Artificial intelligence runs on electricity, water, minerals and land. Its ecological cost is not a footnote to the ethics of AI. It is part of them.

Digital technology presents itself as immaterial. We speak of the cloud, of virtual assistants, of information that seems to exist nowhere in particular. The language is soothing and almost entirely misleading.

A model answering a question draws current from a grid. That grid is fed by generation somewhere, and the somewhere has a fuel source, an emissions profile, and a set of communities living near it. The processors performing the computation sit in buildings that occupy land, consume water for cooling, and require minerals extracted from places where extraction is rarely gentle. When the hardware is superseded, which happens quickly, it becomes waste that must go somewhere.

None of this is hidden. It is simply not visible from the interface, and what is not visible is easy to leave out of the moral accounting.

Industry figures have begun to say plainly that energy will be the binding constraint on how far artificial intelligence can scale. That is an engineering observation, and it is probably correct. But it opens onto a question that engineering cannot answer: if this technology requires a significant share of the world’s power, who decides what that power is spent on, and who bears the cost of generating it?

The Ledger Nobody Keeps

Consider what a large model actually requires across its life.

Training. Months of continuous computation across thousands of specialised processors, consuming electricity at industrial scale before the system produces anything useful at all.

Inference. Every query thereafter. Individually small, collectively enormous, and growing as these systems are embedded into search, productivity software, customer service and public administration. The cumulative cost of ordinary use eventually exceeds the cost of training.

Cooling. Data centres generate heat proportional to their computation, and removing that heat consumes water. Facilities have been sited in water-stressed regions, where the competition between industrial cooling and agricultural or domestic supply is not theoretical.

Hardware. Specialised chips require rare earths, cobalt, lithium and copper. The conditions under which some of these are mined, particularly in the Democratic Republic of Congo, involve labour practices that no ethical framework would defend if they occurred in the countries where the chips are used.

Disposal. Electronic waste is among the fastest-growing waste streams globally, and much of it is processed informally in the Global South by workers without protection.

Almost none of this appears in the discussion of whether a given AI application is ethical. A system can be praised for fairness, audited for bias, and reviewed for privacy while its material footprint is treated as somebody else’s department.

Why This Belongs to Ethics

Laudato Si’ made an argument that is directly applicable here, and it was not primarily an environmental argument. It was that the environmental crisis and the social crisis are one crisis, because both arise from treating the world and the people in it as material for our projects.

The encyclical’s central claim, that everything is connected, is a claim about moral accounting. It means an assessment that examines a technology’s effect on users while ignoring its effect on the miner, the community near the data centre, and the generation that inherits the emissions is not a partial assessment. It is a distorted one, because it counts the people who are visible and omits the people who are not.

This is why the ecological cost of AI cannot be handled as a separate sustainability conversation running alongside the ethics conversation. The same principle governs both. If human dignity is not conditional on visibility, then the dignity of the person mining cobalt weighs as much as the dignity of the person whose loan application is being assessed.

Integral ecology asks for a single ledger, not two.

The Distribution Problem

The difficulty is not that AI consumes resources. Every valuable thing consumes resources, and the question is always whether the value justifies the cost.

The difficulty is that costs and benefits are landing in different places.

The productivity gains accrue largely to firms and consumers in wealthy economies. The mineral extraction happens elsewhere. The electronic waste is processed elsewhere. Climate consequences fall disproportionately on regions that contributed least to emissions and have least capacity to adapt. Water drawn for cooling is water not available locally.

This asymmetry is the pattern that Catholic social teaching has named repeatedly across a century of industrial development, and it is reappearing in a new form. Solidarity is precisely the demand that this pattern be refused: that those who benefit from a system be answerable for what its operation costs others.

The practical question is therefore not whether to build AI infrastructure but where, powered by what, drawing water from where, employing whom on what terms, and with what obligation to the communities affected. Those are governance questions, and they are currently being answered by siting decisions made on the basis of cheap power and permissive regulation.

What Would Count as Serious

A serious response would involve several things that are technically feasible and organisationally inconvenient.

Disclosure. Energy and water consumption reported per model and per service, in terms a non-specialist can evaluate. This is not commercially impossible. It is commercially unattractive, which is different.

Siting standards. Facilities located with regard to grid carbon intensity and local water stress, rather than only to electricity price and tax treatment.

Supply chain accountability. Traceability for minerals, with the same seriousness applied to conflict minerals in other industries, and consequences for firms whose supply chains cannot be traced.

Efficiency as a design value. Smaller models suited to the task rather than the largest available. Much deployed AI capability substantially exceeds what the application requires, and the excess is paid for in power.

Proportionality. The question of whether an application justifies its footprint, asked before deployment. A system that meaningfully improves medical diagnosis in an underserved region and a system that generates marginal engagement gains for an advertising platform are not equivalent uses of the same electricity.

That last point is the one most often avoided, because it requires judgement about which uses are worth their cost. But avoiding the judgement does not mean it goes unmade. It means it is made by whoever can pay.

Not Restraint for Its Own Sake

None of this argues for less technology. AI has genuine potential for ecological benefit: optimising energy grids, improving climate modelling, detecting deforestation, reducing waste in agriculture and logistics. Those applications may well justify substantial resource use.

The argument is for honesty in the accounting, and for the recognition that a technology’s ethics include its metabolism.

Laudato Si’ observed that humanity has developed enormous technological power without developing a corresponding capacity to set limits on it. The remark was not nostalgic. It identified a specific deficiency: we have become far better at building things than at deciding what is worth building.

Artificial intelligence presents that deficiency in a concentrated form. The systems are extraordinarily capable, their resource demands are large and growing, and the decision about what those resources should be spent on is currently distributed across firms optimising for their own returns.

An apparently weightless technology rests on physical systems and human bodies. Any account of its ethics that leaves them out has not finished counting.