The Environmental Impact of AI
Lesson 8.6 — AI and the Environment
Every time you use an AI tool, a data centre somewhere is consuming electricity and water. The environmental impact of AI is real, and there is a meaningful discussion to be had about it. There is also a significant amount of misinformation and disproportionate framing — both from those who dismiss the concern entirely and from those who overstate it to the point of paralysis.
This lesson gives you the actual numbers, compares them fairly to other activities, describes what AI companies are doing in response, and helps you arrive at a proportionate personal perspective.
Energy: The Actual Numbers
AI systems — particularly large language models — require substantial computational power both to train and to run.
Training: Training a large model once is extremely energy-intensive. Training GPT-4 is estimated to have consumed approximately 50 gigawatt-hours (GWh) of electricity — comparable to the annual energy use of around 4,500 average US homes. This is a one-time cost, not an ongoing one.
Inference: Each time you send a message to an AI, the model processes it (inference). A single ChatGPT query is estimated to consume approximately 0.001–0.01 kWh — roughly 5–10x the energy of a standard Google search. This is not enormous for an individual query, but AI is being used at enormous scale — billions of queries per day across all platforms.
Data centres overall: AI is a growing proportion of global data centre energy use, but the picture is complex. Data centre efficiency has improved dramatically as demand has increased — the same amount of computation uses far less energy than it did a decade ago. AI's share of global electricity consumption is currently estimated at roughly 0.5–1%, with projections suggesting it could reach 3–5% by 2030 if growth continues.
Comparison Table: Putting AI Energy in Context
Comparison numbers are approximate and depend on many variables, but they provide a useful sense of scale:
| Activity | Approximate energy / CO₂ |
|---|---|
| One ChatGPT conversation (10 exchanges) | ~0.01 kWh / ~5g CO₂ |
| One Google search | ~0.001 kWh / ~0.4g CO₂ |
| Sending one email | ~0.0004 kWh / ~0.3g CO₂ |
| Streaming one hour of video (Netflix) | ~0.1 kWh / ~36g CO₂ |
| One hour of video conferencing | ~0.1 kWh / ~50g CO₂ |
| A cup of coffee (production + brewing) | ~0.2 kWh / ~70g CO₂ |
| One km driven in a petrol car | ~0.12 kWh / ~170g CO₂ |
| A return economy flight London–New York | ~1,000 kWh / ~300kg CO₂ |
| AI image generation (one image, DALL-E) | ~0.01–0.04 kWh / ~5–20g CO₂ |
| Training GPT-3 (one time) | ~1,300 MWh / ~550 tonnes CO₂ |
What this tells us: A conversation with an AI tool is not environmentally insignificant, but it is comparable to other digital activities we do without much thought. The concern is scale — when billions of queries happen every day, small per-query costs accumulate. Training large models is a significant one-time cost but is not performed continuously.
Water Use: The Less-Discussed Issue
Data centres use water for cooling. Microsoft, Google, and others have disclosed significant water consumption associated with AI workloads.
Microsoft reported that its global water consumption increased by 34% from 2021 to 2022, a period that included significant AI infrastructure investment. Google reported a 20% increase in water use in 2022. In some locations, data centres draw from water systems under stress.
A Microsoft research paper (2023) estimated that training GPT-3 consumed approximately 700,000 litres of fresh water — about the volume needed to produce 370 BMW cars or support a village for several months.
Water use is more geographically variable than energy use — a data centre in a water-stressed region (parts of the US Southwest, for example) has more significant local impact than one in water-rich northern Europe.
What AI Companies Are Doing
The major AI companies have made significant commitments, with varying levels of verifiability:
Microsoft: Committed to being carbon negative by 2030, water positive by 2030, and zero-waste by 2030. Has signed long-term renewable energy agreements. Microsoft also acknowledges that AI is making it harder to meet current targets due to higher energy demand than projected.
Google: Committed to operating on 24/7 carbon-free energy by 2030. Reports detailed environmental metrics annually. Is developing AI-powered tools to manage energy grid efficiency, which it argues will have net positive environmental impacts.
Anthropic (Claude): Has committed to measuring and reducing its environmental footprint, though disclosures are less detailed than Microsoft and Google. Runs on cloud infrastructure (Amazon AWS) whose environmental commitments apply.
OpenAI: Has made fewer public commitments than peers on environmental metrics. The company's rapid infrastructure scale-up has made it more difficult to make firm carbon commitments.
Honest assessment: The commitments are meaningful but are being made harder to meet by the pace of AI growth. The gap between stated ambitions and current trajectories is real and acknowledged by the companies themselves.
The Broader Tradeoff Question
Some researchers argue that AI may have net positive environmental effects that offset its energy use:
- AI is being used to accelerate materials science research for better batteries and solar cells
- Google DeepMind's AlphaFold dramatically accelerated protein structure understanding, with implications for developing drugs and studying biological processes related to climate adaptation
- AI optimisation tools are being applied to energy grid management, reducing waste
- Google reported that AI-based cooling management in their data centres reduced cooling energy by 30%
Whether these benefits outweigh AI's direct environmental costs is genuinely uncertain and depends on which applications scale and how quickly they displace higher-carbon alternatives.
Key takeaway: AI's environmental impact is real and worth taking seriously, but not uniquely or disproportionately harmful compared to other digital activities at individual use level. The concern is concentrated in training large models and in the aggregate scale of inference. Proportionate response matters more than either dismissal or panic.
A Proportionate User Perspective
What does "taking the environmental dimension seriously" actually mean in practice for someone using AI tools?
What probably matters more than your individual AI use:
- Your travel, particularly aviation
- Your diet, particularly beef consumption
- Home energy choices (heating, insulation)
- Consumer purchasing patterns
What is reasonable to do as an AI user:
- Be aware of the environmental cost, especially of very large tasks (generating hundreds of images, running long batch processes)
- Prefer AI providers with credible and verified renewable energy commitments where all else is equal
- Don't run AI tasks unnecessarily — the same discipline that makes you a good prompter (knowing what you actually need) also reduces waste
- Support policy and advocacy that pushes AI companies toward greater transparency and faster decarbonisation
What is not proportionate:
- Refusing to use AI tools on environmental grounds while regularly flying or driving
- Environmental guilt spiralling about a 10-message conversation
The environmental impact of AI is a legitimate systemic concern that deserves serious policy attention. It is not primarily an individual consumption decision in the way that diet and travel are.
Practice Task
Calculate your approximate AI-related energy use this week. If you had around 50 ChatGPT conversations averaging 10 exchanges each, that is roughly 0.5 kWh total — about the same as running a laptop for four or five hours. How does that compare to one other activity in your week? Calibrating these comparisons is more useful than either dismissing the concern or being immobilised by it.