AI is not inherently sustainable or unsustainable. Its impact depends on the model, hardware, data centre, electricity supply, cooling system, frequency of use and what the system replaces or enables. This question sits within the wider field of sustainable technology.
A useful assessment must consider both sides: the direct footprint of developing and running AI, and the indirect environmental benefits or harms created by its application. The answer cannot be reduced to one universal “carbon per prompt” figure.
The short answer
AI can support climate and conservation work, but its rapidly growing infrastructure uses electricity, water, minerals and hardware. Whether a particular use is justified depends on measurable benefits, suitable model size, efficient operation and transparent lifecycle reporting.
AI’s direct environmental impacts
Electricity and emissions
Training and using models require computation in data centres. The electricity demand varies enormously by model and task, while associated emissions depend on the grid and time of use. The International Energy Agency projects global data-centre electricity consumption to reach roughly 945 TWh in 2030 in its base case, with AI an important driver. A projection is not a measured outcome, but it shows the scale of the infrastructure question.
Water
Water may be consumed directly for cooling and indirectly in electricity generation and chip manufacturing. Impacts depend on cooling design, weather, water source and local scarcity. A global average cannot tell a community whether a specific facility is responsible.
Hardware, minerals and e-waste
AI accelerators and data-centre equipment require extraction, fabrication and transport. Their embodied impacts do not appear in operational electricity figures. Rapid replacement can also increase electronic waste. The UN Environment Programme’s lifecycle note recommends looking across critical minerals, manufacturing, use and end of life.
Training is not the whole story
Training a frontier model is energy-intensive, but repeated use—known as inference—can become the larger operational impact when a service reaches enormous scale. Retraining, fine-tuning, data processing and idle infrastructure also belong in the boundary.
This is why dramatic comparisons based on one model, one query or an undisclosed estimate age badly. Providers can improve hardware and models while total demand still rises through greater use.
How AI may support sustainability
- Forecasting renewable generation and electricity demand.
- Detecting methane leaks or environmental change in sensor and satellite data.
- Optimising industrial processes, buildings and transport systems.
- Supporting climate modelling and materials discovery.
- Monitoring biodiversity, land use and illegal extraction.
These are potential benefits, not automatic net savings. A system needs a clear baseline, evidence that its recommendations are acted upon and an assessment of rebound effects. Optimising traffic, for example, may reduce fuel per journey but encourage more journeys.
Why “AI for good” is not enough
An environmentally useful goal does not excuse an inefficient or harmful implementation. Teams should ask whether AI is needed, whether a smaller model or conventional method works, and whether the result creates enough benefit to justify its lifecycle cost.
How developers can make AI more sustainable
- Start with necessity: define the decision or outcome before selecting AI.
- Use the smallest capable model: compare compact, specialised and non-AI approaches.
- Reduce repeated work: cache safe outputs, batch requests and avoid unnecessary generations.
- Measure a useful unit: energy or compute per completed task, not model size alone.
- Choose infrastructure deliberately: consider hardware efficiency, utilisation, electricity and water context.
- Schedule flexible work: shift delay-tolerant training or batch jobs where appropriate.
- Design for longer hardware life: avoid software choices that force needless upgrades.
- Report uncertainty: publish boundaries, estimates and missing data.
The same foundations appear in energy-efficient coding and green software engineering.
What users can do
- Use AI when it adds real value, not by default for trivial tasks.
- Write a clear request to reduce repeated attempts.
- Prefer smaller or task-specific tools when they are sufficient.
- Avoid generating many unused high-resolution images or videos.
- Ask providers for energy, water, emissions and hardware-lifecycle reporting.
- Keep useful devices longer rather than upgrading solely for new AI features.
Questions to ask an AI provider
- What model and hardware serve this task?
- What is the measured energy per useful outcome?
- Which lifecycle stages are included?
- Where and when does computation occur?
- How are electricity emissions calculated?
- What water sources and cooling systems are used?
- How long is hardware used, and what happens at end of life?
- Are claims independently verified?
So, is AI sustainable?
Some AI applications may deliver worthwhile environmental benefits; others add resource demand without proportional value. Sustainable AI means refusing the blanket answer. Define the outcome, compare alternatives, measure the full system, reduce direct impacts and verify that promised benefits occur in the real world.
For the infrastructure underneath AI, continue with our beginner’s guide to data-centre energy use. The OECD AI footprint report provides a policy-level framework for direct and indirect impacts.