The Environmental Cost of AI
Artificial intelligence is often discussed as if it were immaterial — a set of algorithms existing in some abstract digital realm. In reality, AI systems require vast physical infrastructure: data centers filled with specialized chips, cooling systems demanding enormous quantities of water, electricity in quantities that strain regional grids, and supply chains for rare minerals embedded in hardware. The environmental cost of AI has been largely invisible in public discourse, but the research is accumulating rapidly, and the picture it reveals warrants serious attention from anyone thinking seriously about AI's social implications.
The energy footprint of large language models
Training a large language model is one of the most computationally intensive activities in existence. The training of GPT-3, estimated in 2020, consumed approximately 1,287 megawatt-hours of electricity and produced carbon dioxide equivalent emissions roughly comparable to a round-trip transatlantic flight for several people — though these estimates vary significantly by the carbon intensity of the electricity source. More recent models, substantially larger than GPT-3, are likely to have consumed considerably more energy, though most AI companies do not publicly disclose the training costs of their frontier models.
The energy burden of AI does not end with training. Inference — running a trained model to generate outputs in response to user queries — also consumes substantial energy, and at scale, inference may represent the dominant share of AI's total energy use. A single query to a large language model consumes roughly ten times the energy of a Google search, by some estimates, though the comparison is inexact and depends heavily on the model and query type. As AI is integrated into search engines, productivity tools, and consumer applications at scale, the aggregate inference energy consumption grows accordingly.
One of the significant obstacles to accurate accounting of AI's environmental footprint is the lack of transparency from AI companies about the energy and emissions associated with training and serving their models. Unlike, say, the airline industry — where per-seat emissions can be calculated with reasonable precision — the AI industry operates largely without mandatory environmental disclosure. Researchers have called for standardized reporting frameworks, but in the absence of regulatory requirements, most companies have provided limited information, and what is disclosed varies in methodology and scope.
Water consumption and data center infrastructure
Less discussed than energy, but equally significant, is the water consumption of AI data centers. Cooling is essential to data center operation — the computing hardware generates substantial heat and must be kept within safe operating temperatures. Many data centers use water evaporation as a primary cooling mechanism, consuming millions of gallons of water per year per facility.
Research published in 2023 estimated that training GPT-3 consumed approximately 700,000 liters of freshwater — roughly the equivalent of hundreds of millions of bottles of water. Operating large AI models at scale similarly requires continuous water consumption. Microsoft, Google, and other major AI infrastructure operators have disclosed year-over-year increases in their water consumption, attributed in significant part to AI workloads.
The geographic and social dimensions of data center water use matter. Data centers are often sited in regions where land and water are relatively cheap — regions that may also be facing water scarcity. The use of local water resources by energy-intensive facilities owned by distant technology companies, serving users around the world, raises genuine questions about local resource rights, environmental justice, and the distribution of AI's costs and benefits.
Hardware, minerals, and supply chain impacts
AI systems run on specialized hardware — primarily graphics processing units (GPUs) and application-specific chips designed for AI workloads. Manufacturing these chips requires rare minerals: cobalt, lithium, tantalum, and others, extracted in supply chains with documented environmental and human rights concerns. Cobalt, for example, is disproportionately mined in the Democratic Republic of Congo, where mining operations have been linked to child labor, environmental degradation, and community displacement.
The embodied carbon of AI hardware — the emissions associated with manufacturing the chips and servers that AI runs on — is a significant share of total lifecycle emissions that is rarely included in AI environmental assessments focused primarily on operational energy use. As the global stock of AI hardware grows rapidly to meet demand, this manufacturing footprint grows with it.
The disposal of AI hardware raises additional environmental concerns. Computing hardware has a limited operational lifespan, and as AI systems require more powerful and specialized chips, older hardware is retired. Electronic waste (e-waste) is one of the fastest-growing waste streams globally, and improper disposal exposes workers — disproportionately in lower-income countries — to toxic materials.
AI and climate change: the dual relationship
AI's relationship with climate change is genuinely dual: it is simultaneously a significant consumer of energy that may contribute to emissions, and a potentially powerful tool for addressing climate change.
On the contribution side, the rapid growth of AI infrastructure is a material factor in the electricity demand forecasts of major grid operators. Data center electricity consumption in the United States is projected to double or more between 2023 and 2030, driven significantly by AI workloads. If this growth is powered by fossil fuels — as much electricity generation currently is — AI's net contribution to emissions could be substantial. The major AI companies have made commitments to power their operations with renewable energy, but matching actual consumption to renewable generation is technically and contractually complex, and critics argue that "renewable energy certificates" purchased by technology companies do not reliably correspond to actual renewable generation.
On the contribution to solutions side, AI has genuine potential applications in climate change mitigation and adaptation. AI systems are being used to optimize electricity grid management, reducing waste and enabling higher penetrations of variable renewable energy. AI-assisted materials discovery has identified new candidates for solar cells, batteries, and carbon capture. AI-driven climate modeling may improve the accuracy and granularity of climate projections. Precision agriculture applications can reduce the land, water, and chemical inputs required to produce food. These applications are real and in some cases already materially significant.
A historical pattern in technology-driven efficiency gains is the "Jevons paradox": improvements in efficiency often lead to increased overall consumption rather than reduced consumption, because efficiency reduces the cost of an activity, stimulating greater demand. If AI makes energy systems more efficient, but also enables new energy-consuming applications that would not otherwise exist, the net effect on energy consumption is ambiguous. This is not an argument against efficiency improvements, but it is an argument against assuming they automatically translate into reduced environmental impact without accompanying demand management.
Environmental justice and the geography of AI's costs
The environmental costs of AI infrastructure — emissions, water use, land use, noise, heat — are not distributed randomly. Data centers are sited in specific communities, which bear the local environmental burden of serving a global user base. The mining of AI hardware minerals happens in specific places, disproportionately in the Global South. The climate change consequences of AI's energy consumption affect the most vulnerable communities most severely — communities that contribute least to AI's growth and benefit least from its applications.
This geographic and social pattern is a dimension of environmental justice that connects the environmental critique of AI to the broader inequality analysis we examined in module three. The question of who bears AI's environmental costs, and who captures its benefits, is a distributional question with justice implications.
The environmental costs of AI are not fixed — they depend on choices that can be made differently. More efficient model architectures reduce energy use per inference. Siting data centers where renewable energy is abundant and cheap, and where water is not scarce, reduces environmental impact. Mandatory environmental disclosure would enable better accountability. Research into smaller, more efficient models — "model compression," distillation, and efficient inference techniques — is active and promising. And using AI to accelerate the energy transition itself could, if the work is done well, produce net environmental benefits that outweigh AI's direct footprint. The environmental sustainability of AI is a genuine challenge, but it is also a design challenge — one for which better choices exist.