George Markezinis
AI systems rely heavily on data centres that now use more electricity than Saudi Arabia, with each accounting for about 1–1.5 percent of global electricity use. Artificial intelligence can have a significant impact on the environment, particularly when used for specific purposes.
According to a report by Goldman Sachs, for each single AI chat query, there is ten times more power usage than searching on Google. Additionally, training large AI models leaves an enormous environmental footprint. This amount of power could sustain 120 average U.S. homes for a year. Generative AI needs 33 times more energy than traditional software to complete tasks. Tech giants face rising carbon emissions due to this growing power need. For example, Microsoft’s annual emissions jumped 40% between 2020 and 2023.
In previous pieces, we have analyzed how the ongoing development of AI can facilitate business sustainability and help disaster management. However, the environmental aftermath is massive, and it also disrupts various communities.
The more artificial intelligence model training evolves, the greater the costs to the environment. Research from the University of Massachusetts Amherst shows that training just one AI model creates over 626,000 pounds of CO2. This matches what five cars emit in their lifetime. GPT-3’s training phase used up 1,287 MWh of electricity and produced 502 metric tons of carbon emissions. These numbers equal the emissions from 112 gas-powered cars running for a year.
AI’s energy usage keeps growing at an alarming rate. Data centers now make up 2.5 to 3.7 percent of global greenhouse gas emissions. When comparing these numbers with the aviation industry and what it produces, the divergence is evident. We expect most companies to grow their processing requirements at least 50 times more by 2028 as they compete to build the best models.
Through inference operations, AI’s replies to user queries create an environmental issue that isn’t always considered. Google’s research shows that inference operations use 60% of Artificial Intelligence-related energy, while training takes up the other 40%. A single ChatGPT query uses 100 times more energy than a regular Google search. GPT-3’s daily inference operations leave a carbon footprint of 50 pounds of CO2. This adds up to 8.4 tons each year.
AI chip production creates its own set of environmental challenges through heavy water and energy use. Most manufacturing plants sit in areas that rely on fossil fuels and need substantial resources to make chips. New semiconductor facilities lead to more gas-powered energy infrastructure worldwide. The manufacturing process needs complex steps from raw material extraction to chip making. Each step adds to greenhouse gas emissions.
Environmental inequality has become a serious concern in Artificial Intelligence development that affects communities based on their location and economic status. Data centers located in regions that rely heavily on fossil fuels produce higher carbon emissions, which cause environmental problems for local areas.
Different regions face unequal environmental costs from AI infrastructure. Data center energy consumption will reach 6% of total electricity usage by 2026. Water-intensive cooling systems put extra pressure on areas like Arizona where freshwater is already scarce. The construction of new semiconductor facilities has led to more gas-powered energy infrastructure worldwide.
Communities that face discrimination because of their race, ethnicity, and economic status carry the biggest environmental burden. Recent artificial intelligence advances mostly benefit wealthier communities. The less economically advantaged and vulnerable communities remain excluded from AI benefits because they are underrepresented in data. The environmental cost shows up through:
Also, the gap in access to green technology makes existing inequalities worse. Wealthy regions with better infrastructure tend to have lower carbon footprints from their AI operations. Departments that serve vulnerable communities can’t invest in green AI development because they lack funding and support. The shortage of digital skills and access creates obstacles for marginalized communities that try to benefit from AI-driven environmental solutions.
Federal policies on AI’s environmental effects continue to change through executive orders and new laws. The White House released an executive order on advancing United States Leadership in Artificial Intelligence Infrastructure. According to the directive, the Departments of Defence and Energy will use federal locations for the creation of large-scale AI data centres. These projects will be funded by private companies selected through proposals that will also include the implementation of clean energy solutions. The initiative aims to improve the lives of US citizens and national security and help the country maintain its leadership in ArtificiaI Intelligence.
To promote sustainable AI practices and mitigate its ecological footprint, the Artificial Intelligence Environmental Impacts Act was introduced in the U.S. Congress in 2024, mandating the Environmental Protection Agency (EPA) to conduct a comprehensive study on AI’s environmental effects.
Specifically, the law requires:
Industry leaders also created shared initiatives to tackle environmental concerns. The U.S. Artificial Intelligence Safety Institute Consortium, founded by the NIST, brings together more than 280 organizations to develop science-based and empirically backed guidelines and standards for AI measurement and policy to start spreading AI safety across the globe.
Self-regulation by industry is a chance to address specific sector needs better than government rules. But voluntary initiatives still need to prove they work, since many industry-specific groups are just getting started.
Groundbreaking initiatives by global organizations show practical ways to develop sustainable artificial intelligence. The United Nations Environment Program leads these efforts through state-of-the-art monitoring and optimization solutions on its digital platforms.
Smart optimization techniques reduce AI’s environmental footprint without affecting performance. Model pruning and quantization methods need less computing power and lower energy use. Tensor processing units and field-programmable gate arrays work better than traditional hardware. These optimizations have cut energy use by 96% for AI inference tasks in the last two years.
State-of-the-art data centers are the lifeblood of sustainable AI infrastructure. Major cloud providers have launched detailed sustainability measures:
The World Environment Situation Room (WESR) launched in 2022 serves as a central platform to monitor environmental changes. This system tracks real-time data about CO2 in the atmosphere, glacier mass changes, and sea level variations. The International Methane Emissions Observatory (IMEO) adds to this capability by keeping a global database of verified methane emissions.
UNEP’s GEMS Air Pollution Monitoring Platform is the world’s largest air quality network. It collects data from over 25,000 monitoring stations in 140 countries. These systems help the public and private sectors use digital technologies instead of old monitoring methods. WESR will soon work as a mission control center for planetary environmental indicators to enable smooth monitoring and quick responses.
Artificial Intelligence has the capability of tackling some of the world’s greater environmental disasters. Its technology is already being used to transform disaster management and enhance resilience.
Nonetheless, the negative aspect of AI’s explosion with regards to the energy consumption cannot be overlooked. There is a profound need for concerted effort on behalf of technology companies, policymakers, and environmental groups to join forces in the goal of reducing AI’s environmental footprint. We should not interpret this as another demonization of Artificial Intelligence development, but rather as a balanced approach to technology and environmental responsibility, ensuring mutual benefits for society and the planet.
