We often hear about AI as a solution to the environmental crisis: modelling the climate, protecting biodiversity, detecting pollution, optimizing farming and water management, or planning a transition to circular economy systems.
AI can support all of these things, but it also has a huge environmental cost.
Behind every model, chatbot, image generator, prediction tool, or automated analysis system is physical infrastructure: data centres, electrical grids, cooling facilities, servers, chips, minerals, water, land, buildings, and supply chains. AI may feel weightless, something you see on a screen, but it is not immaterial.
The environmental question, then, is not just whether AI can support sustainability, but whether its impacts are measured, mitigated, and governed appropriately, and whether its benefits outweigh these impacts. AI is not by default a green technology just because it is digital.
AI runs on physical infrastructure
The discourse about AI often seems to be in the ether. We hear about algorithms, platforms, clouds, models, and applications. But the systems running them are located in real-world physical infrastructure in real-world places.
Data centres require land, power connections, cooling infrastructure, electrical backup, hardware and network connectivity. Servers rely on chips, memory, storage, cables and controls. These systems depend on the mining, processing, transportation and assembly of raw materials through international supply chains.
This puts AI directly within existing discussions about resource use, waste, and dependence on raw materials. Digital systems, after all, still rely on physical materials, and demands for materials still have consequences.
The more AI expands, the more essential it becomes to understand the broader system behind it.
Electricity is the most visible environmental pressure
The most talked about environmental impact of AI is its power consumption.
Training huge AI systems takes a lot of computing power, and running AI services for millions of users takes power too. As AI gets embedded in search, business software, image generation, coding assistance, customer service, logistics, healthcare, and public administration, we should expect data centres’ power consumption to grow.
This is significant because electricity is not environmentally free.
If new AI electricity demand is supplied from fossil fuel generation, emissions will go up. If it is supplied with clean energy, it might crowd out other urgent clean electrification demands like homes, district heating, public transport, industry or renewables.
So AI’s growing presence in our energy supply chains is part of the broader climate and energy transition. The question is not only how much electricity the industry consumes, but where the electricity comes from, when it is used, and how local grids handle the load.
AI infrastructure planning should be conducted with the same degree of care expected from other major energy consumers.
Renewable power is needed, but not enough
Tech companies regularly tout the importance of renewable power sourcing in relation to AI energy consumption.
Clean electricity matters. Data centres should be running off clean, renewable, and low-carbon energy where feasible. Renewable energy pledges, though, must be scrutinized. While purchasing renewable energy credits, a data centre might still depend on a grid that occasionally generates power using fossil fuels.
It could contribute to local demand in a region that is already struggling to provide sufficient power. It might necessitate additional electricity transmission lines, back-up sources of power or local energy infrastructure. It might also increase demand for renewable energy that could have served other consumers instead.
Ultimately, AI infrastructure either genuinely contributes to a greener energy system, or puts additional strain on it.
Where additional renewable energy capacity is being developed to fuel digital infrastructure, the basic principles of responsible renewable energy planning hold: siting of facilities, biodiversity and water, stakeholder engagement and environmental management.
AI companies should not only purchase clean power, but also help to make new energy demand transparent, additional and compatible with environmental objectives.
Water use is also at stake
AI has a water footprint too.
Many data centres consume water, either directly on-site or indirectly via the electricity generation that powers them. This can be problematic in water-stressed areas.
AI’s water impact is a function of location and energy mix. A data centre in a cool, water-rich area with low-water energy generation would be likely to have a different water footprint to one in a hot, drought-prone area relying on water-intensive cooling and electricity generation.
That means location matters.
Water use must be addressed prior to a data centre being approved and built. Assessments need to take into account local water availability, drought risk, the needs of the ecosystem, the demand of communities and future water availability in a climate-changed world.
Freshwaters are already under pressure from multiple stressors, from pollution and overexploitation to agriculture, industry and climate change, which makes freshwater protection a necessary component to any meaningful debate about AI infrastructure.
AI cannot support environmental policy if it undermines water security.
Raw materials and critical minerals count
AI is built on hardware and hardware depends on materials.
High-powered chips, servers, cooling infrastructure, batteries, power generation, and data-centre facilities all consume metals, minerals and other inputs, which have complicated and environmentally damaging supply chains.
Extraction of critical metals and materials impacts forests, water, land, wildlife and people. Production of electronic equipment often involves hazardous chemicals, high-energy usage and waste. As AI hardware grows more advanced and replacement rates increase, there is a higher need for new hardware.
A digital economy premised on constant growth of hardware cannot be divorced from concerns around extraction, recycling and resource availability.
The environmental footprint of AI does not just lie within the data centres themselves, but also in the mining, manufacture, obsolescence and disposal of hardware.
A more environmentally responsible AI industry would design, produce, maintain and refurbish equipment that lasts longer and is easier to repair, recycle and recover.
E-waste must be tackled
The rapid pace of technological evolution creates electronic waste.
Equipment for servers, chips, data storage and networking can become obsolete and/or less energy efficient. Failing to adequately reuse, refurbish, or recycle old equipment contributes to the overall e-waste problem.
E-waste can have harmful and beneficial qualities; some of it contains materials that can be valuable, others contain hazardous components. Inadequate handling risks exposing workers and people living around the e-waste to pollution; exporting waste and equipment to weaker regulatory regimes just moves the environmental damage elsewhere.
For this reason, we need clear standards for reuse and recycling of hardware, extended producer responsibility and the end-of-life safe management of equipment. Without this, digital advancement may create similar industrial pollution and toxic chemical risks to those environmental regulation was designed to stop.
A responsible AI economy will be one that contains clear rules on reuse and recycling of hardware, extended producer responsibility and safe disposal.
Not all AI uses are equally worthwhile
Not all uses of AI will be equally environmentally worthwhile.
There are clearly uses of AI which would more than outweigh their footprint through benefits such as cutting emissions, preventing pollution, helping to improve energy efficiency, improving response to natural disasters and improving the management of nature and natural systems.
AI which detects illegal deforestation, cuts down energy consumption or improves water management may be delivering far more environmental benefit than the resources it takes up.
Other uses may be harder to justify.
AI which produces large numbers of content or encourages consumption without obvious benefit, or which adds to a more complicated computing system with little social benefit, may contribute to environmental degradation without solving a meaningful problem.
Therefore, we should be looking at AI not just by how efficient it is, but by its purpose. The question is: what problem is the AI solving, and is it proportionate?
Efficiency gains can increase demand
As AI systems become more efficient, this is a good thing. More efficient chips, better cooling, more efficient models and more efficient data centres will mean that the footprint of the AI system itself becomes lower. However, efficiency alone will not reduce overall impact.
As AI becomes more efficient, the cost will come down for a given AI system and companies and countries may use AI more. So the net result can actually be more total demand for data centres and hardware.
We have to be aware of what is called a rebound effect. When things become more efficient, they may in practice not result in reduced overall impact as consumption increases.
We have to be aware of this environmental aspect as well.
So, we should not be looking at AI as just more efficient and cheaper. We should be looking at AI as being more useful, needed and efficient. If we are not looking at things in the bigger context, then we are merely increasing total demand.
Transparency is essential
One of the issues when discussing the environmental impact of AI is that it is not clear or transparent.
It is important that public authorities, researchers and local communities are informed about the environmental impacts of AI and that this information is available. The public needs more information on energy use, water consumption, carbon footprint, the lifecycle of equipment, e-waste generation and disposal, land use, supply chains and so on.
Without this information, there is no way to see whether AI infrastructure can deliver on climate, water and resource objectives.
It is essential for public authorities. Local communities being asked to host data centres need to know what the local environmental impacts would be. Regulators need to know about the water and electricity use. Investors need to know whether the environmental impact claims made are credible, and governments need to understand whether AI deployment fits with infrastructure planning.
Without this information, there is no way to see whether proposed environmental targets for AI can ever be met. The same logic must extend to AI infrastructure.
Regulating AI as infrastructure
While AI is commonly classified as a digital service, its substantial environmental impact demands that it be regulated as infrastructure.
Building large AI models involves planning for energy demand, water usage, land allocation, grid capacity, material sourcing, and waste disposal. These are not just commercial concerns; they are matters of public policy.
Regulators need to ask: do new AI data centres fit with local climate objectives, water availability, renewable energy strategies, community interests, and environmental boundaries?
Planning needs to consider cumulative effects as well as individual projects. A single data centre is easily manageable. But a concentration of data centres in a region can put serious pressure on grid capacity, water use, and land.
We need to think about the way AI infrastructure is sited and the kind of coordination involved in land-use planning to support clean energy and biodiversity.
Features of responsible AI infrastructure
Responsible AI infrastructure requires several principles.
- Data centres should be powered by low-carbon electricity, and ideally help to expand clean energy capacity.
- Water usage must be reported, tracked and restricted, particularly in water-stressed regions.
- Hardware design should aim at extended lifetimes, easier repair and reuse, as well as recycling.
- E-waste must be handled safely and transparently.
- Environmental claims need to be backed up, not used as advertising.
- Local authorities should assess impact on the grid, water, land and the local community.
- AI applications need to be judged by their environmental and social impacts, and the real value created.
- Public authorities must require public reporting on energy, water, emissions, and material impacts.
None of this is to deny AI development; it is to ensure its environmental costs are transparent, mitigated and justified.
How AI must pass the environmental test
AI could be an essential tool to inform environmental policy, to model climate risks, to monitor ecosystems, to inform smarter industry, to support smarter farming and reduce waste.
But it should not be able to expand without its own footprint in mind.
This environmental cost is not just the electrical power. It includes water, land, materials, hardware, supply chains, and e-waste. And it includes the opportunity cost of deploying resources here rather than somewhere else.
A sound environmental policy approach to AI should ask three questions:
- Is the AI use necessary?
- Is the AI being operated as efficiently and cleanly as possible?
- Is it delivering an environmental or social benefit that is worth the environmental footprint?
If the answer is yes, AI can help to drive the transition to a sustainable economy. If the answer is no, it could become just one more source of resource pressure behind the veil of innovation.
AI has a role to play in supporting environmental outcomes. But only if its environmental costs are transparently evaluated right from the start.
Further reading
A longer look at environmental policy issues raised by AI infrastructure:
