What Is AI Costing the Planet — and Who Pays?

AI systems consume electricity at two distinct stages, and both matter. Training — the process of building a model by adjusting its parameters across billions of examples — is the more dramatic consumption event. Training a single large language model can consume electricity comparable to the lifetime energy use of multiple cars. But training happens once. Inference — running the model to respond to user queries — happens billions of times per day across all deployed systems. At the scale of global deployment, inference is the larger sustained energy draw, and it grows with every new user, every new application, and every expansion of AI capability. Kate Crawford, whose comprehensive mapping of the material infrastructure of AI drew on fieldwork at data centres, mines, and logistics facilities across multiple continents, frames the energy question with precision: AI is not a cloud-based abstraction. It is a physical system of machines that require electricity to operate, cooling systems that require water to function, and built infrastructure that required minerals and labour to construct. The energy consumed by those machines comes from somewhere specific — from grids that may or may not be powered by renewable sources, in regions whose communities bear the local consequences of that energy production.

"AI is neither artificial nor intelligent. It is made from natural resources, extracted from the earth, and runs on energy drawn from grids whose carbon intensity varies enormously by location."
— Kate Crawford, Atlas of AI, 2021

AI has a significant and growing environmental footprint — in carbon emissions, water consumption, and the extraction of minerals required to build its hardware. Crawford's analysis establishes the full supply chain: the lithium, cobalt, and rare earth elements required for the chips and batteries that power AI hardware are extracted from specific locations — lithium from the salt flats of Bolivia, Chile, and Argentina; cobalt predominantly from the Democratic Republic of Congo — through processes that damage local ecosystems, displace communities, and concentrate toxic waste near the people who live closest to the extraction sites. Water consumption is a less-discussed but substantial cost. Data centres require enormous quantities of water for cooling — evaporative cooling systems can consume millions of litres per day at large facilities. This consumption occurs at specific locations, drawing on local water supplies that may already be under stress. In arid regions where data centres have been sited for land cost or tax reasons, the water demands of AI infrastructure compete directly with agricultural and residential needs. The counter-argument — that AI applications in energy optimisation, climate modelling, materials science, and logistics can reduce carbon emissions more than AI infrastructure produces — is made seriously by serious researchers. Crawford's response is not that AI cannot reduce emissions in specific applications. It is that this calculation is rarely performed honestly, that the reductions from applications are speculative and distributed while the costs of infrastructure are concrete and localised, and that the communities bearing the environmental costs of AI infrastructure are almost never the same communities that will benefit from AI-driven climate solutions.

"The environmental costs of AI are real, located, and borne by specific communities. The environmental benefits are projected, diffuse, and accrue to different people entirely."
— Kate Crawford, Atlas of AI, 2021

AI is making inequality worse in several distinct and compounding ways. The first is economic: the financial returns from AI accrue overwhelmingly to the organisations that own the infrastructure and the models — predominantly large technology companies and their shareholders in wealthy countries. The labour required to make those models function — the data annotation, content moderation, and training feedback work performed by millions of workers, many in the global south — is paid at rates that reflect the labour market conditions in those locations, not the value those workers contribute to systems worth hundreds of billions of dollars. Crawford's documentation of this labour supply chain is forensic. The "intelligence" of AI systems is produced partly by human workers performing tasks — identifying objects in images, flagging harmful content, rating the quality of AI outputs — that are tedious, sometimes deeply distressing, and compensated at rates that would be considered unacceptable for equivalent work in the countries where the AI systems are deployed. The invisibility of this labour — described by companies as "data work" or "quality assurance" — is structural. Presenting AI as autonomous intelligence requires concealing the human effort that makes it function. The second dimension of AI-driven inequality is geographic: AI systems trained predominantly on English-language data from wealthy, internet-connected populations serve those populations best. The accuracy, utility, and cultural relevance of AI systems degrades for users whose languages, cultural contexts, and information environments are underrepresented in training data. The communities bearing the greatest environmental costs of AI infrastructure — those located near mines, data centres, and manufacturing facilities — are typically the communities least well-served by the systems those infrastructures support.

"AI is a technology of extraction — of data, of labour, of minerals, of energy. What is extracted flows upward. What is left behind stays where it is."
— Kate Crawford, Atlas of AI, 2021

Every AI query begins, physically, in a mine. The graphics processing units that power AI training and inference are manufactured from silicon, requiring highly purified quartz. Their memory systems require rare earth elements. Their power systems require lithium and cobalt. Each of these materials is extracted from specific locations by workers whose labour conditions, environmental exposures, and political circumstances vary enormously and are rarely visible to the end users of the systems their work enables. Crawford traces this supply chain with unusual specificity. The cobalt that enables the battery systems of AI hardware is mined predominantly in the Katanga province of the Democratic Republic of Congo, where artisanal mining operations — including child labour — exist alongside industrial extraction. The lithium that powers portable devices is concentrated in the Atacama Desert of South America, where extraction processes consume water from one of the most arid ecosystems on earth and affect indigenous communities whose land rights over the affected areas are contested. From minerals, the supply chain moves to manufacturing — chip fabrication in facilities in Taiwan, South Korea, and increasingly the United States and Europe — then to logistics, then to the data centres where the hardware is operated. Data centres are physical buildings, located in specific places, consuming electricity and water from local grids and aquifers, and generating heat that must be managed. The siting decisions for these facilities are made on the basis of land cost, energy cost, tax incentives, and regulatory environment — not on the basis of the environmental carrying capacity of the chosen location or the preferences of the communities that will live alongside them.

"To understand AI, you have to follow it all the way down — past the interface, past the server, past the chip, to the mine. That is where it begins. And that is where most of its costs are left."
— Kate Crawford, Atlas of AI, 2021

The benefits of AI accrue overwhelmingly to a small number of technology companies, their shareholders, their high-skill employees, and the users in wealthy countries who can access and effectively use AI tools. The costs — environmental, labour, and displacement — are distributed across a much larger and less powerful set of actors: mining communities, data annotation workers, populations whose jobs are disrupted, and communities bearing the environmental burden of AI infrastructure. This is not a coincidence or a market failure in the conventional sense. It is the predictable outcome of a system in which the people who capture the benefits also make the decisions about how the costs are distributed — and have limited accountability to those bearing the costs. Crawford identifies this as the defining political economy of AI: it is a technology that transfers value upward along existing lines of power while externalising costs downward along the same lines. The geographic dimension is particularly stark. The most capable AI systems are developed in the United States and, to a lesser extent, China and Europe. They are trained predominantly on data generated by users in wealthy, internet-connected countries. They perform best for users whose languages, cultural contexts, and information needs are well-represented in that data. The communities that bear the greatest environmental costs of AI infrastructure — those near mines, data centres, and manufacturing facilities — are typically in the global south, with the least political power to resist those costs and the least access to the benefits the infrastructure supports.

"The geography of AI's costs and benefits is not random. It follows the existing geography of power. Those with the least power bear the most cost and receive the least benefit. This is not a bug. It is how extraction works."
— Kate Crawford, Atlas of AI, 2021

• AI is a physical system of extraction, not a virtual one — every AI query begins in a mine and passes through a supply chain of minerals, manufacturing, logistics, and data centre infrastructure before it reaches a user. The energy and water consumed, and the labour performed, at each stage of that chain are real costs borne by real communities. (Crawford, Atlas of AI, 2021)
• The energy demand of AI is large and growing — training a single large language model can consume electricity comparable to the lifetime energy use of multiple cars, and inference at global scale represents a sustained and expanding energy draw that is already straining grid capacity in several regions. (Crawford, Atlas of AI, 2021)
• The "intelligence" of AI systems is produced partly by invisible human labour — data annotation, content moderation, and training feedback work performed by millions of workers, many in the global south, at compensation rates that do not reflect the value they contribute to systems worth hundreds of billions of dollars. (Crawford, Atlas of AI, 2021)
• AI's environmental costs and benefits are distributed to different communities — the communities bearing the greatest environmental costs of AI infrastructure are typically not the communities that will benefit from AI-driven solutions to environmental problems. The calculation cannot be made honestly by aggregating global costs and global benefits. (Crawford, Atlas of AI, 2021)
• AI transfers value upward along existing lines of power while externalising costs downward along the same lines — changing this distribution requires deliberate governance intervention, not market correction, including accountability for the full supply chain of AI operations and international mechanisms that give affected communities a voice in decisions that impose costs on them. (Crawford, Atlas of AI, 2021)

How much energy does AI use?

AI systems consume electricity at two distinct stages, both of which matter at scale. Training — building a model by adjusting its parameters across billions of examples — is the more dramatic single consumption event. Training a large language model can consume electricity comparable to the lifetime energy use of multiple cars. But training happens once. Inference — running the model to respond to user queries — happens billions of times per day across all deployed systems, and at the scale of global deployment represents the larger sustained energy draw. That draw grows with every new user, every new application, and every expansion of AI capability. Kate Crawford's Atlas of AI (2021) documents the physical infrastructure behind this consumption: data centres that require electricity from local grids, cooling systems that consume water, and built infrastructure that required minerals and labour to construct. The energy consumed comes from grids whose carbon intensity varies enormously by location, and in several documented cases the pace of data centre expansion has outrun the pace of renewable capacity addition, resulting in short-term increases in fossil fuel generation. Major technology companies have made carbon neutrality commitments, but the gap between those commitments and actual operational carbon intensity is a measurement and accountability problem that current reporting requirements do not adequately address.

Is AI bad for the environment?

AI has a significant and growing environmental footprint in carbon emissions, water consumption, and the extraction of minerals required to build its hardware. The lithium, cobalt, and rare earth elements required for AI chips and power systems are extracted from specific ecosystems — lithium from South American salt flats, cobalt predominantly from the Democratic Republic of Congo — through processes that damage local environments and concentrate toxic waste near the communities closest to extraction sites. Data centre cooling systems consume millions of litres of water per day at large facilities, competing with agricultural and residential water needs in regions where those facilities are sited. The counter-argument — that AI applications in energy optimisation and climate modelling can reduce emissions more than AI infrastructure produces — is made seriously, but Crawford's analysis in Atlas of AI (2021) identifies the critical problem with that argument: the environmental costs of infrastructure are concrete and localised, falling on specific communities, while the environmental benefits of applications are projected and diffuse, accruing to different people in different places. The question cannot be answered honestly by aggregating global costs and global benefits. It requires asking who bears which costs and who receives which benefits.

Is AI making inequality worse?

Yes — in several compounding ways. The financial returns from AI accrue overwhelmingly to the organisations that own the infrastructure and the models, predominantly large technology companies and their shareholders in wealthy countries. The labour required to make those models function — data annotation, content moderation, and training feedback performed by millions of workers, many in the global south — is compensated at rates that do not reflect the value those workers contribute to systems worth hundreds of billions of dollars. AI systems trained predominantly on English-language data from wealthy populations serve those populations best, degrading in accuracy and cultural relevance for users whose languages and contexts are underrepresented in training data. The communities bearing the greatest environmental costs of AI infrastructure are typically those with the least political power to resist those costs and the least access to AI's benefits. Crawford's framework in Atlas of AI (2021) identifies the underlying pattern: AI transfers value upward along existing lines of power while externalising costs downward along the same lines. This is not a market failure that will self-correct. It is the predictable outcome of a system in which those who capture the benefits also make the decisions about how costs are distributed, with limited accountability to those bearing them.

1. Crawford, Kate. Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence. 2021.