Chapter 31: The Environmental Cost of AI

Opening: The Weight of Intelligence

Training GPT-3 — the large language model released by OpenAI in 2020 that set off the generative AI era — emitted approximately 552 metric tons of CO2 equivalent. To put that in physical context: 552 metric tons of CO2 is roughly equivalent to driving a gasoline car around the circumference of Earth approximately 70 times, or to the lifetime carbon emissions of five average Americans. It represents the carbon output of approximately 120 US households for an entire year.

And GPT-3 was not especially large by the standards of what came next. GPT-4, released in 2023, is estimated — OpenAI has not published its exact carbon cost — to be substantially larger. The training of models like Google's PaLM 2, Meta's LLaMA, Anthropic's Claude, and Mistral's models, multiplied across dozens of companies training hundreds of models in the 2022–2025 generative AI boom, represents an aggregate carbon cost that remains largely unknown because the companies responsible for it have not disclosed the figures.

Then there is inference: not training the model once, but running it billions of times. A single ChatGPT query consumes approximately 10 times as much energy as a Google search. With hundreds of millions of users making multiple queries per day, the inference-time energy consumption of widely deployed large language models exceeds training costs over the model's lifetime by orders of magnitude.

AI is reshaping the world — reshaping how we write, how we code, how we search for information, how we make decisions in medicine and law and finance. It is also reshaping the climate. And because the climate is a global commons — where the emissions produced in data centers in Virginia, Iowa, and Dublin affect sea levels in Bangladesh, droughts in Sub-Saharan Africa, and floods in Pakistan — the environmental ethics of AI is not merely a corporate responsibility question. It is a question of global justice.

This chapter examines AI's environmental footprint across its dimensions: energy, carbon, water, hardware, and the supply chain that makes AI hardware possible. It examines who pays the costs and who captures the benefits. And it asks what responsible AI development and deployment would look like in a world where the climate is already a crisis.

Section 1: The Energy Appetite of AI

Training vs. Inference — Understanding the Two Energy Dimensions

AI's energy consumption has two fundamentally different components, and conflating them produces misleading analysis. Training — the process of teaching an AI model by iteratively adjusting its parameters on large datasets — is computationally intensive, happens once (or a few times), and represents the most dramatic and easily reportable single energy cost. Inference — running the trained model to answer questions, generate text, process images, or make predictions — happens billions of times daily and in aggregate represents the dominant long-term energy cost of widely deployed AI systems.

Training large language models at the frontier requires enormous computational resources: clusters of thousands of specialized AI accelerator chips (GPUs or TPUs) running for weeks or months. The H100 GPU that Nvidia introduced in 2022 — the chip most widely used for large model training as of 2024 — consumes approximately 700 watts at peak load. A training cluster of 10,000 such GPUs running for three months consumes roughly 1.5 billion watt-hours, or 1.5 gigawatt-hours — enough electricity to power approximately 1,400 average US homes for a year. This is for one training run; frontier AI labs run many training experiments, ablation studies, and model iterations to produce a single released model.

The energy cost of inference is harder to measure but arguably more important for long-term environmental assessment. Unlike training, which produces a model once and is done, inference happens every time a user queries an AI system. A million ChatGPT queries require a million forward passes through a model with hundreds of billions of parameters, each requiring matrix multiplications across massive parameter tensors. The energy per inference is small — estimated at 0.001–0.01 kWh per query, depending on model size and hardware efficiency — but at scale (hundreds of millions of daily queries), the cumulative energy consumption is substantial.

The Google Search vs. ChatGPT Comparison

The most widely cited comparison for communicating inference energy cost is the Google search baseline. A standard Google search query consumes approximately 0.0003 kWh. An AI-assisted search query — using a large language model to generate an answer rather than a ranked list of links — consumes approximately 10 times as much, around 0.003 kWh per query. As Google, Microsoft (Bing), Amazon, and other search providers integrate LLMs into their core search products, and as those products serve billions of queries daily, the aggregate energy cost of search increases substantially.

This comparison is useful but requires context. A Google search query that saves a user 30 minutes of research by immediately answering a question accurately has a very different energy-per-value ratio than a query that generates a plausible-sounding but incorrect response that the user must then verify through additional queries. The absolute energy cost per query matters; so does what that energy achieves. Evaluating AI's energy footprint requires thinking about the denominator (the value delivered) as well as the numerator (the energy consumed).

Cumulative Scale: Data Centers and the AI Build-Out

Beyond individual models, AI's energy demand operates at the level of the global data center industry. AI computing is concentrated in hyperscale data centers — facilities of 100,000 square feet or more, operating at megawatt to gigawatt scale, powered by dedicated electrical infrastructure, and in some cases representing the dominant local electricity consumer.

The International Energy Agency's 2024 assessment estimated that data center electricity consumption was already approximately 200–250 terawatt-hours (TWh) per year globally — roughly 1% of global electricity demand. The AI build-out is driving rapid expansion of this figure. Microsoft, Google, Amazon Web Services, Meta, and a cluster of new AI-specialized cloud providers have announced capital expenditure plans for data center expansion in the 2023–2027 period totaling hundreds of billions of dollars. These investments will translate into significant additional electricity demand. IEA projected that global data center electricity demand could double by 2026, reaching 400–500 TWh annually, with AI as the primary demand driver.

This electricity demand creates physical infrastructure requirements: power plants, transmission lines, substations, and grid connections at scales that strain regional electrical grids. Microsoft's agreement with Constellation Energy to restart the Three Mile Island nuclear facility — signed in 2023 — was driven directly by AI data center power demand. Amazon's similar agreements for nuclear power, and hyperscalers' competition for power purchase agreements with renewable energy developers, reflect the physical reality of AI's energy appetite.

Section 2: Carbon Emissions

The Carbon Cost of Computing

Electricity consumption translates to carbon emissions based on the carbon intensity of the grid supplying the electricity — how much CO2 is emitted per kilowatt-hour of electricity generated. This varies enormously by geography and time of day. Electricity generated from Norwegian hydropower or French nuclear power has very low carbon intensity. Electricity generated from coal-fired power plants in the US Midwest or India has very high carbon intensity. A data center in Washington state powered by Columbia River hydropower has dramatically lower carbon emissions per unit of computation than an identical facility in West Virginia powered by Appalachian coal.

The geographic arbitrage available in data center siting — locating facilities where land and power are cheap, rather than where power is cleanest — has historically been driven by cost optimization. Amazon Web Services, Google, and Microsoft each operate extensive renewable energy programs and have made carbon commitments, but their data centers are also sited in regions where renewable energy availability varies, where their actual carbon impact depends heavily on the time-of-day and season-of-year carbon intensity of the grid.

Documented Carbon Footprints

The carbon footprints of major AI models that have been documented or credibly estimated form a small but important evidence base:

BERT-Large (Google, 2018): Emma Strubell and colleagues' pioneering 2019 paper "Energy and Policy Considerations for Deep Learning in NLP" estimated that training BERT-Large produced approximately 1,438 lbs CO2 equivalent — relatively modest but a wake-up call for the ML research community. The paper was the first to make training carbon costs legible to the research community and sparked the "Green AI" conversation.

GPT-3 (OpenAI, 2020): Patterson et al. (2021) estimated the training carbon cost at approximately 552 metric tons CO2e, based on estimated computation (3.14 × 10²³ FLOPS), hardware specifications, and grid carbon intensity. OpenAI did not publish this figure itself.

Google's LaMDA (2021): Patterson et al. (2021) estimated LaMDA training at approximately 449 metric tons CO2e — comparable to GPT-3 and reflecting similar model scale.

The Measurement Problem: All of these estimates are approximations, because AI companies do not routinely disclose the precise energy consumption, hardware configuration, training duration, and grid location sufficient to calculate accurate carbon footprints. The research community is working from published or estimated FLOP counts, hardware specifications, and grid carbon intensity averages — not direct measurement. The true carbon costs of frontier AI model training are largely unknown to the public because the companies responsible for those costs choose not to disclose them.

Scope 3 Emissions and the Disclosure Gap

Corporate carbon accounting uses a three-scope framework: Scope 1 covers direct emissions from owned operations; Scope 2 covers indirect emissions from purchased electricity; Scope 3 covers indirect emissions throughout the value chain, including suppliers, customers, and the use of sold products. AI companies' carbon disclosures, where they exist, typically cover Scope 1 and 2. Scope 3 emissions — including the carbon embedded in hardware manufacturing, the emissions from customer inference queries, and the supply chain emissions of chip production — are rarely disclosed.

This creates a systematic understatement of AI's climate impact in corporate reporting. A company that buys renewable energy for its own data centers and reports low Scope 2 emissions has not accounted for the carbon cost of manufacturing the GPUs that data center runs; for the carbon in the steel, concrete, and copper in the facility itself; or for the energy consumed by its customers running inference queries on locally deployed model versions. The carbon footprint of the AI ecosystem is substantially larger than the Scope 1 and 2 reporting of its major participants suggests.

Section 3: Water Consumption

The Water Footprint of AI

Carbon emissions receive more public attention, but water consumption is an equally significant environmental dimension of AI — and in some geographies, more immediately consequential. Data centers require water for cooling: the servers and power infrastructure in a large data center generate enormous heat, which must be dissipated to prevent equipment failure. The dominant cooling technology is evaporative cooling — exposing hot air to water, which absorbs heat as it evaporates. A large data center can consume millions of gallons of water per day through evaporative cooling.

The relationship between AI computing intensity and water consumption was systematically analyzed in a 2023 study by Pengfei Li and colleagues titled "Making AI Less 'Thirsty': Uncovering and Addressing the Secret Water Footprint of AI Models." The study estimated that training GPT-3 consumed approximately 700,000 liters (185,000 gallons) of freshwater — enough to fill a large swimming pool. More strikingly, the study estimated that ChatGPT (inference) consumed approximately 500 milliliters (half a liter) of water per conversation — for a system with hundreds of millions of conversations daily, this represents thousands of tons of water per day.

These estimates depend on the water use effectiveness (WUE) of Microsoft's and OpenAI's data centers, the geographic distribution of their infrastructure, and assumptions about query volumes — all parameters that the companies have not fully disclosed. The Li et al. estimates are credible approximations based on available data but should be understood as approximations.

Geographic Impacts in Water-Stressed Regions

The water dimension of AI becomes a justice issue when data centers are sited in regions experiencing water stress. Several major AI data center clusters are located in areas that face current or projected water scarcity:

Arizona: Microsoft, Google, Meta, Apple, and other hyperscalers operate major data centers in the Phoenix metropolitan area, one of the fastest-growing and most water-stressed areas in the United States. The Colorado River, which supplies Arizona with the majority of its water, is overallocated — withdrawals from the river have exceeded its annual average flow for decades. The federal Bureau of Reclamation has imposed cutbacks on Colorado River allocations. Data centers competing for Arizona water against agricultural users and residential users represent a direct resource conflict in an acutely water-limited system.

Netherlands: The Netherlands hosts significant European cloud infrastructure, including Microsoft and Google facilities, in a country that has experienced growing water stress and has implemented temporary restrictions on data center expansion. Amsterdam's metropolitan area has imposed requirements for data centers to demonstrate that they are not using drinking-quality water for cooling and to meet water use efficiency standards.

Chile: A significant planned expansion of data center capacity in Chile — including major facilities by Google, Microsoft, and Amazon — has drawn opposition from water-stressed agricultural communities in the Coquimbo region, where water rights have been a source of social conflict. The Chilean government has faced pressure from communities arguing that data center water use would compete with agricultural water rights on which rural livelihoods depend.

Water Disclosure and Accountability

Unlike carbon emissions, which have established reporting frameworks (the GHG Protocol, CDP disclosure, Scope 1/2/3 categories), water consumption by data centers lacks standardized disclosure requirements. Individual companies publish water stewardship reports with varying levels of detail; many do not report at the facility level sufficient to assess local water impact. Independent analysis of data center water use depends on proxy indicators (facility locations, reported power usage effectiveness, cooling technology descriptions) rather than direct measurement.

Section 4: Hardware and the Supply Chain

The Material Basis of AI

AI is often described as a software phenomenon — as intelligence, as cognition, as language. This framing obscures the material reality: AI is built on physical hardware — chips, servers, cables, power units, cooling systems — that require specific materials extracted from the earth, processed through industrial supply chains, and eventually disposed of. The environmental ethics of AI cannot be complete without examining this hardware supply chain.

The AI chip supply chain begins with rare earth minerals and specialty metals. The H100 GPU that powers frontier AI training contains silicon, cobalt, tantalum, tungsten, and multiple other elements whose extraction is geographically concentrated and environmentally and socially consequential:

Cobalt: Approximately 70% of global cobalt production comes from the Democratic Republic of Congo, where mining conditions have been extensively documented as involving child labor, unsafe mining practices, and environmental contamination. Cobalt is used in lithium-ion batteries (including the batteries in AI server UPS systems) and in certain semiconductor applications. The DRC's cobalt supply chain extends through Chinese processing facilities to global electronics manufacturers.

Lithium: The lithium triangle of Chile, Bolivia, and Argentina contains the world's largest lithium reserves, in high-altitude salt flats (salares) where extraction requires pumping large volumes of brine that can affect local hydrology and the communities that depend on it. Lithium production for AI data center battery backup systems is a small fraction of total lithium demand (dominated by electric vehicle batteries), but it is part of the same supply chain with the same social and environmental issues.

Silicon and semiconductor fabrication: AI chips are fabricated at semiconductor foundries — primarily TSMC in Taiwan, Samsung in South Korea, and SMIC in China — using processes that require vast quantities of ultrapure water, specialty chemicals (many toxic), and electricity. Taiwan's semiconductor cluster is located in a water-stressed area; the Hsinchu Science Park's water use has been a subject of policy attention. The chemicals used in chip fabrication, if improperly managed, can contaminate local water supplies and soil.

E-Waste

The end-of-life dimension of AI hardware represents a growing environmental challenge. AI chip generations are advancing rapidly: the H100 GPU (2022) was followed by the H200 (2024) and subsequent generations with each new generation delivering meaningfully higher performance. As AI labs and cloud providers upgrade to the latest hardware for competitive performance, older hardware is retired — sometimes within three to five years of deployment. This rapid hardware turnover generates electronic waste (e-waste) containing toxic materials including lead, mercury, cadmium, and hexavalent chromium.

Global e-waste generation reached approximately 62 million metric tons in 2022 and is growing. The majority of global e-waste is inadequately documented and a significant fraction is exported to lower-income countries for informal recycling — where workers extract valuable metals through processes (burning, acid leaching) that release toxic substances. The AI industry's contribution to e-waste is not separately tracked, but the rapid hardware upgrade cycles driven by AI performance competition are a meaningful contributor.

Section 5: The Rebound Effect

Jevons' Paradox and AI Efficiency

The history of energy technology includes a recurring and counterintuitive pattern known as the Jevons paradox or the rebound effect: improvements in energy efficiency tend to increase rather than decrease total energy consumption, because lower cost per unit encourages greater total use. When steam engines became more fuel-efficient in the 19th century, coal consumption increased because more economical steam power enabled expanded industrial applications. When automobile fuel efficiency improved, total vehicle miles traveled increased. When computing became cheaper and more efficient, more computing was done.

AI deployment is reproducing this pattern in real time. As AI makes computation more efficient — allowing more tasks to be performed per dollar of computing — organizations respond by computing more. A company that previously could afford to analyze only a subset of customer service transcripts now uses AI to analyze all of them. A drug discovery program that previously screened thousands of compounds can now screen billions. A content platform that previously recommended based on limited behavioral signals can now incorporate vastly more data. Each of these applications represents genuinely valuable use of AI — and each consumes additional energy that would not have been consumed without the AI capability.

The rebound effect means that efficiency improvements in AI computation — which have been substantial, with performance per watt improving significantly with each hardware generation — do not necessarily translate into reduced environmental impact. They may translate into greater AI utilization that increases total energy and environmental cost even as the cost per unit of AI capability falls.

AI Making Itself Cheaper

A particularly notable instance of the rebound effect in AI is AI being used to optimize AI hardware and model development. Chip design AI (used by Intel, NVIDIA, and others to optimize chip layouts) reduces the cost of producing the next generation of AI chips. Model compression and distillation techniques (making smaller models that approximate larger ones) reduce the cost of inference. Neural architecture search (using AI to find more efficient model architectures) reduces training costs. Each of these efficiency improvements reduces the cost barrier to AI deployment — potentially enabling more total AI use.

Section 6: AI for Climate

The Positive Contribution

Having established AI's environmental costs, intellectual honesty requires examining AI's potential positive environmental contribution. The same computational capability that consumes energy can, in some applications, help address the climate crisis that its energy use contributes to.

Energy grid optimization: AI systems are being used to optimize electricity grid dispatch — predicting demand, managing variable renewable energy inputs, reducing transmission losses, and enabling more efficient integration of solar and wind generation. Google DeepMind's work with UK grid operator National Grid on neural network-based grid management demonstrated measurable reductions in balancing costs. AI-optimized smart grids could accelerate the transition to renewable energy by making variable renewables more manageable.

Climate modeling: Advanced AI is enabling higher-resolution, faster climate models that improve understanding of regional climate impacts, extreme weather event probabilities, and feedback loop dynamics. DeepMind's GraphCast weather prediction model achieved more accurate 10-day weather forecasts than traditional numerical methods while running 1,000 times faster — with implications for climate science and weather-dependent industries.

Materials discovery: AI is accelerating the discovery of new materials for solar cells, batteries, and other clean energy technologies. DeepMind's AlphaFold has shown what AI-accelerated scientific discovery can look like in protein structure prediction; analogous approaches are being applied to materials science for energy applications.

Agricultural efficiency: AI precision agriculture tools — crop monitoring via satellite and drone imagery, AI-optimized irrigation, disease and pest prediction — can reduce the environmental footprint of agriculture, which is a major source of greenhouse gas emissions and freshwater consumption globally.

The Tension

The tension between AI's climate costs and climate contributions is real and cannot be resolved by simply asserting that beneficial applications outweigh harmful ones. Whether any specific AI application produces net climate benefit depends on: how much energy it consumes, whether that energy is from low-carbon sources, what climate benefit the application produces, and whether the application would exist without AI or would instead be achieved through less energy-intensive means.

For many applications described as "AI for climate," this cost-benefit analysis has not been rigorously conducted. The field needs the same evidence rigor that characterizes the best climate policy analysis — quantified emissions reductions, measured against counterfactuals, with uncertainty ranges — rather than narrative arguments about transformative potential.

Section 7: Corporate Responsibility and Disclosure

The Gap Between Commitment and Disclosure

Major AI companies have made prominent voluntary carbon commitments:

Microsoft committed in 2020 to become carbon negative by 2030, removing all its historical emissions by 2050. It reports on Scope 1, 2, and some Scope 3 emissions annually.

Google committed to operating on carbon-free energy on a 24/7 basis by 2030 — meaning matching its electricity consumption to carbon-free sources in the same time and place, not just on an annual average basis. It has published detailed data center efficiency metrics.

Meta claims to have achieved net zero in its operations and 100% renewable energy since 2020.

Amazon (AWS) committed to 100% renewable energy by 2025 and net zero by 2040 through its Climate Pledge.

These commitments are meaningful but should be evaluated against several considerations:

Renewable energy claims rely on RECs: Many companies' "renewable energy" claims are achieved through purchasing Renewable Energy Certificates (RECs) — certificates representing that a MWh of renewable energy was generated somewhere on the grid, at some time, and credited to the purchaser. RECs do not guarantee that the electricity actually powering a data center at any given moment is carbon-free. The 24/7 carbon-free energy standard that Google has committed to is more stringent and more meaningful; REC-based claims are considerably less so.

Scope 3 emissions remain largely undisclosed: The carbon embedded in hardware manufacturing, supply chains, and customer use is not adequately captured in most AI company reporting. For hardware-intensive businesses like AI, Scope 3 emissions from chip manufacturing alone may be substantial.

Absolute emissions may be rising despite efficiency improvements: Companies can reduce carbon intensity (emissions per unit of output) while increasing absolute emissions if output is growing faster than efficiency improves. Rapid AI capacity expansion means absolute electricity consumption and absolute carbon emissions are growing even as per-unit efficiency improves.

Section 8: Regulatory Framework

EU AI Act and Sustainability Requirements

The EU AI Act, adopted in 2024, includes provisions addressing the environmental impact of AI systems. High-risk AI systems are required to document their computational resource consumption and energy use. The Act requires that AI systems be designed with energy efficiency in mind, though it does not specify quantitative energy limits. The European Commission is directed to develop measurement methodologies and assessment frameworks for AI's environmental impact — acknowledging that current methods are insufficient for meaningful disclosure.

The EU AI Act's environmental provisions represent a regulatory beginning — establishing the principle that AI systems' environmental impact should be measured and disclosed — but fall well short of comprehensive mandatory disclosure with standardized methodologies.

SEC Climate Disclosure Rules

The US Securities and Exchange Commission adopted climate disclosure rules in 2024 that require public companies to disclose their Scope 1 and 2 greenhouse gas emissions, and Scope 3 emissions where material or included in company targets. These rules were subject to legal challenges and implementation uncertainty, but if fully implemented would require AI companies to disclose emissions data that has been largely voluntary. The "materiality" assessment for Scope 3 creates an opportunity for tech companies to argue that supply chain and customer use emissions are not material to their business, limiting the disclosure requirement's reach.

EU Corporate Sustainability Reporting Directive

The EU Corporate Sustainability Reporting Directive (CSRD), effective from 2024 for large companies and phased in for smaller companies, requires detailed disclosure of a wide range of sustainability metrics including climate impact, water consumption, biodiversity impact, and supply chain sustainability. Unlike the SEC rules, CSRD requires disclosure on a "double materiality" basis — both how sustainability risks affect the company (financial materiality) and how the company's activities affect society and the environment (impact materiality). This is a significantly more demanding standard that would require AI companies to disclose their water consumption, supply chain carbon, and other environmental impacts that current voluntary reporting largely omits.

The "Right to Know"

A useful frame for evaluating disclosure obligations is the public's "right to know" about AI's environmental impact. When governments, researchers, and citizens cannot assess how much energy, water, and carbon AI systems consume — because the companies responsible for these systems choose not to disclose the figures — they cannot make informed decisions about whether the benefits of AI justify its environmental costs, which companies are making genuine progress on sustainability, or what regulatory interventions are needed.

The parallel to food labeling is instructive: consumers have a right to know what is in the food they eat; investors have a right to know the environmental liabilities of companies they own; citizens have a right to know the environmental impact of publicly beneficial and publicly harmful technologies. This principle supports mandatory, standardized, independently verified environmental disclosure from AI companies — not as a burden on innovation, but as accountability to a public that bears the environmental costs of a technology's development.

Section 9: Sustainable AI Research

The Green AI Movement

The "Green AI" research agenda, catalyzed by Strubell et al.'s 2019 paper and formalized by Roy Schwartz and colleagues' 2020 paper "Green AI" in Communications of the ACM, argues that AI research should measure and optimize for energy efficiency and computational cost alongside accuracy and performance. The paper observed that the computational and therefore energy requirements of state-of-the-art AI research had been growing exponentially (doubling approximately every few months at peak), while the research community rarely reported the computational cost of its results — creating incentives for ever-larger models without accountability for the environmental cost.

The Green AI agenda proposes: - Mandatory reporting of computational cost in research papers, alongside accuracy metrics - Community standards for model comparison that account for efficiency, not just performance - Research investment in model compression, knowledge distillation, and efficient architecture design - Evaluation frameworks that give credit for achieving similar performance at lower computational cost

Efficient ML Techniques

The Green AI movement has been accompanied by substantial technical progress in efficient ML:

Knowledge distillation: Training smaller "student" models to replicate the behavior of larger "teacher" models. DistilBERT, developed by Hugging Face, achieved 97% of BERT's performance with 40% fewer parameters and 60% faster inference — a significant efficiency gain. GPT-4 can in principle be distilled into smaller models that achieve most of its performance at a fraction of its inference energy cost.

Model compression: Quantization (representing model weights in fewer bits), pruning (removing less important parameters), and sparsification (making most weights zero) can reduce model size and inference cost substantially without proportionate performance degradation.

Efficient architecture design: EfficientNet (Google, 2019) demonstrated that systematic compound scaling of network dimensions (width, depth, resolution) could achieve state-of-the-art accuracy at dramatically lower computational cost than prior scaling approaches. Similar principles applied to language models have produced competitive models at lower scale.

Mixture of Experts (MoE): Architecture designs in which only a subset of the model's parameters are activated for any given input — routing each input to the most relevant "expert" subset — allow very large total model capacity with much lower inference cost per query.

The Green Software Foundation

The Green Software Foundation, established in 2021 with Microsoft, GitHub, Thoughtworks, and other technology companies as founding members, develops standards and tools for measuring and reducing the environmental impact of software. Its Software Carbon Intensity (SCI) specification provides a methodology for measuring the carbon intensity of software, analogous to miles-per-gallon for automobiles, enabling consistent comparison across systems.

The foundation's work represents industry recognition that software (including AI) has environmental impact that should be systematically measured and reduced — a significant shift from treating computation as environmentally costless.

Section 10: The Justice Dimension

Who Emits, Who Suffers

The deepest ethical dimension of AI's environmental impact is its justice dimension: the people and communities who contribute most to AI's carbon and water costs are not the same as those who bear its climate consequences most severely.

The global emissions geography: AI development and deployment is concentrated in the United States, Europe, China, and a small number of other wealthy countries. The data centers, the companies, the investors, and the consumers who benefit from AI are concentrated in the world's high-income economies. The carbon and water consumed by this infrastructure contributes to greenhouse gas concentrations and freshwater stress that are global.

The climate impact geography: The most severe near-term consequences of climate change — more frequent and intense storms, sea level rise, agricultural disruption, extreme heat, and displacement — fall disproportionately on communities in the Global South that are most exposed, have the least adaptive capacity, and have contributed the least to the emissions causing the crisis. A person living in coastal Bangladesh who faces displacement from sea level rise attributable in part to AI data center emissions has received no AI benefit and bears significant AI climate cost. This is a classic externality: those generating the harm bear its cost elsewhere.

Climate refugees: Climate change is already contributing to displacement — the Internal Displacement Monitoring Centre estimated that over 30 million people were displaced by weather-related disasters in 2021 alone. Projections of future climate displacement range from tens to hundreds of millions. The AI industry's contribution to the emissions causing climate displacement is not separately quantifiable, but it exists and is growing. Business ethics that grapples seriously with AI's environmental impact cannot ignore this nexus.

Environmental Justice in AI Deployment

Environmental justice concerns arise not only at the global level but locally. Data centers are frequently sited in communities with less political power to resist them, where land and power are cheap, and where local authorities may prioritize economic development over environmental protection. The water, noise, and electricity infrastructure impacts of large data centers fall on neighboring communities. The economic benefits — primarily construction jobs and tax revenue — are often more modest and time-limited than initial projections suggest.

Communities adjacent to data centers have organized against their environmental impacts in Virginia, Georgia, Ireland, and the Netherlands. In Virginia's Prince William County, residents and environmental advocates contested a planned Amazon data center development on land that includes historically significant battlefield sites and adjacent to communities that did not want the industrial infrastructure. In Ireland, where approximately 20% of the national electricity supply is consumed by data centers, concerns about power grid impact, renewable energy competition, and land use have generated significant political controversy.

What Responsibility Looks Like

What would AI companies taking their environmental justice responsibilities seriously actually do?

Mandatory, standardized disclosure: Publish energy consumption, carbon emissions (Scope 1, 2, and 3), and water consumption at the facility level, with consistent methodology enabling comparison and verification. Do not rely on REC-based renewable claims that obscure actual grid carbon intensity.

24/7 carbon-free energy: Commit to and achieve time-matching of electricity consumption to carbon-free sources, not just annual average matching. Invest in grid-scale storage and demand flexibility to make this viable.

Water stewardship in stressed regions: Commit to net-zero water consumption in water-stressed regions — returning to local watersheds what is withdrawn — and disclose facility-level water use. Where water stress is acute, reconsider facility siting or invest in alternative cooling technologies (air cooling, seawater cooling) that do not compete with local freshwater.

Supply chain accountability: Require suppliers of hardware to meet environmental and labor standards, including responsible mineral sourcing. Publish supply chain audits.

Community benefit: In communities hosting data centers, provide genuine benefit — meaningful employment, local energy access, infrastructure investment — proportional to the land, power, and water consumed.

Climate finance: Contribute to climate adaptation and mitigation finance proportional to the company's historical and ongoing emissions — not as greenwashing but as genuine accountability to the global communities bearing the environmental cost of the AI industry's growth.

Section 11: Biodiversity and Land Use — The Overlooked Dimension

Data Centers and Ecological Footprint

Carbon, energy, and water receive the most attention in discussions of AI's environmental impact, but data center infrastructure also has land use and biodiversity consequences that have received less scrutiny. The construction of hyperscale data centers consumes significant land — large facilities occupy 100 to 500 acres — and the associated infrastructure (power lines, roads, cooling water sources, staff facilities) extends the footprint further. When data center development occurs on or adjacent to ecologically significant land, the biodiversity impact can be substantial.

In Northern Virginia — the "data center alley" that is the most densely concentrated data center region in the world — rapid expansion of data center infrastructure has consumed agricultural and wooded land at a pace that has drawn criticism from environmental advocates and local governments. The Prince William County controversy of 2022–2024, in which Amazon proposed a data center campus on land including historically significant Civil War battlefield terrain and adjacent to wooded areas providing wildlife corridor connectivity, illustrated the land use tensions that data center expansion creates when it encounters ecologically or historically valued land.

In Ireland, where data centers have expanded dramatically to serve European cloud infrastructure, concerns about the environmental footprint extend beyond electricity demand (which is substantial — Irish data centers consumed approximately 20% of national electricity in 2023) to the land use impacts of large-scale industrial development in the Irish countryside. Water abstraction permits for data center cooling have drawn regulatory attention from the Environmental Protection Agency.

The Embodied Carbon in Data Center Construction

Beyond the energy and water consumed in data center operations, the construction of the data center itself — the concrete, steel, copper, aluminum, and engineered materials that constitute the physical facility — embeds substantial carbon. This "embodied carbon" in construction materials is estimated to represent 20–40% of a building's lifetime carbon footprint before it operates for a single day.

For hyperscale data centers — built at extraordinary scale with specialized materials for power distribution, cooling, and fire suppression — embodied carbon can be substantial and is not captured in operational energy reporting. Companies that report carbon-free operational energy may nonetheless have significant embodied carbon in their construction programs. As the AI build-out continues to drive new data center construction, the embodied carbon of that construction is a meaningful environmental cost that should be included in comprehensive environmental accounting.

Section 12: The Ethics of AI Environmental Decision-Making

Who Decides, Who Bears the Consequences

The environmental decisions associated with AI deployment are made by a relatively small number of large technology companies, informed by their own financial interests and subject to limited external accountability. The communities that bear the consequences — neighboring communities facing noise, traffic, and visual impact from data center construction; water users in data center regions facing resource competition; and the global population facing climate consequences of AI's carbon emissions — have minimal formal input into those decisions.

This governance gap mirrors broader patterns in environmental governance: those who generate externalities are typically better organized, better resourced, and better connected politically than those who bear them. The regulatory frameworks that address this imbalance — environmental impact assessment requirements, community consultation processes, carbon pricing mechanisms, water rights regulation — are unevenly applied to data center development, and the specific dimension of AI's environmental impact is even less well-addressed by regulation than data center impacts generally.

The Role of Business Leadership

For business professionals, the environmental ethics of AI is not merely an external regulatory compliance question — it is a leadership question. Organizations that deploy AI make choices that have environmental consequences. Those choices include: which AI services to use and from which providers; how much AI is appropriate for a given task versus less computationally intensive alternatives; how to account for AI's environmental cost in sustainability reporting; how to engage with suppliers and partners on environmental performance; and how to exercise voice as large customers of AI infrastructure companies.

Large enterprises are among the most significant customers of cloud AI services, and enterprise procurement decisions aggregate into market signals that influence AI company behavior. Organizations that require their AI vendors to disclose energy and carbon metrics, that prefer vendors with credible renewable energy programs, and that include AI environmental costs in their sustainability accounting create market incentives for better environmental performance that complement regulatory requirements.

The question is not whether environmental considerations should be part of AI deployment decisions — they already are, whether acknowledged or not. Every AI deployment has energy, carbon, and water consequences. The ethical question is whether those consequences are accounted for explicitly, with appropriate attribution of cost and responsibility, or whether they are allowed to remain invisible externalities that someone else — communities, future generations, the global climate system — bears without compensation.

The Principle of Environmental Proportionality

A useful organizing principle for the ethics of AI environmental decisions is proportionality: the environmental cost of AI deployment should be proportional to the value delivered, should be borne by those who capture the benefit, and should be disclosed transparently enough that proportionality can be assessed.

By this standard, AI applications that deliver high value — medical diagnosis that detects cancer earlier, materials discovery that enables better solar cells, climate modeling that improves policy — may justify higher environmental costs than AI applications that deliver marginal value — generating marketing copy of mediocre quality, producing visual content that replaces adequately-functional human creative work, automating customer service in ways that reduce service quality while cutting costs. This is not a call for AI rationing; it is a call for honest accounting that prevents AI's genuine high-value applications from subsidizing its low-value ones by obscuring the aggregate environmental cost.

For business professionals: when evaluating AI deployment, ask not just "what is the return on AI investment?" but "what is the environmental cost per unit of value delivered, and is that cost proportionate?" This is the same cost-benefit discipline that responsible organizations apply to other resource decisions. The absence of a market price for AI's environmental externalities does not make those externalities irrelevant to sound business decision-making; it makes the organization's own accounting more important.

The Trajectory and What It Requires

Looking forward, the environmental trajectory of AI depends on the interaction of several forces: the pace of AI adoption and the breadth of applications, the efficiency of AI hardware and model architectures, the carbon intensity of the electricity grid, the regulatory framework for environmental disclosure, and the market signals that shape AI company behavior.

On the trajectory question, the most important variable is electricity decarbonization. If the electricity supplying AI data centers decarbonizes rapidly — through the growth of renewable energy and storage, nuclear power, and grid modernization — AI's carbon footprint can decrease even as its energy consumption grows. The major AI companies have significant market power to accelerate this trajectory: their demand for renewable energy drives power purchase agreements that finance new renewable capacity, and their political engagement on energy policy can support the grid investment that decarbonization requires. The choice between using this market power to secure access to renewable energy and actually advancing grid decarbonization — through advocacy for grid-scale storage, transmission expansion, and market design that values carbon-free electricity — is a genuine choice with different environmental implications.

On the regulatory front, the trajectory toward mandatory environmental disclosure for AI systems is clear even if the timeline is uncertain. The EU CSRD, the SEC's climate disclosure rules, and the EU AI Act's environmental provisions collectively represent a regulatory direction that will require more comprehensive and standardized environmental reporting from AI companies. Organizations that develop measurement and reporting infrastructure now — before requirements are mandatory — will be better positioned to comply and to demonstrate leadership when the requirements take effect.

The fundamental ethical insight that this chapter's analysis supports is simple but consequential: there is no free computation. Every AI query, every model training run, every chip manufactured, every data center built has environmental consequences that are real even when they are invisible in current accounting. The ethical responsibility of AI developers, deployers, and users is to make those consequences visible, to account for them in decision-making, and to invest in reducing them to levels proportionate to the genuine human value that AI enables. That responsibility is not discharged by purchasing renewable energy certificates or announcing long-term carbon neutrality goals; it is discharged by the hard work of measurement, disclosure, investment in genuine efficiency, and accountability to the communities and global climate system that bear the costs of AI's environmental impact.

Summary

AI's environmental footprint is significant, growing, and inadequately disclosed. Training large models produces substantial carbon emissions. Operating them at scale produces water consumption measured in millions of gallons per day. The hardware supply chain involves materials extraction with serious environmental and human rights implications. The companies responsible for these costs have made voluntary commitments that are meaningful but insufficient — and have not been required to provide the disclosure necessary for those commitments to be independently verified.

The justice dimension is not peripheral: the communities bearing the costs of AI's environmental impact are not the communities capturing its benefits. Global carbon emissions fall on those least able to adapt; local water consumption falls on communities least able to challenge powerful corporate neighbors; hardware supply chains externalize costs to mining communities in the Global South.

The "green AI" technical agenda — efficient architectures, model compression, mandatory computational cost reporting — offers genuine promise but faces the rebound effect challenge: efficiency gains enable expanded deployment that may offset environmental gains. Technical solutions require complementary regulatory frameworks that create accountability for environmental disclosure, mandatory efficiency standards, and genuine pricing of carbon and water externalities.

Business professionals making AI deployment decisions should incorporate environmental costs into those decisions — not as a checkbox item in a sustainability report, but as a genuine factor in the cost-benefit analysis of AI investment. The energy, water, and carbon costs of AI are real costs, currently externalized to communities and the global climate system. Ethical AI deployment requires accounting for them.

This chapter concludes the section on AI's societal impacts. Chapter 32 examines the governance of AI at the organizational, national, and international levels — what institutions, rules, and accountability mechanisms are needed to ensure that AI serves humanity.