribute to energy savings.
• Explore Renewable Energy Sources: AI companies should explore opportunities to power their data centers with renewable energy sources. This can significantly reducethe carbon footprint of AI. Investing in on-site renewable energy generation, such as solar panels, or purchasing renewable energy credits (RECs) can help offset the carbon emissions from data center operations. Collaborating with utility companies to increase the availability of renewable energy in the region can also have a positive impact.
• Invest in Water Conservation: Data centers should invest in water conservation technologies to minimize water usage. This includes using closed-loop cooling systems and rainwater harvesting. Closed-loop cooling systems recirculate water, reducing the amount of water needed for cooling. Rainwater harvesting can provide a source of water for cooling and other non-potable uses. Implementing water leak detection systems and optimizing cooling system efficiency can also help reduce water consumption.
• Transparency and Reporting: AI companies should be transparent about the environmental impact of their models and report key metrics such as energy consumption, water usage, and carbon emissions. Providing clear and accessible information about the environmental footprint of AI models allows users to make informed choices and encourages developers to prioritize sustainability. Standardized reporting frameworks and certifications can help ensure comparability and accountability.

Recommendations for Policymakers:

• Incentivize Green AI: Policymakers should incentivize the development and deployment of green AI technologies through tax credits, subsidies, and other incentives. Tax credits can encourage AI companies to invest in energy-efficient hardware, renewable energy sources, and water conservation technologies. Subsidies can support research and development of new green AI technologies.
• Regulate Data Center Energy Consumption: Policymakers should regulate data center energy consumption to ensure that data centers are operating as efficiently as possible. Setting energy efficiency standards for data centers and requiring regular energy audits can help reduce energy consumption. Providing incentives for data centers to adopt best practices for energy efficiency can also be effective.
• Promote Renewable Energy Adoption: Policymakers should promote the adoption of renewable energy sources to reduce the carbon intensity of power grids. Investing in renewable energy infrastructure, providing tax incentives for renewable energy development, and setting renewable energy targets can help accelerate the transition to a cleaner energy system.
• Support Research and Development: Policymakers should support research and development into new technologies that can reduce the environmental impact of AI. Funding research into more energy-efficient AI algorithms, hardware, and cooling technologies can help address the environmental challenges of AI. Supporting research into the life cycle assessment of AI models can provide a more comprehensive understanding of their environmental impacts.

Recommendations for End-Users:

• Be Mindful of AI Usage: End-users should be mindful of their AI usage and avoid unnecessary or frivolous queries. Each AI query consumes energy, so reducing the number of unnecessary queries can help reduce the overall environmental impact of AI. Before using AI, consider whether the task can be accomplished efficiently through other means.
• Choose Eco-Friendly AI Models: When possible, end-users should choose AI models that are known to be more energy-efficient. Some AI models are designed to be more energy-efficient than others. Choosing these models can help reduce the environmental impact of AI. Look for information about the energy efficiency of AI models before using them.
• Support Sustainable AI Practices: End-users can support sustainable AI practices by choosing AI products and services from companies that are committed to environmental responsibility. Supporting companies that are transparent about their environmental impact and are taking steps to reduce their carbon footprint can help encourage the development of more sustainable AI practices.

Future Research Directions

The study highlights the need for further research into the environmental impact of AI. Future research should focus on the following areas:

• Life Cycle Assessment: Conducting a comprehensive life cycle assessment of AI models, from development to disposal, to identify all potential environmental impacts. This would involve analyzing the environmental impacts associated with the extraction of raw materials, manufacturing of hardware components, energy consumption during training and inference, and the disposal of e-waste.
• Impact of Training: Investigating the environmental impact of training AI models, which can be significantly higher than the impact of inference. Training large AI models often requires massive amounts of computational resources and energy, leading to significant carbon emissions. Research should focus on developing more energy-efficient training methods and optimizing training infrastructure.
• Impact of AI on Other Sectors: Examining the impact of AI on other sectors of the economy, such as transportation and manufacturing, to understand the overall environmental consequences of AI adoption. AI can have both positive and negative environmental impacts in these sectors. For example, AI can improve energy efficiency in transportation and manufacturing, but it can also lead to increased consumption and waste.
• Development of New Metrics: Developing new metrics to assess the environmental impact of AI, such as metrics that account for the embodied energy and materials in AI hardware. Current metrics primarily focus on energy consumption during operation. Developing metrics that capture the environmental impacts associated with the manufacture and disposal of AI hardware would provide a more comprehensive picture of the environmental footprint of AI.

Conclusion

The environmental impact of LLMs is a complex and multifaceted issue that requires careful consideration. The findings of this study provide valuable insights into the energy, water, and carbon costs associated with popular AI tools. By understanding these costs, AI developers, policymakers, and end-users can take steps to minimize the environmental footprint of AI and ensure that AI technology is developed and deployed in a sustainable manner. As AI becomes more integrated into our lives, it is crucial to prioritize sustainability and work together to create a future where AI benefits society without harming the environment. The development and deployment of AI should not come at the expense of our planet, and a concerted effort is required to mitigate its environmental consequences.