Definition
Every chatbot answer, every generated image and every model trained runs on electricity in data centres, which also need water for cooling. As AI use explodes, its environmental footprint is becoming a serious question, and also something AI can help reduce.
The environmental impact of AI includes the electricity used to train and run models (and the carbon emitted to generate it), water used to cool data centres and generate power, the materials and manufacturing behind chips and servers, and electronic waste. Impact depends on model size, hardware efficiency, how often the model is used and the carbon intensity of the local electricity grid. AI can also reduce environmental harm, for example by optimising energy grids, forecasting weather and monitoring deforestation.
In simple words
Think of AI like a huge kitchen that cooks answers. The bigger the stove and the more meals it cooks, the more gas it burns. A smaller stove for simple meals, clean fuel and cooking only what is needed all cut the bill and the smoke.
Overview
- AI uses energy, water and materials to train and run models.
- Training one large model can consume over a thousand megawatt-hours of electricity.
- Patterson et al. (2021) estimated GPT-3's training at about 1,287 MWh and 552 tonnes of CO₂-equivalent.
- Running models for millions of users (inference) can use more energy over time than training.
- Data centres use water for cooling and depend on the local electricity mix.
- AI can also help the environment, for example in climate modelling, energy grids and agriculture.
Why it matters
AI's energy demand is growing quickly while the world is trying to cut emissions, and India's grid still relies heavily on coal. Developers and users can make choices that greatly reduce AI's footprint without losing much benefit.
Key terms
- Training energy
- Electricity used once to train a model, which can be very large for frontier models.
- Inference energy
- Electricity used every time a model answers; over millions of uses it can exceed training.
- Carbon intensity
- Grams of CO₂ emitted per kilowatt-hour of electricity, which depends on the grid's energy sources.
- PUE
- Power usage effectiveness: total data-centre energy divided by computing energy; 1.0 is ideal.
- Water footprint
- Water consumed for cooling data centres and generating their electricity.
- Green AI
- Designing and using AI with efficiency and environmental cost in mind.
Types
AI's environmental impact has several types. Each can be measured and reduced in its own way.
1. Training energy and emissions
Electricity used and carbon emitted while training a model.
- How it works: Energy equals GPU power × number of GPUs × hours × data-centre overhead (PUE); emissions depend on the grid's carbon intensity.
- Example: The same training run emitting several times more CO₂ on a coal-heavy grid than on a hydro-powered one.
- When to use it: Measure it with tools such as CodeCarbon and report it in model cards.
2. Inference energy
Energy used each time a model answers a request.
- How it works: Each query uses little, but billions of queries add up; larger models and longer outputs use more.
- Example: Routing simple questions to a small model instead of a huge one to save energy.
- When to use it: Reduce it with smaller models, caching, quantisation and shorter outputs.
3. Water use
Water consumed to cool data centres and to generate their electricity.
- How it works: Evaporative cooling consumes fresh water; usage varies greatly with climate, season and cooling design.
- Example: Concerns about new data centres in water-stressed regions.
- When to use it: Consider it when choosing data-centre locations and providers. Ask providers to report water use per unit of compute, and prefer sites with air cooling or recycled water.
4. Hardware and electronic waste
The materials, manufacturing emissions and waste of chips and servers.
- How it works: Making GPUs needs mining and energy-intensive factories; short upgrade cycles create e-waste.
- Example: Retiring servers after a few years because newer chips are much faster.
- When to use it: Reduce it by extending hardware life, reusing and recycling equipment. Choose efficient models so fewer chips are needed in the first place.
5. AI for sustainability
Using AI to cut emissions and protect the environment.
- How it works: AI forecasts renewable energy output, optimises cooling and traffic, monitors forests and improves crop yields with less water.
- Example: DeepMind's AI reducing the energy used to cool Google data centres by up to 40% (2016).
- When to use it: Use it to weigh AI's benefits against its footprint when choosing projects.
Explained in detail
How big is the footprint?
A 2021 study by David Patterson and colleagues estimated that training GPT-3, a 175-billion-parameter model, used about 1,287 megawatt-hours of electricity and emitted about 550 tonnes of CO₂-equivalent, while efficient choices of hardware, data-centre location and model design could cut such emissions by large factors. Since then models have grown and usage has exploded, so the energy used to answer billions of queries now often exceeds the energy used for training. The International Energy Agency's 2025 report "Energy and AI" estimated that data centres used about 415 terawatt-hours in 2024, around 1.5 per cent of global electricity, and projected this could roughly double to about 945 terawatt-hours by 2030, with AI the biggest driver. Water use matters too: data centres use water for cooling, and researchers have estimated that a short conversation with a large chatbot can indirectly consume an amount of water comparable to a small bottle, depending on location and season.
Where emissions come from
| Factor | Lower footprint | Higher footprint |
|---|---|---|
| Model size | Small or distilled model suited to the task | Largest model for every request |
| Hardware | Modern, efficient accelerators | Old, inefficient servers |
| Data centre | Low PUE, efficient cooling | High PUE, wasteful cooling |
| Grid | Solar, wind, hydro or nuclear power | Coal-heavy power |
| Usage | Caching, batching, shorter outputs | Repeated, unnecessary calls |
AI for the environment
AI is also a tool for sustainability. Machine-learning weather models now produce forecasts that rival traditional systems at a fraction of the computing cost. AI helps grid operators forecast solar and wind output and balance demand, detects methane leaks and illegal deforestation from satellite images, optimises building cooling and logistics routes, and helps farmers use water and fertiliser more precisely. Google reported in 2016 that DeepMind's system cut the energy used for cooling its data centres by up to 40 per cent. The responsible path is to use AI where its benefits clearly outweigh its costs, measure and report energy use, and choose the most efficient options.
Transparency helps here too. Few companies publish the energy and water used by their models, which makes it hard for users and policymakers to compare options. Tools such as CodeCarbon estimate the emissions of training runs, and researchers increasingly report compute and energy in papers. Asking providers for this information, and preferring those who publish it, encourages the whole industry to become more efficient.
How it works
- Choose the smallest model that meets the quality you need.
- Reduce wasted computation: cache results, batch jobs and cap output length.
- Train and run where electricity is cleaner, and prefer efficient data centres.
- Measure energy and estimate emissions for training and usage.
- Report the footprint in your model card and weigh it against benefits.
In code
def emissions_kg(gpu_count, hours, watts_per_gpu=400, pue=1.3, grid_g_per_kwh=700):
kwh = gpu_count * hours * watts_per_gpu / 1000 * pue
return kwh, kwh * grid_g_per_kwh / 1000
print("Fine-tuning a small model: 4 GPUs x 6 hours")
for grid, g in [("coal-heavy grid (about 700 g/kWh)", 700), ("mostly renewable grid (about 50 g/kWh)", 50)]:
kwh, kg = emissions_kg(4, 6, grid_g_per_kwh=g)
print(f" {grid:<40} {kwh:6.1f} kWh {kg:6.1f} kg CO2")
print("Training a larger model from scratch: 256 GPUs x 30 days")
kwh, kg = emissions_kg(256, 30 * 24)
print(f" {kwh:,.0f} kWh {kg / 1000:,.1f} tonnes CO2 on a coal-heavy grid")
# Inference adds up: a million requests a day for a year
wh_per_request = {"large model": 3.0, "small model": 0.3} # illustrative watt-hours per answer
for name, wh in wh_per_request.items():
year_kwh = wh * 1_000_000 * 365 / 1000
print(f"{name:<12} 1M requests/day for a year: {year_kwh:,.0f} kWh, {year_kwh * 0.7 / 1000:,.0f} tonnes CO2")Output
Fine-tuning a small model: 4 GPUs x 6 hours coal-heavy grid (about 700 g/kWh) 12.5 kWh 8.7 kg CO2 mostly renewable grid (about 50 g/kWh) 12.5 kWh 0.6 kg CO2 Training a larger model from scratch: 256 GPUs x 30 days 95,846 kWh 67.1 tonnes CO2 on a coal-heavy grid large model 1M requests/day for a year: 1,095,000 kWh, 766 tonnes CO2 small model 1M requests/day for a year: 109,500 kWh, 77 tonnes CO2
Try it yourself
A greener homework helper
An edtech startup's homework bot answers 2 million questions a day using a very large model.
Raises or lowers the footprint?
Decide for each choice.
Generating ten long images when one would do
Using the largest model for every simple question
Fine-tuning a pre-trained model instead of training from scratch
Running in a data centre powered by renewables
Caching answers to repeated questions
Common mistakes
- Ignoring inference energy because training gets the headlines.
- Defaulting to the largest model when a smaller one is good enough.
- Forgetting water use and local grid carbon intensity.
Check yourself
Summary
- AI's footprint includes electricity, carbon, water, hardware and e-waste, for both training and use.
- Model size, hardware, data-centre efficiency, grid carbon and usage patterns decide the impact.
- Right-size models, avoid waste, use clean power, measure and report, and use AI to help the environment.
Finished reading?
Sources: International Energy Agency, Energy and AI (2025) · Patterson et al. (2021), Carbon Emissions and Large Neural Network Training