Table of Contents
134 Colo. Env’t L.J. Vol. 37: Issue 1
Introduction
In 2023, Google launched Gemini, its most sophisticated Artificial Intelligence (“AI”) models yet.1 According to the company, the Gemini AI models possess state-of-the-art capabilities as a result of rigorous testing and performance evaluations for different tasks.2 Gemini Ultra was the first AI model to outperform human experts on a test covering subjects like math, ethics, and medicine.3
The AI models are multimodal models, which means they can understand text and also “natively understand, operate on, and combine other kinds of information like images, audio, videos, and code.”4 The models are able to analyze and answer queries about complex documents, no matter their size, due to their uniquely long context window.5 Additionally, by training the AI models on all forms of data at once, the models can
1 Sundar Pichai & Demis Hassabis, Introducing Gemini: Our Largest and Most Capable AI Model, GOOGLE: THE KEYNOTE (Dec. 6, 2023), https://blog.google/technology/ai/google-gemini-ai/#performance.
4 Harry Guinness, What is Google Gemini?, ZAPIER (Aug. 8, 2024),https://zapier.com/blog/google-gemini/.
5 Id.; “The context window . . . of a large language model (LLM) is the amount of text, in tokens, that the model can consider or ‘remember’ at any one time. A larger context window enables an AI model to process longer inputs and incorporate a greater amount of information into each output.” Dave Bergmann, What is a context window?, IBM (Nov. 7, 2024), https://www.ibm.com/think/topics/context-window.
2026 The Environmental Cost of Intelligence 135
purportedly “seamlessly understand and reason about all kinds of inputs from the ground up.”6 At the time of launching, this characteristic was quite unique.7
Recently, Google integrated its AI models into the search engine.8 Currently, when a consumer searches a topic on Google, an “AI Overview” appears at the top of the search results.9 This overview provides a summary and answer to the search through the use of Google’s generative AI.10 Google also recently introduced an “AI Mode” to its search engine.11 The company stated that “[t]his new Search mode expands what AI Overviews can do with more advanced reasoning, thinking and multimodal capabilities.”12 The new feature assists with answering more nuanced questions that previously required multiple queries.13
All in all, Google’s launch of its Gemini models and subsequent search engine features seems to provide promise in further developing AI’s use in everyday life. However, the promise of AI’s integration and increasing usage does not come without a cost.
Google’s 2024 Environmental Impact Report highlights concern over the negative environmental impact of AI.14 The report notes that Google has seen a 48% increase in its greenhouse gas emissions since 2019.15 Additionally, in 2023— the year Google launched Gemini—the company produced 14.3 million tons of carbon dioxide equivalent, which is the same as the emissions produced by 3.3 million gasoline-powered cars
6 Sundar Pichai & Demis Hassabis, supra note 1.
8 Robby Stein, Expanding AI Overviews and Introducing AI Mode, GOOGLE: THE KEYNOTE (Mar. 5, 2025), https://blog.google/products/search/ai- mode-search/.
14 See GOOGLE, 2024 ENVIRONMENTAL REPORT (July 2024),
https://www.gstatic.com/gumdrop/sustainability/google-2024-environmental- report.pdf.
136 Colo. Env’t L.J. Vol. 37: Issue 1
driven in one year.16 This was the highest emission of greenhouse gases the company had seen in the last five years.17 The company noted that this increase was mainly due to the increase in data center energy consumption.18 Additionally, Google explained that further integration of AI into its products may threaten its goals to reduce emissions “due to the increasing energy demands from the greater intensity of AI comput[ing], and the emissions associated with the expected increases” in the company’s technical infrastructure development.19
While Google works to contract with companies to purchase carbon removal credits, it does so knowing its demand for energy will likely continue to increase.20 Despite Google’s hopes to meet its goal of net-zero emissions by 2030, recent AI development and integration may threaten the company’s ability to do so.21
AI is emerging at a time when countries are making strides toward reducing carbon emissions and achieving net-zero goals.22 As electric vehicles rise in popularity and states implement greenhouse gas reduction statutes, the world takes one more step closer to addressing the global impact of greenhouse gas emissions.
The development and integration of AI into everyday life will likely enable individuals to solve complex problems in a matter of seconds. AI shows great promise in helping the environment through mapping deforestation, estimating carbon levels in forests, determining the rate at which icebergs are melting, and detecting the location of ocean waste.23 However,
16 Id.; Greenhouse Gas Equivalencies Calculator, EPA, https://www.epa.gov/energy/greenhouse-gas-equivalencies-calculator#results (last updated Nov. 2024).
22 For a Livable Climate: Net-zero Commitments Must be Backed by Credible Actions, UNITED NATIONS, https://www.un.org/en/climatechange/net- zero- coalition#:~:text=As%20of%20June%202024%2C%20107,a%20high%2Dlevel
%20government%20official (last visited Apr. 5, 2025).
23 Victoria Masterson, 9 Ways AI is Helping Tackle Climate Change,
2026 The Environmental Cost of Intelligence 137
AI also shows great promise in harming the environment.24 As Google’s implementation of AI highlights, AI is energy hungry and demands vast amounts of power to function properly.25 This energy is often provided through the use of non-renewable sources which produce greenhouse gases and contribute to climate change.26 Implementation of AI also requires the use of data centers, which need water and land to properly function.27 Accordingly, the unregulated expansion of AI use poses a significant risk of further environmental degradation.28
This Note will focus on the environmental impact of AI and the possible regulatory solutions that would prevent further environmental degradation as the AI industry continues to grow. Part I of this Note will discuss the expected impact of AI on the environment. It will discuss in depth AI’s effects on energy, land and water usage, grid reliability, and environmental equity. In addition, Part I will provide a background of the current environmental regulatory regime of the United States. This includes a discussion of the Clean Air Act, Clean Water Act, National Environmental Policy Act, Inflation Reduction Act, and the current state laws that are directed towards lowering carbon emissions. With this background, Part II will provide suggestions on how to approach and mitigate the environmental effects of AI before they become an even larger problem in the United States. These suggestions include (1) developing technologies and requirements for emission transparency in the industry; (2) implementing the newest and best technologies for efficiency of AI technology; (3) further developing renewable energy power sources; (4) implementing careful planning of AI development across regions; (5) creating AI specific federal regulations targeting the
WORLD ECONOMIC FORUM (Feb. 12, 2024),
https://www.weforum.org/stories/2024/02/ai-combat-climate-change/.
24 David Berreby, As Use of A.I. Soars, So Does the Energy and Water It Requires, YALE ENV’T 360 (Feb. 6, 2024),
https://e360.yale.edu/features/artificial-intelligence-climate-energy-emissions.
138 Colo. Env’t L.J. Vol. 37: Issue 1
environmental impacts of the technology; and (6) further developing state regulation of greenhouse gas emissions.
As AI expansion is inevitable, it is crucial for both federal and state governments to develop specific legislation to mitigate the technology’s potential emissions, water usage, and environmental inequities.
Background
Artificial Intelligence refers to “the ability of a digital
computer or computer
controlled robot to perform tasks commonly associated with intelligent beings.”29 In recent years, AI development and use grew at an unprecedented rate.30 This growth, in part, is due to the vast amount of data available, which is necessary to train AI models.31 AI models are trained on large datasets where they learn to “identify patterns, relationships and trends.”32 Once trained, these models can analyze new data and make predictions based on prior datasets.33
There are four types of AI models: (1) machine learning; (2) supervised learning; (3) unsupervised learning; and (4) deep learning.34 All four types rely on data to “learn” patterns, enabling AI to determine the probability of various outcomes.35 The different models simply deploy different methods of training
29 B.J. Copeland, Artificial intelligence, BRITANNICA.COM https://www.britannica.com/technology/artificial-intelligence (last visited Sep. 5, 2025).
30 AI has an environmental problem. Here’s what the world can do about that, UNITED NATIONS ENV’T PROGRAMME (Sep. 21, 2024),
https://www.unep.org/news-and-stories/story/ai-has-environmental-problem- heres-what-world-can-do- about#:~:text=The%20proliferating%20data%20centres%20that,which%20are
%20often%20mined%20unsustainably.
32 Artificial intelligence: What it is, How it Works and Why it Matters, INT’L ORG. FOR STANDARDIZATION, https://www.iso.org/artificial- intelligence#toc1 (last visited Apr. 5, 2025).
34 David Brault, What Are the Different Types of AI Models?, MENDIX (Aug. 14, 2025), https://www.mendix.com/blog/what-are-the-different-types-of- ai-models/.
2026 The Environmental Cost of Intelligence 139
the AI, using data sets.36 For example, unsupervised learning models structure data analysis on the information’s similarities, differences, and patterns, but do not require data scientists because the model is designed to work without instructions.37 In contrast, deep learning models identify complicated patterns in text, images, and sounds.38
Many popular AI models are large language models (“LLMs”), which are trained on extensive datasets using machine learning.39 Essentially, LLMs process vast amounts of data, enabling them to recognize and interpret language as well as other complex information.40
With AI programs becoming widely available to individual consumers, the use of AI continues to grow. The AI market is expected to grow at a compound annual rate of 31.5% from 2025 to 2033.41 AI is now being deployed for use in advertising and media, law, agriculture, and healthcare.42 However, this rapid expansion comes at a cost. The increasing demand for AI requires vast computational power, contributing to significant energy consumption and environmental degradation.
The rising popularity of AI significantly increases energy consumption, which threatens to increase (1) carbon emissions;
(2) grid instability; (3) land development; (4) water use; (5) indirect waste pollution; and (6) environmental inequity.
Carbon Emissions
Data centers, physical locations that house technological infrastructure, typically provide the energy and resources AI needs to run efficiently.43 Data centers are responsible for
39 What is a Large Language Model (LLM)?, CLOUDFLARE, https://www.cloudflare.com/learning/ai/what-is-large-language-model/ (last visited Apr. 5, 2025).
41 Artificial Intelligence Market Size, Share & Trends Analysis Report, SOL. TECH., https://www.grandviewresearch.com/industry-analysis/artificial- intelligence-ai-market (last visited Apr. 5, 2025).
43 Stephanie Susnjara & Ian Smalley, What is a data center?, INT’L BUS.
140 Colo. Env’t L.J. Vol. 37: Issue 1
processing large amounts of information.44 Every time a data point changes, a small amount of electricity is consumed and produces heat as a byproduct.45 Because data centers process vast changes in data points, the technology is prone to overheating. Thus, the data center must be kept cool using air conditioning or other cooling methods.46 Most of the energy used by data centers goes to powering air conditioning units.47
AI’s reliance on data centers is expected to drive up electricity consumption in both the United States and globally.48 Data centers worldwide could require around fourteen gigawatts of additional energy capacity by 2030.49 In the United States, the increased use of AI is projected to double data centers’ electricity needs by 2030.50 This would cause more than nine percent of America’s energy usage to stem from data centers.51
In addition to a skyrocketing energy demand, data centers are primarily powered by fossil fuels, which threaten the environment. 52 The burning of fossil fuels leads to emission of greenhouse gases into the atmosphere. 53 As data centers emit greenhouse gases into the atmosphere, they trap radiation from
MACH. (Sep. 4, 2024), https://www.ibm.com/think/topics/data-centers; Can We Mitigate AI’s Environmental Impacts?, YALE SCH. OF THE ENV’T (Oct. 10, 2024), https://environment.yale.edu/news/article/can-we-mitigate-ais-
environmental-impacts.
44 Josh Mahan, Understanding Data Center Energy Consumption, C&C TECH., GRP., https://cc-techgroup.com/data-center-energy-consumption/ (last updated June 8, 2023).
48 Taiba Jafari et al., Projecting the Electricity Demand Growth of Generative AI Large Language Models in the US, CTR. ON GLOB. ENERGY POL’Y AT COLUMBIA UNIV. SCH. OF INT’L AND PUB. AFFS. (July 17, 2024),
https://www.energypolicy.columbia.edu/projecting-the-electricity-demand- growth-of-generative-ai-large-language-models-in-the-us/; Brian Martucci, Data Center, AI load growth could threaten grid reliability: Conference Board, UTIL. DIVE (June 20, 2024), https://www.utilitydive.com/news/data-center-ai- load-growth-grid-reliability-conference-board/719380/.
49 JAFARI ET AL., supra note 48.
52 YALE SCH. OF THE ENV’T, supra note 43.
2026 The Environmental Cost of Intelligence 141
the sun, warming the planet’s surface at an unnatural rate.54 Accordingly, as AI requires more energy development, the use of fossil fuels grows as well. These emissions are an alarming byproduct of AI expansion.
AI’s actual increase in carbon emissions is hard to predict.55 This uncertainty is, in part, due to the different AI models, which require different computing power.56 For example, the training of OpenAI’s LLM, GPT-3, produced about 500 tons of carbon dioxide.57 A study from the University of Massachusetts Amherst found that training a single LLM can emit over 313 tons of carbon dioxide.58 In contrast, smaller, focused AI models take less time to train, consuming less energy overall.59 For instance, the BLOOM model, a similar sized LLM to OpenAI’s GPT-3, generated only 30 tons of carbon dioxide during its training process.60
Uncertainty carries into the implementation of the AI models after training. When an AI models a request, or query, it uses energy to provide a response.61 While there is limited understanding of the emissions from a single query, some estimates found it to be four to five times higher than that of a search engine query, raising concerns as AI usage expands.62
In addition to the carbon emissions from training models and queries, AI development is likely to require more data
54 What are greenhouse gases and how do they affect the climate?, U.S. ENERGY INFO. ADMIN. (MAY 2, 2024),
https://www.eia.gov/tools/faqs/faq.php?id=81&t=11.
55 Jude Coleman, AI’s Climate Impact Goes beyond Its Emissions, SCI. AM. (Dec. 7, 2023), https://www.scientificamerican.com/article/ais-climate-impact- goes-beyond-its-emissions/.
57 Id.
58 Renée Cho, AI’s Growing Carbon Footprint, COLUMBIA CLIMATE SCH.: STATE OF THE PLANET (June 9, 2023),
https://news.climate.columbia.edu/2023/06/09/ais-growing-carbon-footprint/.
60 Kate Saenko, A Computer Scientist Breaks Down Generative AI’s Hefty Carbon Footprint, SCI. AM. (May 25, 2023), https://www.scientificamerican.com/article/a-computer-scientist-breaks-down- generative-ais-hefty-carbon-footprint/.
142 Colo. Env’t L.J. Vol. 37: Issue 1
centers, and in turn, produce more emissions.63 Data center power demand is expected to see an increase of 160 percent by 2030.64 If natural gas were used to meet sixty percent of this demand, it would lead to increased emission of 215–220 million tons of carbon dioxide worldwide.65 If companies choose to rely on natural gas powered data centers, the carbon emissions worldwide could rise at an alarming rate to meet power demands.66
Further developing AI threatens to prevent countries, states, and individual companies from reaching carbon neutrality.67 For example, Microsoft, which vowed to be carbon- neutral by 2030, saw its carbon emissions grow in recent years due to advancements in AI.68 Moreover, utilities have already seen an up-tick in carbon emissions. Umatilla Electric Cooperative’s carbon emissions quadrupled since Amazon data centers were built in its service area. 69 As more companies and countries strive to allow AI development and meet new energy needs, carbon emissions will likely grow despite previous promises to decrease their presence in the atmosphere.70
Grid Instability
As the need for energy grows with the use of AI, concerns about the stability of the power grid are brought to the forefront of the conversation. The increasing power demand leaves traditional power grids struggling to function properly, which can cause blackouts and unreliability.71 Data centers, which
63 See Is nuclear energy the answer to AI data centers’ power consumption?, GOLDMAN SACHS (Jan. 23, 2025), https://www.goldmansachs.com/insights/articles/is-nuclear-energy-the- answer-to-ai-data-centers-power-consumption.
65 Id.
66 Eleni Kemene et al, AI and energy: Will AI help reduce emissions or increase power demand? Here’s what to know, WORLD ECON.F. (July 22, 2024), https://www.weforum.org/stories/2024/07/generative-ai-energy-emissions/.
67 See MARTUCCI, supra note 48.
71 AI Data Centers and Grid Instability: A Growing Concern, COLORADO’S RELIABLE ELEC. (Jan. 30, 2025), https://co-reliableelectric.com/ai-data-centers- and-grid-instability-a-growing-concern/.
2026 The Environmental Cost of Intelligence 143
require immense amounts of power, produce a large strain on the grid simply through their operation and maintenance.72 For example, the Microsoft data center in Kenosha, Wisconsin, consumes “more power than the entire residential population of Milwaukee.”73 Electric grids not designed for such large loads may face outages and struggle to meet increasing demand. AI’s demand for more power going to data centers may only exacerbate this issue.74
Power outages may not be a cause for concern with the current energy demand. However, if AI demands as much power as projected, the grid will have to find a way of meeting a possible 160% increase while supplying energy to its current users as well.75 Thus, the further growth of AI may undermine the current stability of the electricity grid.
Land Usage
Building more data centers in the United States to facilitate AI usage also poses concerns regarding land development. While smaller data centers can be between 5,000 and 10,000 square feet, larger data centers often significantly exceed this size range.76 This size difference is due to the need to ensure the processing equipment does not overheat.77
A requirement for more land may threaten previously undeveloped or agricultural areas.78 For example, in Prince William County, Virginia, two technology companies are attempting to purchase 2,100 acres of rural land to build 23 million square feet of data centers.79 This land was previously used for agriculture and maintained a scenic value to the locals
74 Id.
75 GOLDMAN SACHS, supra note 63.
76 Josh Mahan, Data Center Sizing Essentials: Your Guide to Efficient Infrastructure Planning, C&C TECH. GRP., https://cc-techgroup.com/data- center-sizing/ (last updated Feb. 12, 2024).
78 Issie Lapowsky, The Coming Tsunami of AI Data Centers, THE COAL.
TO PROTECT PRINCE WILLIAM CNTY. (Dec. 21, 2023, 5:54 AM EST),
https://protectpwc.org/2023/12/21/business-insider-inside-ais-giant-land- grab/.
144 Colo. Env’t L.J. Vol. 37: Issue 1
in the area.80 Similarly, a Denver-based technology company recently purchased 2,100 acres of land in Arizona.81 The company plans to create one of America’s largest data centers in existence.82 While just ten years ago, companies purchased fifteen to twenty acres for data centers, they are now purchasing thousands of acres.83 The rapid expansion of data centers not only consumes vast amounts of land but also threatens agricultural and recreational spaces, raising questions about sustainable infrastructure planning.
Water Availability
Data centers require large amounts of water, impacting the availability of water in areas where they are sited.84 To dissipate the heat produced by servers, data centers often use cooling towers or outside air.85 Both methods need water.86 Cooling towers use water evaporation to chill the water while the outside air method requires water in dry or hot outdoor conditions.87Additionally, generating energy to run data centers requires water.88 Nuclear reactors require water for cooling purposes, and hydroelectric power leads to expedited water evaporation.89
Technology companies are already seeing an increase in water usage with the rise of AI.90 In 2022, Google’s data centers, which host generative AI systems, used twenty percent more water than the previous year.91 In the same year, Microsoft’s
81 Han Lung, Arizona Land Deal Shows AI-Fueled Demand for Data Centers, CRE DAILY (Aug. 28, 2024), https://www.credaily.com/briefs/arizona- land-deal-shows-ai-fueled-demand-for-data-centers/.
84 See Shaolei Ren, How much water does AI consume? The public deserves to know, OECD (Nov. 30, 2023), https://oecd.ai/en/wonk/how-much- water-does-ai-consume.
2026 The Environmental Cost of Intelligence 145
servers, which host an LLM, increased by thirty-four percent.92 Additionally, AI’s global annual water consumption is projected to be between 4.2 billion and 6.6 billion cubic meters by 2027.93 This is more than four times the annual water use of Denmark.94
The water used for data centers is typically treated with chemicals that render it unusable for human consumption or agriculture.95 Further, data centers lose large portions of their water to evaporation.96 Accordingly, while these data centers require and use vast amounts of water, they do not provide for an opportunity to recycle the water for other purposes afterwards.97 As water scarcity continues to grow in many regions of the world, AI threatens to exacerbate this issue by removing available drinking water from the water cycle altogether.98
Indirect Environmental Concerns
Data centers rely on graphics processing units (“GPUs”) whose production and disposal have significant environmental consequences. GPUs are the hardware used for parallel processing in data centers.99 They can assist with training AI technology while optimizing and operating complex algorithms.100 Using GPUs aids in speeding up computer
93 Ana Pinheiro Privette, AI’s Challenging Waters, UNIV. OF ILL. URBANA- CHAMPAIGN CEE MAG., Winter 2024, at 15, 16,
https://ws.engr.illinois.edu/sitemanager/getfile.asp?id=8033.
95 Eric Olson, Anne Grau & Taylor Tipton, Data centers draining resources in water stressed communities, UNIV. OF TULSA: NEWS (July 19, 2024), https://utulsa.edu/news/data-centers-draining-resources-in-water-stressed- communities/.
96 Miguel Yañez-Barnuevo, Data Centers and Water Consumption, ENV’T AND ENERGY STUDY INST.: ARTS. (June 25, 2025),
https://www.eesi.org/articles/view/data-centers-and-water-consumption.
97 Ben Murray & Mia DiFelice, Artificial Intelligence: Big Tech’s Big Threat to Our Water and Climate, FOOD AND WATER WATCH (Apr. 9, 2025), https://www.foodandwaterwatch.org/2025/04/09/artificial-intelligence- water-climate/.
99 Why Data Center GPUs Are Essential to Innovation, INTEL, https://www.intel.com/content/www/us/en/products/docs/discrete-gpus/data- center-gpu/what-is-data-center-gpu.html (last visited Apr. 5, 2025).
146 Colo. Env’t L.J. Vol. 37: Issue 1
processing, which is important when training AI models.101 GPUs require raw materials like tungsten, palladium, cobalt, and tantalum.102 Mining and extracting these materials from the environment presents concerns about ethical and environmentally friendly mining operations.103 Dirty mining techniques and releases of toxic chemicals present issues of further environmental degradation.104 Collection of raw materials used to make GPUs present environmental concerns which is particularly alarming as 3.85 million GPUs were shipped to data centers by the top three producers in 2023 alone.105
Further, the life span of a GPU is only about 3–5 years, assuming the chip is maintained properly, meaning that GPU use creates large amounts of waste.106 Once a GPU is no longer functioning, it can be recycled, incinerated, or sent to a landfill.107 However, recycling is unlikely as only about twenty percent of “e-waste” is recycled.108 A majority of the waste ends up in landfills where the metals have the opportunity to leach into the ground and potentially reach and pollute groundwater.109 Thus, while indirectly, AI development threatens to increase waste of technology and further pollute the environment.
Environmental Equity
A final concern is AI’s threat to environmental equity.110 Burning fossil fuels produces local air pollution, thermal
102 Stephanie Kirmer, Environmental Implications of the AI Boom, TOWARDS DATA SCI. (May 2, 2024),
https://towardsdatascience.com/environmental-implications-of-the-ai-boom- 279300a24184/.
104 Adam Zewe, Explained: Generative AI’s environmental impact, MIT NEWS (Jan. 17, 2025), https://news.mit.edu/2025/explained-generative-ai- environmental-impact-0117.
110 Environmental equity is defined as “achieving fairness and balance in
2026 The Environmental Cost of Intelligence 147
pollution in nearby bodies of water, and sometimes hazardous waste.111 Additionally, data centers can strain freshwater supplies in areas already facing water scarcity.112 AI’s negative environmental impacts disproportionately affect communities that are already vulnerable to environmental harms.113 For example, Google’s data center in Finland operates on ninety- seven percent renewable energy, while its data centers in Asia rely primarily on fossil fuel generated energy.114 Such congestion can raise energy costs, burdening local residents.115
Although the full extent of AI’s impact on environmental inequity remains uncertain, it is a crucial issue in assessing AI’s environmental footprint. Environmental inequity currently exists around the world. In the United States, for example, African Americans “are exposed to 21% more pollution even though they produce 23% less pollution than the average” group.116 In Houston, Texas, a neighborhood with primarily Non-White is located within one mile of twenty-one industrial and toxic waste facilities.117 Google’s planned data centers in areas where water is scarce, like Chile and Uruguay, highlight AI’s threat to furthering the already existing environmental inequity around the world. 118 Without proactive regulatory oversight, AI’s environmental footprint will continue to grow, disproportionately harming vulnerable communities and further
access to environmental resources … in bearing environmental burdens … and in participating in environmental decision-making.” Equity and Environmental Justice at EDF: Vision and Principles, ENV’T DEF. FUND, https://www.edf.org/about/equity-and-environmental-justice-edf (last updated May 18, 2021).
114 Id.
116 Sophie Dulberg, Black and Latino residents live in more pollution than they cause. This is clearer in Houston more than anywhere else. TEXAS HOUSERS (Mar. 21, 2019), https://texashousers.org/2019/03/21/study-black-latino- pollution-consumption-exposure/.
117 Garrett Sansom et al., Confirming the Environmental Concerns of Community Members Utilizing Participatory-Based Research in the Houston Neighborhood of Manchester, Int’l J. of Env’t Rsch. and Pub. Health, Aug. 23, 2016, at 1, 2.
148 Colo. Env’t L.J. Vol. 37: Issue 1
straining global resources. Addressing these challenges requires urgent policy intervention and sustainable AI development practices.
Analysis
Without a comprehensive legislative framework at both the federal and state levels, AI’s rapid growth will significantly accelerate environmental degradation. The current environmental protections do not adequately address and mitigate areas which AI will likely affect, including greenhouse gas emissions and environmental inequity. Additionally, existing laws may inhibit the development of renewable energy systems that could support AI’s growth in a more sustainable manner. Addressing the pressing issues that AI presents will require developing industry-wide transparency, implementing energy-efficient technologies, encouraging and developing renewable energy sources quickly, carefully planning new data center development, and creating new federal and state laws targeting carbon emissions and AI’s specific environmental impacts.
Problems with the Current Environmental Regulatory Scheme
Although the current regulatory framework addresses some environmental concerns, it is insufficient to mitigate AI’s future environmental impacts. While there are regulations in place regarding emissions of pollutants into the environment, discharge of pollutants into water, and impact studies for development of certain projects, it is not enough. This Note provides a brief description of five different statutes at play and their potential to lessen AI’s negative impacts on the environment.
The Clean Air Act (“CAA”) regulates the emission of pollutants into the atmosphere.119 The CAA’s purpose is to prevent pollution at the federal, state, and local levels of
119 Summary of the Clean Air Act, EPA (July 25, 2025), https://www.epa.gov/laws-regulations/summary-clean-air-act.
2026 The Environmental Cost of Intelligence 149
government.120 However, regulation of carbon emissions for stationary pollution sources is rather weak. In Utility Air Regulatory Group v. EPA, the court held that requiring permits for stationary sources based only on greenhouse gas emissions is “incompatible” with the CAA.121 Only when the source is over the pollution limit for a National Ambient Air Quality Standard (“NAAQS”) pollutant can the Environmental Protection Agency (“EPA”) also regulate the source’s greenhouse gas emissions, including carbon dioxide.122
Data centers are subject to CAA regulations regarding the non-greenhouse gas pollutants that they produce, such as sulfur dioxide, nitrogen dioxide, and particulate matter.123 Emissions of NAAQS pollutants would require a permit from the EPA and could potentially trigger technology requirements for the data centers to ensure that the centers implement the best available control technology to control pollution of NAAQS pollutants and greenhouse gases.124 While regulation of these pollutants is crucial to preventing further air pollution, this regulation does not always address the biggest pollutant produced by data centers: greenhouse gases. The EPA can regulate greenhouse gas emissions when data centers qualify as a major emitting facility under the CAA for NAAQS air pollutants.125 Thus, some data centers are regulated under the CAA while others, which do not meet the threshold to be considered a major emitting facility, are not. Accordingly, regulating data centers and generation sources based solely off carbon emissions is currently not a viable option.
Although some data centers may be subject to greenhouse gas emissions regulations, it is important as AI develops to
120 42 U.S.C. § 7401(c) (2018).
121 Util. Air Regul. Grp. v. EPA, 573 U.S. 302, 322 (2014).
122 Id. at 334 (“Specifically, the Agency may not treat greenhouse gases as a pollutant for purposes of defining a ‘major emitting facility’ (or a ‘modification’ thereof) in the PSD context or a ‘major source’ in the Title V context.).
123 NAAQS Table, EPA (July 31, 2025), https://www.epa.gov/criteria-air- pollutants/naaqs-table.
124Clean Air Act (CAA) and Federal Facilities, EPA, https://www.epa.gov/enforcement/clean-air-act-caa-and-federal-facilities (last updated May 6, 2025).
150 Colo. Env’t L.J. Vol. 37: Issue 1
ensure that all the data centers are subject to regulation. Without proper control of the greenhouse gas emissions that data centers produce, global climate change could continue to worsen. Accordingly, current enforcement of the CAA is not enough to address AI’s greenhouse gas emissions.
The Clean Water Act (“CWA”) focuses on regulating the discharge of pollutants into “waters of the United States.”126 The CWA’s goal is to eliminate the discharge of pollutants into navigable waters by 1985, which the EPA is still working to achieve today.127 Data centers must follow the CWA if they plan to dredge and fill wetlands, produce stormwater runoff, or discharge water into municipal water districts or waters of the United States.128
However, data centers’ impact on water scarcity does not face regulation. While the CWA is crucial for reducing pollution in waterways, it does not regulate water availability. The EPA does not currently regulate water scarcity or water allotments at the local and state levels.129 Rather, the EPA provides drought assistance through identification, coordination, and collaboration in threatened communities.130 The EPA attempts to help areas with water scarcity address supply concerns and prevent future water crises.131 Thus, federal environmental regulations do not currently address drought and water scarcity issues, which AI development will likely worsen.
The National Environmental Policy Act (“NEPA”) requires federal agencies to perform environmental assessments prior to making decisions on permit applications, federal land management, and the development of highways and other public
126 33 U.S.C. § 1251(a)(1) (2025).
128 Clean Water Act (CWA) Compliance Monitoring, EPA (Sep. 2, 2025), https://www.epa.gov/compliance/clean-water-act-cwa-compliance-monitoring.
129 See Drought and Water Scarcity Initiatives, EPA (Mar. 18, 2025), https://www.epa.gov/climate-change-water-sector/drought-and-water-scarcity- initiatives.
2026 The Environmental Cost of Intelligence 151
facilities.132 NEPA does not require federal agencies to make decisions based on the environmental assessment, nor does it require federal agencies to choose the most environmentally friendly alternative.133 In the energy sector, projects like fossil fuel development, power plant construction, and siting transmission lines are all subject to NEPA review.134 Such projects are typically subject to NEPA review because they are sited on federal land or affect waters of the United States.135
NEPA review poses a fair number of challenges to developing new generation sources, whether renewable or not.136 For example, wind and solar projects attempting to gain rights of ways on land managed by the Bureau of Land Management face unpredictable timelines regarding the NEPA process.137 Because NEPA requires agencies to determine the environmental impact of a federal project, this causes delays in development of new energy projects.138 As the need for energy has rapidly increased in recent years, federal agencies have a backlog of applications for siting, financing, construction and operation of new energy-related projects.139
Although not directly targeted at regulating greenhouse gas emissions, the Inflation Reduction Act plays a role in reducing carbon emissions .140 The Act offers funding and incentives to
132 What is the National Environmental Policy Act?, EPA, (Apr. 11, 2025) https://www.epa.gov/nepa/what-national-environmental-policy- act#:~:text=The%20National%20Environmental%20Policy%20Act%20(NEPA
)%20was%20signed%20into%20law,actions%20prior%20to%20making%20dec isions.
133 COUNCIL ON ENV’T QUALITY, A CITIZEN’S GUIDE TO THE NEPA 5 (Dec.
2007) https://ceq.doe.gov/docs/get-involved/citizens_guide_dec07.pdf.
134 DAVID LAZERWITZ & MATTHEW BOSTICK, NEPA PROCESSES FOR ENERGY PROJECTS: UNIQUE CHALLENGES AND NEW DIRECTIONS, Nat’l Envt’l Pol’y Act 11-
1 (Rocky Mt. Min. L. Fdn. Paper 2010).
140 Summary of Inflation Reduction Act Provisions Related to Renewable Energy, EPA, (July 29, 2025), https://www.epa.gov/green-power- markets/summary-inflation-reduction-act-provisions-related-renewable- energy.
152 Colo. Env’t L.J. Vol. 37: Issue 1
individuals and companies implementing renewable energy to encourage a faster transition in the United States.141 Taxpayers are able to obtain investment tax credits or production tax credits to offset the cost of installing renewable energy systems.142 An additional bonus credit can be earned if specific environmental justice criteria are met in an effort to reach disadvantaged communities.143 In early 2025, the Clean Energy Production Tax Credit and the Clean Electricity Investment Tax Credit replaced previous tax incentives.144 Such credits allow further implementation of renewable energy as there is no technology specific requirement associated with them.145 The Act provides a way to incentivize and encourage development and implementation of renewable energy into the grid without forcing such implementation. Encouraging data centers built to support AI to run on renewable energy with monetary incentives could provide a path to preventing further AI emissions.
While federal regulation plays an important role in environmental protection, states have begun passing laws placing stricter requirements on pollution sources within their borders. For example, some states chose to adopt greenhouse gas emission targets.146 Twenty-two states developed targets for reducing greenhouse gas emissions or reaching net zero emissions by a specific year. 147 Some states, like Colorado, have enacted statutory targets to reduce greenhouse gas emissions by 100 percent by 2050.148 Other states, like New Mexico, have succeeded in setting greenhouse gas emission limits through executive orders.149 Although not all states currently have such
146 U.S. State Greenhouse Gas Emissions Targets, CTR. FOR CLIMATE AND
ENERGY SOLS. (Aug. 2025), https://www.c2es.org/document/greenhouse-gas- emissions-targets/.
148 Id.; see Colo. Rev. Stat. § 25-7-102 (2025).
149 CTR. FOR CLIMATE AND ENERGY SOLS., supra note 146.
2026 The Environmental Cost of Intelligence 153
goals in place, state officials are clearly taking the reduction and regulation of carbon emissions seriously.
Further, twelve states “price” carbon through mandatory cap-and-trade programs.150 Cap-and-trade programs create an emissions “cap” and set emissions reduction goals.151 Emission sources are then able to “trade” their pollution allowances with other polluters.152 This system enables one entity to reduce its emissions while allowing another to exceed its limit, balancing overall emissions.
Along with other state initiatives to reduce carbon emissions, thirty states have implemented renewable portfolio standards.153 Portfolio standards require that a specified percentage of electricity sold is produced from renewable resources.154 The goal is to encourage diversification of energy resources and promote clean energy.155 Most states with renewable portfolio standards have set goals of reaching at least forty percent renewable energy within a set time frame.156 Such diversification of energy sources seeks to help states reach and maintain their emissions goals through slow conversion of the electric grid. State-level emissions regulations demonstrate a promising manner of addressing climate change at a local level. As these reduction plans become more popular among the states, they may help to address AI’s environmental impact by forcing companies to power their data centers with renewable energy.
150 U.S. State Carbon Pricing Policies, CTR. FOR CLIMATE AND ENERGY SOLS (Jan. 2025), https://www.c2es.org/document/us-state-carbon-pricing- policies/.
151 What is Emissions Trading?, EPA (Nov. 22, 2024), https://www.epa.gov/emissions-trading/what-emissions-trading.
153 State Renewable Portfolio Standards and Goals, NCSL (Aug. 13, 2021), https://www.ncsl.org/energy/state-renewable-portfolio-standards-and- goals#:~:text=Thirty%20states%2C%20Washington%2C%20D.C.%2C,set%20 voluntary%20renewable%20energy%20goals.
154 Colo. Env’t L.J. Vol. 37: Issue 1
Possible Policy Solutions to AI Environmental
Threats
The first step in mitigating AI’s environmental impact is ensuring transparency from companies regarding the emissions produced by AI technology. A significant challenge in assessing AI’s environmental impact is the lack of data on emissions and water usage.157 Greater transparency and reporting from AI companies would enable regulators to make more informed policy decisions.158
Technologies that can efficiently measure emissions from AI technology could also provide for more transparency in the industry. 159 Without proper measuring devices, companies are not entirely able to disclose emissions data because it is not readily available. Further, standardizing the reading of such data to allow for application across all different hardware platforms would provide companies with the capability to meet transparency needs.160
The first step in mitigating AI’s environmental impacts is to implement legislation that develops the at environmental impact. The Senate introduced the Artificial Intelligence Environmental Impacts Act of 2024 to accomplish just that.161 Although the bill has not been reintroduced in the new Congress, the goal of the bill can still point the way towards development of new legislation. For example, the bill proposed a scheme for voluntary reporting of AI emissions from companies within the United States.162 Taking this scheme one step further, companies should be required to disclose the environmental
157 OCED, Measuring the Environmental Impacts of Artificial Intelligence Compute and Applications, 20 (2022) https://www.oecd.org/content/dam/oecd/en/publications/reports/2022/11/meas uring-the-environmental-impacts-of-artificial-intelligence-compute-and- applications_3dddded5/7babf571-en.pdf?utm.com.
159 Kylie Foy, New Tools are Available to Help Reduce the Energy that AI Models Devour, MIT NEWS (Oct. 5, 2023), https://news.mit.edu/2023/new-tools- available-reduce-energy-that-ai-models-devour-1005.
161 S. 3732, 118th Cong. (2024).
2026 The Environmental Cost of Intelligence 155
impacts of their AI for the government to properly understand the full weight of the potential environmental threat AI poses. Required disclosures would force full transparency and provide a baseline knowledge for further regulation.
Implementing energy-efficient technologies could help decrease data centers’ carbon emissions. Advocates for “Green AI” suggest that implementing optimized and efficient AI models can help reduce AI’s environmental footprint.163 A recent study showed that implementing the A100 GPU (an extremely quick GPU developed by NVIDIA164) reduced AI models’ carbon emissions by eighty-three percent compared to other processing technologies.165 A100 GPU also decreased the training time required by 62.2 percent.166 Moreover, one model trained on the A100 GPU emitted less carbon dioxide and performed better than models trained using different technology.167 The study supports the conclusion that employing lighter AI models and GPUs with faster processing can significantly lower AI’s carbon emissions.168
Using GPUs that cap the power used for processing can reduce energy consumption by twelve to fifteen percent depending on the model being trained.169 Implementing GPUs that cap power usage would likely increase the time required to train an AI model, but it would allow control over the amount of energy used in the process.170 This may also lead to a reduction in the cooling requirements for data centers.171 When GPU caps were introduced to supercomputers, the supercomputers were
163 See VIVIAN LIU & YIQIAO YIN, Green AI: Exploring Carbon Footprints, Mitigation Strategies, and Trade Offs in Large Language Model Training, 4 DISCOVER ARTIFICIAL INTELLIGENCE 49 (2024).
164 NVIDIA A100 TENSOR CORE GPU, NVIDIA 1 (2021)
https://www.nvidia.com/content/dam/en-zz/Solutions/Data- Center/a100/pdf/nvidia-a100-datasheet-us-nvidia-1758950-r4-web.pdf.
165 LIU & YIQIAO, supra note 163.
156 Colo. Env’t L.J. Vol. 37: Issue 1
around thirty degrees cooler than before, which reduced stress on the data center’s cooling system.172 Reducing cooling system stress could lessen AI’s environmental impact because cooling often requires either energy or water consumption.
Optimizing what applications get to run, where they are run, and when they run can help AI companies save energy.173 As a starting point, moving AI energy consumption from mid- afternoon to early morning could reduce the energy required for the workload.174 Another promising method is implementing training performance estimates.175 This process reduces energy consumption by evaluating a training hyperparameter configuration’s performance against the performance of other models before training is complete.176 Thus, the program does not have to run for the full time it takes to train the model with non-optimal parameters, but rather can be stopped early if the preemptive evaluation proves to be non-optimal.177
Regarding AI’s energy consumption after training, deployment of GPU portioning and mixing low- and high-quality AI model variants may reduce AI’s post-training energy consumption.178 GPU partitioning shares the GPU among multiple workloads.179 Mixing AI models allows for the best model variant to be utilized at the time the program runs.180 An AI program by the name of CLOVER demonstrated that implementing both methods together can reduce emissions by seventy-five percent.181 Running a system with both low- and high-quality AI model variants would allow for optimal energy
173 Reducing AI’s Climate Impact: Everything You Always Wanted to Know but Were Afraid to Ask, UNIV. CAL. BERKELEY BEGIN (Sept. 13, 2024) https://begin.berkeley.edu/reducing-ais-climate-impact-everything-you- always-wanted-to-know-but-were-afraid-to-ask/.
174 Id.
179 NVIDIA MULTI-INSTANCE GPU AND NVIDIA VIRTUAL COMPUTE
SERVER, NVIDIA 2 (2020), https://www.nvidia.com/content/dam/en- zz/Solutions/design-visualization/solutions/resources/documents1/Technical- Brief-Multi-Instance-GPU-NVIDIA-Virtual-Compute-Server.pdf.
180 UNIV. CAL. BERKELEY BEGIN, supra note 173.
2026 The Environmental Cost of Intelligence 157
efficiency as smaller models, which require less energy, could perform routine queries while larger models would be deployed only when necessary. Accordingly, this would save energy compared to using a singular LLM model to perform all queries despite their level of complexity.
Reducing the complexity of the mathematical operations AI computes may also reduce AI’s footprint.182 Decreasing the size and complexity of the numbers, matrices, and networks AI uses achieves computational savings without a significant loss in accuracy.183 By using less computational power, AI uses less energy overall, lowering the carbon emissions.184
Another technology-based method for reducing carbon emissions is carbon capture and storage (“CCS”). CCS is the process of collecting carbon dioxide generated through burning fossil fuels before the carbon dioxide is emitted into the atmosphere.185 The carbon dioxide is then stored, often deep underground.186 Implementing CCS at generation plants that supply energy to AI data centers could significantly reduce AI’s carbon footprint.
Implementing efficient cooling technologies could further reduce AI’s environmental impact, particularly its water usage. For example, free air-cooling methods use minimal amounts of water; however, they are only efficient in cool climates.187 Accordingly, placing data centers in climates that are cool for a significant portion of the year could reduce the water needed for AI. Additionally, submersion cooling, where the data center’s hardware is placed in dielectric fluid which transfers heat without conducting electricity, can eliminate the need for regular air conditioning.188 Locating data centers in cooler climates could significantly reduce their water consumption.189
185 Howard Herzog, Carbon Capture, MIT CLIMATE PORTAL (Jan. 20, 2023), https://climate.mit.edu/explainers/carbon-capture.
187 UNIV. CAL. BERKELEY BEGIN, supra note 173.
158 Colo. Env’t L.J. Vol. 37: Issue 1
As the analysis above demonstrates, AI companies can run data centers more efficiently. However, the federal government needs to act to either encourage or force companies to continue to implement the most efficient technology available to lessen AI’s environmental impact. To do so, the federal government should develop legislation that requires companies to adopt specific technology available at the time of construction. Developing such legislation would force companies to adopt current best practices while allowing flexibility of efficiency to continue to grow and be implemented in the future. Without such legislation that focuses on forcing energy efficiency, companies decide what technology to implement and may fail to choose the most efficient technology due to cost or resource limitations.
Implementing renewable energy is yet another tool that could significantly reduce AI’s environmental impact. As AI- driven data centers proliferate, transitioning them to renewable energy sources would significantly reduce carbon emissions. Deploying rapidly scalable energy sources like solar and wind can create cleaner, cost-competitive data centers.190 Developments using clean energy are necessary to meet future carbon-emission goals.191 Some companies are already turning toward development of renewable energy data centers. For example, Microsoft recently announced plans to re-open Three Mile Island, a nuclear power plant in Pennsylvania.192 The company will use the energy generated to serve their data centers in response to growing AI energy needs.193 By using solar, wind, and geothermal energy to power data centers around the country, AI’s carbon emissions are reduced. Additionally, the use of nuclear energy provides a carbon-free
190 Clean Energy Resources to Meet Data Center Electricity Demand, U.S. DEP’T OF ENERGY, https://www.energy.gov/gdo/clean-energy-resources-meet- data-center-electricity-demand (last visited Apr. 6, 2025).
192 Max Hauptman, Microsoft Announces Plan to Reopen Three Mile Island Nuclear Power Plant to Support AI, USA TODAY, Sept. 20, 2024. https://www.usatoday.com/story/money/energy/2024/09/20/three-mile-island- nuclear-plant-constellation-microsoft-deal/75307770007/.
2026 The Environmental Cost of Intelligence 159
energy source which can help reduce “the amount of generation capacity, storage, and transmission needed to ensure grid reliability.”194 Development of renewable and nuclear energy- based data centers also sets a precedent for other companies and encourages further fossil-fuel free projects.
Incentives and regulatory changes which target the increase in nuclear energy development could reduce the emissions of AI as well. Nuclear energy, for example, comes with expensive construction and operation costs.195 Additionally, such development projects are met with licensing and regulation requirements which can be difficult to overcome.196 For instance, nuclear projects are required to conduct an environmental impact statement while other non-nuclear projects may opt for a less intensive environmental assessment.197 The Nuclear Regulatory Commission currently charges annual fees to license holders along with hourly fees for the review of applications and other oversight measures.198 The current nuclear permitting and licensing framework discourages significant industry expansion. Reforming regulations to facilitate nuclear energy expansion while safeguarding environmental and public health could significantly reduce AI’s carbon footprint.
Furthering renewable development may require slashing some of the permitting challenges for the approval of such
194 5 Ways the U.S. Nuclear Energy Industry is Evolving in 2024, U.S. DEP’T OF ENERGY, OFF. OF NUCLEAR ENERGY (Sept. 30, 2024),
https://www.energy.gov/ne/articles/5-ways-us-nuclear-energy-industry- evolving- 2024#:~:text=Nuclear%20Complements%20Renewable%20Energy%20Source s&text=Nuclear%20energy%20can%20provide%20clean,needed%20to%20ens ure%20grid%20reliability.
195 Advantages and Challenges of Nuclear Energy, U.S. DEP’T OF ENERGY, OFF. OF NUCLEAR ENERGY (June 11, 2024),
https://www.energy.gov/ne/articles/advantages-and-challenges-nuclear- energy.
196 JOHN JACOBS ET. AL., LICENSING AND PERMITTING REFORMS TO ACCELERATE NUCLEAR ENERGY DEPLOYMENT, 1 (Jan. 2024),
https://bipartisanpolicy.org/download/?file=/wp- content/uploads/2024/01/BPC_Nuclear-Permit-License-Reform-Issue- Brief.pdf.
160 Colo. Env’t L.J. Vol. 37: Issue 1
projects.199 There are a few steps the federal and state governments could take to increase the efficiency of renewable energy permitting and construction.200 First, the government could streamline agency review and encourage interagency coordination by implementing tools that allow the different agencies involved in the permitting process to view the review process status.201 Consolidating environmental review into a single process without repetitive steps would lessen the time agencies need to approve projects.202 Moreover, without proper funding and support to meet the demand of the current applications, the permitting process slows.203 Thus, providing additional funding to the agencies responsible for NEPA review could assist with streamlining the environmental impact studies and permitting.204 It is critical to streamline NEPA while maintaining a thorough environmental review process.
Encouraging renewable energy development is key to transitioning to a more sustainable grid. However, focusing on only the development and permitting of such projects overlooks a crucial issue: transmission. Transmission is required to provide electricity to the end user. Currently, the country faces a steep obstacle to developing the current transmission system. Projections indicate that the United States must expand its electricity transmission capacity by sixty percent by 2030 to meet growing clean energy demands.205 There are currently many clean energy proposals that are gridlocked due to wait
199 Pablo Alvarez, Speeding Up the Green Transition: Proposed Permitting Reforms for Faster Renewable Energy Development, COLUM. CLIMATE L. BLOG (Aug. 8, 2024),
https://blogs.law.columbia.edu/climatechange/2024/08/08/speeding-up-the- green-transition-proposed-permitting-reforms-for-faster-renewable-energy- development/.
200 Id.
203 Nicole Pavia, Beyond NEPA: Understanding the complexities of slow infrastructure buildout, CLEAN AIR TASK FORCE (Aug. 1, 2024), https://www.catf.us/2024/08/beyond-nepa-understanding-complexities-slow- infrastructure-buildout/.
205 Queued Up… But in Need of Transmission, U.S. DEP’T OF ENERGY, OFF. OF POL’Y, https://www.energy.gov/policy/queued-need-transmission (last visited Apr. 6, 2025).
2026 The Environmental Cost of Intelligence 161
time and costs of connecting transmission lines.206 Implementing efficient planning and processing methods to grant interregional transmission projects will be crucial to producing cleaner energy at a higher rate.207 Additionally, if state and federal governments work together, they can lessen gridlock and provide for quicker development of transmission.208 While transitioning to renewable energy is essential, ensuring equitable distribution of AI’s environmental impact require careful planning of data center locations and workloads.
To incentivize new clean energy developments, the federal government should continue to deploy an incentives scheme and tailor additional legislation specifically to data centers. Encouraging big AI companies to run on renewable–energy– powered data centers would help ensure the development of this energy for AI, which would ultimately reduce the overall impact of AI related emissions. Thus, incentives that provide tax breaks or credits to companies taking advantage of the renewable energy could help prevent further environmental degradation. Additionally, streamlining parts of the NEPA process to allow for interagency communication during each step of the environmental impact studies would allow for a smoother and quicker permitting process.
Geographical load balancing—distributing AI workloads across data centers strategically—can promote environmental equity.209 Utilizing a load balancing system that is “equity- aware,” meaning that it considers the carbon and water footprints on geographical areas where data centers are located when distributing the workload among such data centers, may produce significant advancements in preventing environmental
207 Lori Bird et al., US Clean Power Development Sees Record Progress, As Well As Stronger Headwinds, WORLD RES. INST. (Feb. 21, 2025), https://www.wri.org/insights/clean-energy-progress-united-states.
209 See PENGFI LI ET AL., TOWARDS ENVIRONMENTALLY EQUITABLE AI VIA
Geographical Load Balancing, e-Energy (2024),
https://dl.acm.org/doi/pdf/10.1145/3632775.3661938.
162 Colo. Env’t L.J. Vol. 37: Issue 1
inequity caused by AI.210 While current research points to the need for an algorithm which will effectively achieve this goal, this seems to be a promising approach to considering and mitigating environmental inequity in the most affected areas.211
Strategically siting data centers in regions with ample water resources can prevent AI from exacerbating water scarcity in drought-prone areas.212 Rather than locating data centers in developing regions with limited water resources, siting them in more water-secure areas ensures a more equitable distribution of AI’s environmental impact.213 This allows AI’s water footprint to be balanced across regions and prevent further burdening already threatened areas.214 Considering the placement of future data centers across America is crucial to balancing the environmental impact amongst all communities rather than those already susceptible to environmental degradation.
Developing data centers in politically stable regions can also help mitigate environmental inequity. Such placement would prevent already burdened areas from facing further threats exacerbated by a need to produce more energy to meet the data center’s needs and to provide substantial amounts of water (which may have to come from available drinking water). Passing legislation incentivizing data center siting in environmentally stable regions—particularly those with secure water supplies—could help mitigate environmental inequities both domestically and globally.
States should encourage community engagement, develop a requirement for equity assessments, and monitor air quality and health in all areas of the state when developing energy and data center projects. This could assist with environmental equity.215 By allowing community engagement when approving new plans for data centers or permits for fossil fuel power plants, the individuals most affected are given a voice and a platform to be
212 PRIVETTE, supra note 93, at 16.
215 Serena Li et al., How US States Can Lead on Carbon Removal Policy, WORLD RES. INST. (Oct. 31, 2024), https://www.wri.org/technical- perspectives/us-state-carbon-dioxide-removal-policies.
2026 The Environmental Cost of Intelligence 163
heard. This allows for interaction and development that takes into consideration environmental impacts and the impact on citizens’ health and welfare. Accordingly, it may allow underprivileged areas to gain the power to address their concerns and mitigate some of the environmental inequity they face. By integrating environmental equity considerations and encouraging community engagement into AI’s expansion strategy, policymakers can ensure that the benefits of AI innovation do not come at the cost of exacerbating existing environmental injustices.
While adapting provisions from existing environmental laws offers a starting point, AI’s unique energy consumption patterns may require a tailored regulatory approach that addresses its specific carbon, water, and land-use impacts. For example, as the CAA requires new pollution sources to employ certain technologies to improve efficiency and control of said pollution, something similar could be deployed for AI.216 As discussed above, AI-related technology is continuously improving in energy efficiency.217 Creating a regulatory scheme which allows for enforcement of best available control technologies, such as those energy efficient technologies currently available, could potentially solve some problems that AI currently presents.
Despite limited concrete data on AI’s environmental impact, a flexible regulatory framework that can adapt without statutory amendment is essential to address ongoing knowledge gaps. Forcing AI companies to implement available pollution control technologies for greenhouse gas emissions would not overburden the administration. It would provide a legal mechanism to protect the population from AI’s large carbon footprint with the limited data available. Such legislation would require future data centers to employ energy efficient technology which may yet to be developed. This type of legislation allows for
216See Clean Air Act (CAA) Compliance Monitoring, EPA, https://www.epa.gov/compliance/clean-air-act-caa-compliance-monitoring (last updated Sep. 23, 2025).
217 UNIV. CAL. BERKELEY BEGIN, supra note 173.
164 Colo. Env’t L.J. Vol. 37: Issue 1
a more dynamic solution that improves with time and continues to adapt to a changing industry.
Another way to mitigate emissions in relation to generation of electricity needed for AI would be to create incentives for using clean or renewable energy. Tax incentives, like those provided by the Inflation Reduction Act have seen tremendous results in targeting greenhouse gas emissions.218 Millions of families claimed tax credits for residential clean energy and home efficiency under the Inflation Reduction Act.219 Expanding tax incentives for data centers and AI firms to adopt clean energy could mitigate AI’s environmental impact before it significantly affects the climate.
To maintain grid stability and ensure the electricity grid’s growth, governments and utilities will need to work together to develop clean energy and modernize the current grid.220 To bring the grid into this century, deployment of cutting-edge technologies, equipment, and controls that can work to provide more reliable and efficient electricity is crucial.221 Grid modernization would allow for more available energy, a reduction in peak energy loads, and potentially lower operational costs.222 Legislating grid modernizations and mandating advanced technology adoption could enhance the stability and efficiency of the United States’ power grid.
It is important to note that although federal legislation could provide a promising path to mitigating the environmental impacts of the AI industry, environmental protection is currently a controversial topic in the United States. While environmental protection is an important concern for many
218 How Tax Credits are Driving Clean Energy Growth Two Years into Inflation Reduction Act, DEP’T OF ENERGY, OFF. OF POL’Y (Aug. 16, 2024), https://www.energy.gov/policy/articles/how-tax-credits-are-driving-clean- energy-growth-two-years-inflation-reduction-act.
220 Manav Mittal, From Code to Current: How to Keep AI Data Centers in Check for a Sustainable Grid, UTIL. DIVE (Jan. 3, 2025), https://www.utilitydive.com/news/ai-data-centers-sustainable-renewable- energy-demand- response/736387/#:~:text=By%20focusing%20on%20energy%20efficiency,abo ut%20balancing%20progress%20with%20sustainability.
2026 The Environmental Cost of Intelligence 165
voters, “partisanship continues to be a major factor in attitudes about the environment and climate change.”223 Since 2017, increasing environmental concern primarily came from Democratic voters.224 This ideological divide is further highlighted by the fact that Republican and Republican-leaning independents do not prioritize global climate change as much as Democrats and Democrat-leaning independents.225
The current attempts to shrink the administrative state could hinder further development of federal environmental protection as well.226 For example, the recent effort to remove diversity, equity and inclusion programs from the federal government resulted in the Environmental Protection Agency placing over 160 environmental justice workers on leave.227 The Department of Justice has seen similar cuts to their environmental protection sector.228 Recent judicial decisions, such as Loper Bright Enterprises v. Raimondo, may limit administrative agencies’ authority to enact environmental regulations.229 As agencies, like the Environmental Protection Agency may no longer be able to rely on a “reasonable” or “permissible” view of a statute, legal challenges to new regulations could be more successful. Accordingly, while legislation and regulation on the federal level will be important to protect against AI’s environmental impacts, the current political climate may hamper the development of new statutes and agency rules. The new developments in administrative law
223 AS ECONOMIC CONCERNS REcede, ENVIRONMENTAL PROTECTION RISES ON THE PUBLIC’S POLICY AGENDA, PEW RESEARCH CTR. 5 (2020),
https://www.pewresearch.org/politics/wp- content/uploads/sites/4/2020/02/PP_2020.02.13_Political- Priorities_FINAL.pdf.
225 Id. at 8.
226 Valerie Volcovici et al., Trump administration cuts environmental justice programs at EPA, DOJ, REUTERS (Feb. 6, 2025, 4:39 PM), https://www.reuters.com/world/us/trump-administration-cuts-environmental- justice-programs-epa-doj-sources-say-2025-02-06/.
229 Environmental Law Implications of Loper Bright and the End of Chevron Deference, SIDLEY (July 2, 2024), https://environmentalenergybrief.sidley.com/2024/07/02/environmental-law- implications-of-loper-bright-and-the-end-of-chevron-deference/.
166 Colo. Env’t L.J. Vol. 37: Issue 1
may serve to protect against the dismantling of current environmental regulations but may also serve to prevent new agency rules and regulations.
While action at the federal level could assist with protecting the environment from AI’s impact, state initiatives could be a source of assistance as well. As discussed above, many states have already implemented statutes or executive orders aimed at lowering carbon emissions.230 However, meeting these goals is harder than setting them. California, which pledged to lower emissions by forty percent of their 1990 levels by 2030, is not currently on track to meet their target.231 In order to meet their 2030 target, California would need to increase their emissions cutting from 1.5% each year to around 4.4% each year.232
The effectiveness of greenhouse gas reduction efforts depends on the policy mechanism used to implement them.233 While climate action plans are popular policies deployed at the state level, they have no significant effect on lowering the planet’s carbon emissions.234 Planet-level emissions can be reduced more effectively through setting caps for greenhouse gas emissions.235 Effective emissions policies establish specific, measurable goals.236 State-led carbon removal initiatives offer promising opportunities to further reduce emissions. State action can be beneficial for lowering carbon emissions and thus
230 CTR. FOR CLIMATE AND ENERGY SOLS., supra note 146.
231 Melody Petersen, California unlikely to meet landmark goals for reducing greenhouse gas emissions, L.A. TIMES (Mar. 16, 2024, 3:00 AM), https://www.latimes.com/environment/story/2024-03-16/california-behind-on- goals-for-reducing-greenhouse-gases; California Green Innovation Index 2023, NEXT 10 (2023), https://greeninnovationindex.org/2023-edition/.
233 Don Grant et al., Effectiveness of U.S. state policies to reduce CO2 emissions from power plants: Research brief, THE JOURNALIST’S RES. (Oct. 15, 2014), https://journalistsresource.org/environment/effectiveness-u-s-state- policies-powerplant-c02-emissions/.
2026 The Environmental Cost of Intelligence 167
lowering the impact of AI on the environment. However, such policies must have clear directives on how to reach those goals.
State-led carbon removal initiatives offer promising opportunities to further reduce emissions. Such plans that attempt to go “beyond net-zero,” are seeking to scale up carbon removal while also integrating state action into broad climate plans.237 New York introduced legislation to make carbon reduction goals binding.238 The state proposed a plan to begin purchasing carbon removal credits and double the amount purchased each year.239 While existing state plans are making headway, further developing such plans can mitigate carbon emission impacts even more as the rate of emissions may grow in the technology sector.240 States without existing carbon reduction goals must develop comprehensive plans to ensure national progress, rather than limiting impact to isolated regions.241 By strengthening both federal and state regulatory frameworks, the United States can help ensure that AI’s rapid expansion aligns with sustainability goals rather than exacerbating environmental challenges.
Conclusion
AI’s rise is still in the early stages. While this technology offers innovation and efficiency, it also presents significant environmental challenges. AI has the potential to increase carbon emissions, burden water scarce areas with further water strains, threaten grid stability, require more mining of precious metals, and lead to further electronic waste due to its GPU requirements.
The current environmental regulatory regime primarily addresses pollutants other than greenhouse gases. The Clean Air Act does not allow the EPA to cap carbon emissions unless there are other pollutants being released into the air at rates above current limits. Expanding state-level greenhouse gas caps and setting clear, enforceable reduction targets would allow AI
237 Li et al., supra note 215.
168 Colo. Env’t L.J. Vol. 37: Issue 1
to grow sustainably while incentivizing industry-driven environmental solutions. While state-level policies can drive emission reduction and grid improvements, a cohesive federal strategy is necessary to regulate AI’s broader environmental footprint.
The federal government could potentially pass legislation that allows for regulation of greenhouse gas emissions, reducing their continued release into the environment. This would require targeted legislation addressing emissions from AI operations, data centers, and fossil fuel power plants. As the Supreme Court held that regulating greenhouse gas emissions for all stationary sources would be an administrative burden, it is likely the legislation would have to be tailored narrowly. However, even narrowly tailored legislation could allow the government to enforce efficient technology requirements, among other things, which could significantly lessen the carbon emissions AI threatens to create.
Water concerns in the AI space mostly relate to the amount of water needed for cooling data center equipment. Both the federal and state government can work cooperatively to encourage new data centers to be built in areas where drought is not present. Governments could use incentives or legislation to restrict data center permits in regions experiencing severe water scarcity. In addition to preventing further stress on areas already in peril, this can help balance the environmental load the country must take on to supply more data for AI’s continued development.
Continuing to provide incentives for renewable energy development will assist in further modernizing our grid. However, state governments must collaborate to approve the transmission infrastructure necessary for integrating renewable projects. Without allowance for connection, new energy developments are ready to go with nothing to do.
To encourage development of base-load power, the federal government should reform regulatory barriers for nuclear energy development. By creating new legislation which focuses on allowing projects to be approved in shorter amounts of time but still focuses on the safety of such projects, nuclear energy
2026 The Environmental Cost of Intelligence 169
could provide consistent energy to the grid without an extreme environmental burden.
Providing federal incentives in the form of tax credits, or otherwise, to encourage development of data centers and power plants in the United States can help encourage environmental equity. These incentives would reduce data center development in foreign countries by companies who primarily work in the United States. However, such development must be thoughtful and done in areas that are not already disproportionately affected by industrial development and its environmental impacts.
AI has the potential to set the United States back in terms of its efforts to protect the environment. However, being able to address these concerns now with a critical eye on the full effect of them is key. Creating legislation that can be dynamic and adaptable to new obstacles would provide the most environmental protection as the industry expands. This includes legislation that focuses on efficient technology requirements, requires community cooperation when approving new data centers and power plant projects, and focuses on modernizing the grid to prevent future instability through addition of renewable and nuclear energy. Taking these steps would help mitigate the current issues AI is imposing on the environment and those that could be discovered in ten or fifteen years. A proactive regulatory approach will ensure that AI’s rapid expansion aligns with environmental sustainability, rather than exacerbating existing ecological challenges. By implementing adaptable policies today, policymakers can safeguard the environment while fostering responsible AI innovation.