The cloud feels weightless. We send a prompt, receive an answer and continue with the working day.
But every prompt lands somewhere physical.
It is processed by servers, carried through networks, powered by electricity and cooled by industrial systems. Behind the convenience are buildings, cables, minerals, water supplies, land and eventually discarded hardware.
The honest question is not whether AI is good or bad for the planet. It is what resources we are using, what benefits we are receiving and whether the trade-off is justified.
What actually sits behind the cloud
Cloud computing is not an abstract space floating above us. It is a network of data centres filled with specialist equipment.
An AI service may depend on:
- High-performance servers and graphics processors
- Storage systems holding vast quantities of data
- Networking equipment and fibre connections
- Backup generators and batteries
- Cooling systems, pumps and ventilation
- Buildings, substations and physical land
- Minerals used in chips, batteries and electrical components
- Engineers, maintenance teams and replacement parts
AI workloads can be particularly demanding because they often run on dense clusters of specialised processors. These systems generate substantial heat. They also create rapid changes in power demand as workloads start, stop or increase in complexity.
The International Energy Agency describes the situation plainly. Its latest analysis says the speed of AI development is increasingly contrasting with the speed of the physical systems that support it, including electricity networks, chip manufacturing and planning processes.
The energy question
The global numbers are significant, although they should be treated as projections rather than certainties.
The IEA expects global data centre electricity consumption to roughly double from around 485 terawatt-hours in 2025 to around 950 terawatt-hours by 2030. That would represent approximately 3% of global electricity demand.
This does not mean all of that electricity is used by AI. Data centres also support websites, online storage, streaming, business software and many other services. However, AI-focused facilities are growing faster than the wider data centre sector.
The UK is facing its own infrastructure questions.
According to reporting on Ofgem’s assessment, around 140 proposed UK data centre projects could require up to 50 gigawatts of power. That figure exceeds current national peak demand. The National Grid DSO data centre impact study also warns that data centre demand could reach approximately 11% of Great Britain’s electricity demand by 2035.
These figures describe proposed capacity, not guaranteed future consumption. Some projects may be delayed, reduced or never built. Even so, they show the scale of the planning challenge.
The UK Government has also revised its estimates for AI compute emissions sharply upwards. Its corrected Compute Evidence Annex estimates cumulative emissions from UK AI compute between 2025 and 2035 at 34 to 123 million tonnes of carbon dioxide.
That is a very wide range. It reflects uncertainty about:
- How quickly AI adoption will grow
- How efficient models and processors will become
- How much the electricity grid will decarbonise
- How much infrastructure will be built
- Which direct and indirect emissions are included
Uncertainty does not make the issue unimportant. It means organisations should avoid presenting one neat number as the complete answer.
There is also a useful counterpoint. Energy use per AI task is falling rapidly. The IEA says simple text queries now typically use less electricity than running a television for the same period. More complex tasks, including video generation, advanced reasoning and agentic systems, can consume hundreds or thousands of times more energy than simple text generation.
Efficiency is improving. Usage is expanding faster.
Water, the footprint nobody mentions
Electricity is only one part of the environmental footprint. Data centres also need to manage heat, and cooling can require substantial quantities of water.
Evaporative cooling works by using water to carry heat away as some of the water changes into vapour. Other systems use air cooling, chilled water, direct-to-chip cooling or immersion cooling. Some operate in closed loops and recirculate water, while others rely more heavily on evaporation.
Each approach changes the balance between energy and water.
A system that uses less water may require more electricity. A water-efficient data centre may therefore reduce its direct water use while increasing the indirect water associated with power generation. There is no single solution that removes every trade-off.
The UK Government’s report on water use in data centres and AI notes that cooling can account for a substantial share of data centre energy use and that AI creates additional pressure through higher-density servers and continuous workloads.
Location matters as well. Global Action Plan’s report estimates that 84% of proposed UK data centres are located in areas projected to be water-stressed by 2040.
That does not mean every proposed facility will consume the same amount of water. It does mean that a national digital strategy intersects with local water availability, especially in already pressured catchments.
Per-query estimates vary enormously. A short text prompt may use a fraction of a millilitre under some conditions, while generating an image or video sequence can require far more. The UN University research gives examples ranging from approximately 29 millilitres for an AI-generated image to 4.1 litres for a complex video, depending on the assumptions and system involved.
Other estimates are lower or higher because they use different models, locations, cooling systems and boundaries. This is why any single headline figure should be treated cautiously.
The reliable point is not that every prompt uses a fixed amount of water. It is that water use scales with demand and remains almost completely invisible to the person making the request.
The waste problem
AI hardware has a physical lifecycle.
Processors, servers, batteries and cooling equipment must be manufactured, transported, maintained and eventually replaced. Rapid performance improvements can encourage shorter refresh cycles, particularly where older hardware cannot run the latest models efficiently.
The United Nations University Institute for Water, Environment and Health estimates that AI infrastructure could generate up to 2.5 million tonnes of electronic waste annually by 2030. Across different studies, estimates range from approximately 1.2 to 5 million tonnes.
These are projections, but the direction is clear. More AI infrastructure means more specialist hardware entering the system.
The National Grid DSO study also highlights a less visible issue. Rapid AI inference peaks can place additional stress on backup power batteries, potentially shortening their useful life and increasing the frequency of replacement.
Decommissioned equipment does not simply disappear. Some hardware can be refurbished, resold or redeployed. Some components can be recycled. Other equipment is dismantled in places where environmental controls and worker protections may be weaker.
This is where sensible maintenance matters. Extending the useful life of a server, laptop or network component is not always possible, but replacing equipment earlier than necessary creates its own environmental cost.
Why this is genuinely difficult to reason about
AI’s environmental footprint is difficult to understand because the infrastructure is distant, shared and constantly changing.
A person can see a television drawing electricity from a wall socket. They cannot see the data centre processing an AI request, the water used to cool it or the minerals contained in the processor.
A 2026 study published in Electronic Markets found that people tend to diffuse responsibility for AI’s energy use. Participants placed more responsibility on developers and policymakers than on users themselves.
That reaction is understandable. Users do not control data centre design, model architecture or electricity procurement. However, individual usage still contributes to total demand, particularly when multiplied across millions of people and organisations.
There is also a rebound effect.
If AI becomes more efficient and cheaper to use, people may use it more often. A saving per query can be outweighed by a much larger number of queries. The same applies to longer outputs, higher-resolution images, repeated revisions and automatically generated content that nobody ultimately uses.
Efficiency matters. Efficiency without sensible demand can still produce rising total consumption.
What a fair comparison looks like
AI may replace some physical activity.
It can reduce certain journeys, support remote collaboration, assist with research and reduce the need to create some physical drafts or prototypes. In some situations, it may replace printing, travel or repeated manual processing.
It can also add consumption.
An AI-generated image may be produced, revised several times, stored indefinitely and shared across multiple systems. A digital meeting may avoid travel but still involve cameras, microphones, networks, cloud recording and transcription. A tool that saves staff time may make it economically attractive to produce far more material than before.
The fair comparison is therefore not “AI versus nothing”. It is:
- What activity is being replaced?
- What new activity is being created?
- What infrastructure supports both?
- How long will the equipment remain useful?
- Who benefits from the service?
- Which communities carry the energy, water, land and waste burden?
The UNU report makes a similar argument. Professor Kaveh Madani described it as “a call for using it responsibly and addressing its unintended impacts proactively to make it sustainable and equitable.”
That is a more useful position than either greenwashing or doom. We need measurement that includes energy, carbon, water, land, minerals and waste rather than relying on one attractive sustainability claim.
What organisations can sensibly do
Organisations cannot solve the entire infrastructure problem through individual prompts. They can, however, make more informed choices.
High priority actions
- Use the lightest suitable tool
Do not default to the largest model or highest-output format for a simple task. A short text classification task does not require an image or video system.
- Define when AI is genuinely useful
Use AI where it improves a clear process, reduces avoidable work or supports a considered outcome. Avoid generating content simply because it is easy to produce.
- Ask suppliers for evidence
Review published information about data centre location, energy sourcing, water use, hardware lifecycles, emissions boundaries and electronic waste. Treat vague claims as incomplete rather than automatically reassuring.
- Use existing systems before adding more
Review the tools your organisation already pays for. Adding another AI platform may duplicate features, increase storage and make lifecycle management more difficult.
Medium priority actions
- Set sensible defaults
Limit unnecessary output length, image resolution, video duration, automatic transcription and duplicate generation where the tool allows it.
- Review storage and retention
Create clear rules for backups, recordings, generated files and archives. Storage that grows indefinitely still requires servers, electricity and hardware.
- Extend equipment life where sensible
Use maintenance, cleaning, upgrades, refurbishment and redeployment where they are safe and practical. A lifecycle review should consider reliability and security rather than focusing only on the newest specification.
- Keep technology decisions documented
Our managed IT services include the sort of monitoring, asset records, backup reviews and lifecycle planning that can help organisations understand what technology they already operate before adding more. The wider principle is simple. Good technology management should reduce unnecessary replacement and make resource use visible.
Lower priority actions
- Track usage patterns
Record which AI tools are being used, for what tasks and at what scale. This can reveal duplication and unnecessary high-compute workloads.
- Include environmental questions in procurement
Ask whether suppliers publish energy, water and hardware information. Also ask how they handle retired equipment and whether their figures cover inference as well as training.
- Review the position regularly
AI models, hardware and supplier disclosures are changing quickly. A decision that was reasonable this year may need reviewing as better information becomes available.
The takeaway
The cloud is not weightless. AI depends on physical infrastructure that consumes electricity, uses water, occupies land and requires minerals, buildings, cables and specialist hardware.
The central points are these:
- Global data centre electricity use could roughly double from 485 TWh in 2025 to 950 TWh in 2030.
- UK data centre proposals could place substantial pressure on national electricity planning.
- UK AI compute emissions estimates now range from 34 to 123 MtCO₂ between 2025 and 2035.
- Water use varies widely according to the task, model, location and cooling system.
- 84% of proposed UK data centres are projected to be in water-stressed areas by 2040.
- AI infrastructure could generate up to 2.5 million tonnes of electronic waste each year by 2030.
- Efficiency gains are valuable but may be offset if usage grows faster.
- Organisations should measure trade-offs honestly rather than relying on simple claims that AI is either clean or destructive.
Public concern is rising. The AP-NORC September 2026 poll found that 53% of Americans were extremely or very concerned about the environmental consequences of AI, up from 41% in 2025.
That concern does not require rejecting AI. It requires better information, clearer responsibility and more deliberate use.
If digital consumption does not appear on the bill we personally pay, how should we decide what level of energy, water and material use is acceptable?