Is AI bad for the environment?

It depends on which quantity you mean. By Google's own count, the median text prompt to its Gemini app used 0.24 watt-hours, 0.03 grams of CO2e and 0.26 millilitres of water in May 2025, which is very little. Across all data centres the International Energy Agency estimates 415 terawatt-hours in 2024, about 1.5 per cent of the world's electricity, and projects around 945 by 2030, naming AI as the most important driver. Both figures are sound as far as they go. They answer different questions, and neither covers training, other providers or the places where demand concentrates.

One text prompt costs about a quarter of a watt-hour#

Google's 2025 measurement paper is the most complete per-prompt count from an operator. It reports a median of 0.24 Wh, 0.03 gCO2e and 0.26 mL of water for a Gemini Apps text prompt. The boundary is stated: active accelerator, processor and memory energy, idle machines held for reliability, and data centre overhead. Counting only the active accelerator gives 0.10 Wh, so the fuller boundary is 2.4 times larger, which shows how far a headline number depends on what is left out. Google also reports that per-prompt energy fell 33-fold and emissions 44-fold over the twelve months to May 2025.

The task changes the answer by orders of magnitude. Luccioni, Jernite and Strubell measured 1,000 inferences on open models and found a mean of 0.002 kWh for text classification, 0.047 kWh for text generation and 2.907 kWh for image generation. For question answering, the most efficient task-specific models emitted 0.3 g of CO2e per 1,000 inferences and multi-purpose models 10 g. Those are 2023 measurements of open models, so they show the spread between tasks and not what a commercial chatbot uses now.

The grid total is where the growth shows#

The Energy and AI report from the IEA estimates data centres at around 415 TWh in 2024 and projects around 945 TWh by 2030 in its Base Case, just under 3 per cent of world electricity. Electricity use in accelerated servers, mainly driven by AI, is projected to grow 30 per cent a year. Data centres account for around one-tenth of global electricity demand growth to 2030, less than industrial motors, air conditioning or electric vehicles, and China and the United States account for nearly 80 per cent of the growth in data centre consumption. The agency also notes that data centres cluster in particular places, which makes them harder for a grid to absorb than their national share suggests. Emissions from data centre electricity are projected to rise from 180 Mt to 300 Mt by 2035 in the Base Case, still under 1.5 per cent of energy sector emissions.

Water follows the same pattern of small per unit and large in total. Li, Yang, Islam and Ren's modelling study projects 4.2 to 6.6 billion cubic metres of water withdrawal for AI demand in 2027. That is a projection under the authors' assumptions, and withdrawal is not the same as consumption.

Where the numbers stop#

• One operator, one service. The Google figures cover the median text prompt on Gemini Apps. They exclude training, users' own devices, external networking, and image and video prompts, and they cannot be rebuilt from outside the company.
• A median hides the tail. Long reasoning runs and automated agent workflows are not what the median describes, and no source cited here reports their footprint.
• Data centres are not AI. The IEA totals include conventional cloud, storage and streaming. AI is named as the main driver of growth, and its share of the total is not separated in the figures above.
• The projections are wide. The IEA's 2035 demand runs from 700 to 1,700 TWh across its cases.
• Local effects are outside a national share. A small share of a national grid can still be a large share of one region's, and none of the sources cited here measures that.
• Nothing here prices what AI might save elsewhere. A claim that AI reduces emissions in other sectors needs its own evidence.

A footprint is only bad against a comparison#

This section is interpretation, kept apart from the evidence above.

Two opposite mistakes are common. One treats your own chatbot use as the problem, when the best per-prompt count is a fraction of a watt-hour. The other treats the total as negligible because each prompt is, when the IEA expects the total to roughly double in six years. The two numbers are compatible, and the gap between them is volume.

That points at where the decision sits for an organisation. A person asking a chatbot a question is a small draw. A pipeline that calls a model thousands of times a day, or an agent that runs continuously, multiplies a small draw by a large number, and the sources cited here do not measure that case. The estate's wider argument is that any AI use should be judged by what human capability it builds or removes. The footprint adds a second column to the same ledger: a cost per use that is small alone and real at volume. That framing is mine, and none of the studies tested it.

Questions worth putting to a supplier before a large deployment#

• Which measurement boundary do your per-query figures use, and do they include idle capacity and cooling?
• Is the figure a median or a mean, and for which task and model?
• Where is the workload run, and what is the electricity mix and water source at that site?
• Do the figures include training, and is the model retrained or fine-tuned for us?
• Can the task be done by a smaller or task-specific model, given the spread Luccioni and colleagues measured?

Key sources

• Elsworth, C. and colleagues (2025). Measuring the environmental impact of delivering AI at Google Scale. arXiv:2508.15734, 21 August 2025. Google-authored. Graded entry.
• International Energy Agency (2025). Energy and AI. Graded entry.
• Luccioni, A. S., Jernite, Y. and Strubell, E. (2024). Power Hungry Processing: Watts Driving the Cost of AI Deployment?. ACM FAccT 2024. Graded entry.
• Li, P., Yang, J., Islam, M. A. and Ren, S. (2025). Making AI Less Thirsty. Communications of the ACM, accepted per the arXiv record. Graded entry.

Related SuperSkills research#

On the wider risk question, is AI dangerous and neither hype nor doom. On whether the use is worth its cost, does AI actually make people more productive. On automation at volume, should I let an AI agent act on my behalf.

About this research#

Rahim Hirji is the author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company. The IEA, Google, Luccioni and Li figures were read at source on 30 September 2026 and are graded in the evidence base. Nothing on this page is a SuperSkills coinage.

How this research works  ·  Reviewed quarterly  ·  Found an error? Tell me and it is corrected on the page.