Energy and Environment

Ask a chatbot to summarise an email and the reply seems to come from nowhere. It does not. Somewhere — perhaps in Virginia, Dublin, or a new industrial park outside Johor Bahru — racks of chips switched on for a fraction of a second, fans and pumps carried their heat away, and a power station somewhere burned gas or caught sunlight to supply them. Is that cost trivial or alarming? The honest answer is “both, depending on what you count”. Learning what to count is the point of this page.

The physical stack behind a prompt

“The cloud” is a metaphor for very solid things. A data centre is a warehouse of servers. For AI, most of the work runs on accelerator chips such as GPUs or Google’s TPUs. Peel the stack from the outside in.

Two numbers describe a building’s efficiency. Power usage effectiveness (PUE) divides total facility energy by the energy reaching the computers; 1.0 would mean no overhead at all. The IEA’s 2025 report Energy and AI says cooling takes about 7% of electricity in efficient “hyperscale” data centres but over 30% in less efficient enterprise ones. Water use is measured in a similar way, as litres per kilowatt-hour of computing.

What is measured, and what is estimated

The most careful global figures come from the IEA, an intergovernmental agency. Its Energy and AI report (April 2025) estimated that data centres used about 415 terawatt-hours (TWh) of electricity in 2024, around 1.5% of world electricity As of 2025. The United States accounted for about 45% of that and China about 25%. Consumption had grown about 12% a year over the previous five years.

Speculative In its central “Base Case”, the IEA projects data-centre demand roughly doubling to about 945 TWh by 2030 — slightly more than Japan’s whole electricity use today and just under 3% of the world’s — with AI the most important driver. Its other scenarios span a wide range: by 2035, a “Headwinds” case plateaus around 700 TWh, while a “Lift-Off” case passes 1,700 TWh.

Notice what kind of number each one is. The 2024 figure is itself an estimate built from chip shipments, server counts and company reports, because most operators do not publish per-site energy use and almost none split out AI from other computing. The 2030 figure is a scenario, not a measurement. Good reports say this openly; headlines often do not.

Training, inference and the cost of one prompt

AI uses energy twice. Training happens once per model and is enormous: a widely cited 2021 study by David Patterson and colleagues at Google and Berkeley estimated that training GPT-3 used about 1,287 megawatt-hours, roughly the yearly electricity of more than a hundred American homes. Inference — answering prompts — is tiny per request but happens billions of times. A 2022 follow-up by Patterson’s team reported that across Google in 2019–2021 about three-fifths of machine-learning energy went to inference and two-fifths to training. As usage grows, inference tends to dominate.

Per-prompt estimates have varied tenfold, and the differences are mostly about assumptions, not dishonesty. Open the cards.

Consensus Researchers broadly agree that the type of task matters more than the brand of chatbot. A 2023 study by Sasha Luccioni, Yacine Jernite and Emma Strubell measured 88 models and found general-purpose generative models “orders of magnitude” more costly than small task-specific ones, with image generation especially energy-hungry. A long reasoning answer, a generated video or an agent that runs for an hour can cost many times a short text reply. A single median figure is useful; it is not your typical day.

Water and the carbon built into hardware

Data centres use water in two places: on site, when evaporative cooling towers carry heat away, and off site, at power stations that need water for cooling. Pengfei Li, Shaolei Ren and colleagues (first posted 2023, published in Communications of the ACM) estimated that training GPT-3 in Microsoft’s US data centres could have directly evaporated about 700,000 litres of fresh water, and projected that global AI demand could account for 4.2–6.6 billion cubic metres of water withdrawal in 2027 Speculative. What matters locally is where and when: a litre drawn in a rainy region is not the same as a litre drawn during a drought. Air-cooled designs use less water but more electricity — a genuine trade-off, not a free win.

Embodied carbon is the emissions released making hardware and buildings before they compute anything. When Luccioni, Sylvain Viguier and Anne-Laure Ligozat studied the 176-billion-parameter BLOOM model (2022), the chips’ electricity during final training came to about 24.7 tonnes of CO₂-equivalent, but counting manufacturing, idle machines and the wider infrastructure roughly doubled the total, to about 50.5 tonnes. Accelerators are often replaced within a few years, so embodied carbon is a growing share — and on a very clean grid it can be the largest share.

Efficiency gains and the rebound question

Each chip generation does more work per watt, and software tricks — smaller models, quantisation, better scheduling — cut the energy per answer further. Google’s reported 33-fold fall per prompt in a year is an extreme example. So why do total figures keep rising?

The answer may be the Jevons paradox. In The Coal Question (1865) the economist William Stanley Jevons observed that more efficient steam engines had increased Britain’s total coal use, because cheaper power found countless new uses. When the Chinese lab DeepSeek released an unusually efficient model in January 2025, Microsoft’s chief executive Satya Nadella posted “Jevons paradox strikes again!”, predicting that cheaper AI would mean far more AI. Google’s own 2025 environmental report shows the tension: its data-centre electricity use rose 27% in 2024 while it reported a 12% fall in data-centre energy emissions, thanks to efficiency and clean-energy contracts As of 2025.

Debated Whether efficiency will ultimately shrink or grow AI’s footprint is genuinely open.

Malaysia’s data-centre boom

Malaysia, and Johor in particular, has become one of the fastest-growing data-centre markets in Asia, helped by land, power, fibre links and proximity to Singapore, which paused new data-centre approvals in 2019 before reopening selectively. The broader policy picture is on AI in Malaysia; here are the resource facts.

In a written reply to Parliament reported on 20 October 2026, the government said data-centre projects had secured approval for 12.3 gigawatts of electricity supply capacity as of August 2026, and projected peak data-centre demand in Peninsular Malaysia of 8.39 GW by 2030, with Johor about 4.42 GW of it As of October 2026. For comparison, approved capacity is not the same as actual use: operators reserve more than they draw at first. On water, the reply said data centres in Selangor, Johor and Negeri Sembilan had approved allocations of 445.04 million litres a day between 2025 and 2030, of which facilities already operating accounted for 63.90 million litres a day.

Policy has tightened as the numbers grew As of October 2026:

• Tariffs: an Ultra-High Voltage electricity tariff has applied since 1 July 2025 to data centres supplied above 50 kV, and a higher connected-load charge penalises booking capacity that is not used — both intended to stop costs being passed to other consumers.
• Renewables: data centres are required to source at least 30% of demand capacity from renewable energy, for example through the Corporate Renewable Energy Supply Scheme (CRESS).
• Water: a higher water tariff for data centres took effect on 1 August 2025, and the national water regulator SPAN has issued efficiency guidelines.
• Johor: in late 2025 the state reportedly stopped approving the water-hungry categories it calls Tier 1 and Tier 2 data centres. In March 2026 the Menteri Besar, Onn Hafiz Ghazi, said only air-cooled facilities in industrial zones using reclaimed water would be approved.

Debated Whether the boom is good for Malaysians is argued over. Supporters, including state and federal governments, point to investment, construction jobs and a chance to move up the digital value chain. Critics, including environmental groups and some economists, note that data centres employ few people once built, that much of the grid is still fossil-fuelled, and that water and grid upgrades are partly a public cost. Both sides agree the outcome depends on rules — on clean supply, water and who pays for the grid.

Keeping perspective, and what you can do

How big is AI’s footprint next to other things? The bars below use shares of global carbon dioxide emissions. They are drawn relative to aviation, a sector people often compare with.

• Aviation≈ 2.5% of global CO₂
  Hannah Ritchie, Our World in Data (2020), using 2018 data.
• Data centres, 2035 (“Lift-Off” scenario)≈ 1.4%
  IEA, Energy and AI (2025). A forecast, not a measurement.
• Data centres, 2035 (Base Case)≈ 1%
  IEA (2025). A forecast; about 300 million tonnes.
• All data centres today, AI and non-AI≈ 0.5%
  IEA (2025): about 180 million tonnes from electricity, of roughly 35,000 million tonnes from fuel combustion in 2024.

The denominators differ slightly (all CO₂ versus combustion CO₂), so read the bars as orders of magnitude, not precise ratios.

This guide’s framing Two things are true at once. Globally, data centres are a small slice of emissions, and the IEA expects them to account for less than 10% of the growth in world electricity demand to 2030 — air conditioning, electric vehicles and industry matter more. Locally, they can be huge: a cluster of sites can strain one state’s grid and water supply, as Johor shows. And the trend, not the level, is what worries planners. Engineers who want the economics of serving models can read MLOps and Serving; the compute behind training is on Pretraining and Scaling Laws.

What can you do? Individual prompts are not where most of the impact lies, so guilt is less useful than good habits — and builders and organisations have far more leverage than users.

1. Use the right toolA search engine, calculator or spreadsheet is cheaper than a chatbot for a fact or a sum. Save heavy features — image and video generation, long “deep research” runs — for when they are worth it.
2. Ask well, onceA clear prompt avoids five rounds of regeneration. That saves your time, which is the main benefit, and a little energy.
3. Builders: choose the smallest model that worksTest whether a small or task-specific model meets your quality bar before defaulting to the largest. Cache repeated answers and set sensible output lengths.
4. Builders: choose where and whenCloud regions differ greatly in grid carbon intensity; batch jobs can run when and where power is cleaner. Ask your provider for energy and water figures.
5. Organisations: measure and discloseCount AI use in sustainability reports, include embodied carbon, and prefer suppliers that publish methods rather than single flattering numbers.
6. Citizens: watch the rulesThe largest levers are public: clean-power requirements, water permits and who pays for grid upgrades. These are decided by governments and regulators, and they respond to informed voters.

Key terms on this page

Sources & further reading

• International Energy Agency, Energy and AI (2025), especially “Energy demand from AI” and “AI and climate change”.
• Elsworth, C. et al. (Google), Measuring the environmental impact of delivering AI at Google Scale (2025).
• Li, P., Yang, J., Islam, M. A. & Ren, S., Making AI Less “Thirsty” (2023; Communications of the ACM, 2025).
• Luccioni, A. S., Viguier, S. & Ligozat, A.-L., Estimating the Carbon Footprint of BLOOM (2022).
• Luccioni, A. S., Jernite, Y. & Strubell, E., Power Hungry Processing: Watts Driving the Cost of AI Deployment? (2023).
• Patterson, D. et al., Carbon Emissions and Large Neural Network Training (2021).
• Patterson, D. et al., The Carbon Footprint of Machine Learning Training Will Plateau, Then Shrink (2022).
• You, J. (Epoch AI), How much energy does ChatGPT use? (2025).
• de Vries, A., “The growing energy footprint of artificial intelligence”, Joule (2023).
• Ritchie, H., Climate change and flying: what share of global CO2 emissions come from aviation?, Our World in Data (2020).
• The Star, “12.3GW energy capacity approved for data centres as of August, Parliament told” (20 October 2026).
• Jevons, W. S., The Coal Question (1865).