Electricity, water and carbon.
Research reviewed 13 September 20262024 estimate · includes AI and non-AI services
All data centres Other uses
Each square = 1%. This is not AI’s share alone.
The remaining 98.5% includes factories, homes, shops, hospitals, farms and transport. A sector breakdown is shown alongside.
Final electricity use · 2022 · UN global energy balance
1 TWh = 1 billion kWh. Other = agriculture and other uses. Shares calculated from rounded UN figures. Data centres are already inside these sectors; do not add 1.5% to this chart.
The two charts answer different questions: all electricity demand in 2024 versus final use by sector in 2022. Final use excludes grid losses and the energy industry’s own electricity use.
Where do AI, streaming, gaming and social media fit? →
Global electricity · 2024 estimates · TWh
Lighting excludes industry and agriculture. End uses can overlap sector accounts; these are scale comparisons, not slices to add to the UN chart.
U.S. electricity retail sales · 2023 · TWh
Retail sales total 3,874.3 TWh; another 136.9 TWh was used directly by generators. EV charging is distributed across sectors. All U.S. data centres used an estimated 176 TWh in 2023 and overlap these categories.
All data centres worldwide · TWh per year
Redrawn from IEA figures: ¹2025 report; ²2026 update. Stripes mark the central projection, not a measured outcome.
2030 central projection · AI-focused facilities
IEA projects roughly half at AI-focused facilities. That is a building classification, not a global measurement of individual AI workloads.
Shares of global electricity used by all data centres · 2024
U.S.: 45% of the global data-centre total, not 45% of U.S. electricity.
Bubble areas show shares; positions are regional anchors. Other = remainder. Geography: Natural Earth / world-atlas ↗.
Local grid capacity and electricity costs still matter where facilities are concentrated. IEA ↗
All data-centre electricity · worldwide · 2024
This measures equipment, not AI’s exact share. Some AI runs on ordinary servers; accelerated servers also run non-AI science.
Redrawn from IEA Box 2.1. The 85% remainder includes infrastructure supporting AI, so it is not “85% non-AI.”
Streaming and social feeds can also use AI. These labels overlap; they cannot form a clean AI-versus-everything-else pie chart.
IEA finds no comprehensive global electricity breakdown by application. Data traffic is not an electricity meter.
One hour · PS5 CFI-2016 · console electricity only
Calculated as measured watts × one hour. Sony tests, January 2024. Excludes TV, router, networks and data centres; not a cloud-gaming estimate.
| Activity | Reported energy | Boundary |
|---|---|---|
| Gemini text prompt | 0.24 Wh / prompt | Google median · May 2025; serving system |
| Streamed TV | 112.7 Wh / device-hour | UK model · 2021; viewing devices, networks and data centres |
A prompt and an hour of viewing are different activities. Neither number establishes total global use, and the Gemini result does not cover every model or task.
Social media uses storage, networking and computing, with AI increasingly mixed into those services. The IEA does not provide an audited global social-media electricity share. Platform-wide sustainability reports also include offices, hardware and other products.
A precise AI-versus-social-media ranking would require consistent workload-level reporting. We leave that share unquantified.
All U.S. data centres · AI and non-AI services · 2023 estimates
Water consumed · billion litres per year
The second figure is nearly 800 billion litres, a grid-based estimate. Both cover all U.S. data centres.
Consumed = not returned to the immediate environment, often because it evaporates. USGS ↗
Consumption · billion litres per year
Context, not a same-year ranking. Textile water includes global supply chains; data-centre cooling counts onsite use. Different years, populations and boundaries prevent a like-for-like efficiency comparison.
A litre consumed in a dry watershed has different consequences from a litre consumed where water is plentiful. Site and season matter. Research ↗
Water withdrawn is taken from a source. Water consumed is not returned to the immediate environment, for example through evaporation.
Contiguous U.S. · average, 2010–2020
Consumed Returned
Redrawn from USGS consumption and withdrawal estimates, rounded. Water quality and where it returns also matter.
The EEA textile estimate includes fibre production and manufacturing worldwide; about 85% of its water footprint occurs outside Europe. It excludes rainwater.
L’Oréal identifies water use in farming ingredients, manufacturing and rinsing products such as shampoo. That is a wider boundary than cooling a data centre. Its Waterloop factories reuse treated water for industrial processes; this is a company example, not a figure for the entire cosmetics industry.
Google’s reported median Gemini Apps text prompt, May 2025. The water estimate excludes electricity generation. A product-specific median is not a universal prompt footprint or a basis for multiplying out total usage.
Cooling water is generally reported for a facility or company, not separately for each application. These data-centre totals cannot reliably be divided into AI, social media, streaming and gaming. A facility’s cooling design and local grid also change the result.
Of fuel-combustion CO₂, from electricity supplying all data centres in 2024.
About 180 million tonnes CO₂ · AI and non-AI workloads
Excludes hardware, construction and backup generation. This is not AI’s complete carbon footprint.
Of energy- and process-related emissions comes directly from steel and cement production.
IEA 2025 report. This includes industrial process emissions and excludes purchased electricity; the data-centre figure uses electricity emissions and a fuel-combustion denominator. Context, not a like-for-like lifecycle ranking.
Company-wide emissions · million tonnes CO₂e
Annual renewable-energy matching does not eliminate the company’s emissions.
Operations, electricity and supply chains across all of Google, not AI alone. GHG Protocol-aligned totals with market-based electricity accounting; chart values rounded.
61 million tonnes CO₂e: estimated electricity-related emissions from all U.S. data centres in 2023, using regional grids.
1,000 kWh →
Illustration only. Electricity × intensity ÷ 1,000 = kg CO₂e; excludes hardware and other lifecycle emissions.
BLOOM’s final training used an estimated 24.7 tonnes CO₂e from dynamic power; adding equipment and other operations raised the estimate to 50.5 tonnes. The larger number includes the smaller and excludes lifetime inference.
A 2024 model projects 1.2–5.0 million tonnes of cumulative GenAI e-waste for 2020–2030. Longer use and circular strategies lower it by 16–86% across scenarios. These are possible futures, not waste already produced.
Practical responses · recommendations based on the linked research
Test smaller, task-specific models and tune GPU power limits. An MIT-led training example saved about 15% energy with a small runtime increase; savings depend on the workload.
Compress models, keep useful equipment in service and measure energy per completed task. MIT’s TinyChat reports memory and speed gains; these are not equivalent carbon savings.
Shift flexible workloads around grid conditions. Buy additional clean supply with hourly matching; Princeton modelling finds stronger emissions benefits than annual matching alone.
Use recycled water or closed-loop cooling where suitable. Publish local withdrawal, consumption and water stress; include any electricity tradeoff.
Report operations, electricity and supply-chain emissions separately. Reduce them before relying on removals for residual emissions; publish absolute totals alongside efficiency gains.
Total energy: of the starting level
50% less per task × 3× as many tasks = 50% more total energy.
Arithmetic illustration, not a prediction. Total = energy per task × task count.
Company disclosures · progress, limits and methods worth sharing
Annual electricity matched with renewable purchases in 2025.
Company-wide emissions still totalled 18.85 million tonnes CO₂e on its GHG Protocol-aligned measure. Its separate “ambition-based” total uses exclusions and additional instruments.
Practice: shifting flexible computing workloads around grid conditions.
Of capacity served by closed-loop cooling at the featured Fairwater site.
A site design disclosure, not a measured fleet-wide water saving. Initial fill and other water uses remain.
Limit: Microsoft reported total emissions up 25% year on year in its 2026 report, despite annual renewable matching.
Data centres using reclaimed water for cooling in its 2025 report.
Nearly 850 million litres of potable water substituted. That is not 850 million litres less total consumption.
Practice: reserve drinking-quality water by using treated wastewater where feasible.
Microsoft’s next-generation cooling announcement projects over 125 million litres avoided per facility per year against its FY2024 average. It is a design comparison, not an observed fleet saving, and comes with a small electricity increase. Its schedule described 2026 pilots and new sites from late 2027.
Paraphrased views, not quotations or a claim of consensus.
Coordinate computing with the grid so flexible work can follow available clean power.
Efficiency per task can improve while growing use raises the total footprint.
Primary university studies, government and intergovernmental reports, ethical-AI research and company disclosures published from September 2021–September 2026. This is a curated guide, not an exhaustive systematic review. Older underlying data are labelled.
Estimates, forecasts, surveys and observed results are kept distinct. Company claims are identified as disclosures. Charts are redrawn from cited data, with rounded values and calculated shares labelled; no missing years are interpolated.
Industry comparisons show scope and date because electricity, onsite water and supply-chain footprints are different measures. Findings do not automatically generalize beyond the studied people, products or settings.
Each “Notes” button gives the finding, scope and limits.
Download source ledger ↗International Energy Agency
International Energy Agency
Lawrence Berkeley National Laboratory
Google researchers
Luccioni, Viguier & Ligozat
Wang et al. · Nature Computational Science
International Labour Organization
World Economic Forum
Stanford Digital Economy Lab
Brynjolfsson, Li & Raymond · Quarterly Journal of Economics
Stanford HAI
Otis et al. · Management Science / Berkeley Haas
International Telecommunication Union
Lund University · The Lancet
Hofmann et al. · Nature
Costello, Pennycook & Rand · Science / MIT Sloan
National Institute of Standards and Technology
Google · Michael Terrell
Microsoft · Brad Smith & Melanie Nakagawa
Microsoft · Steve Solomon
Microsoft · Scott Guthrie
Amazon · Tyler Greenawalt
McDonald et al. · MIT Lincoln Laboratory and collaborators
MIT Han Lab · Yang et al.
Luccioni, Jernite & Strubell · FAccT
Xu et al. · Princeton / Joule
Science Based Targets initiative
Science Based Targets initiative
Harvard Business School · Rachel Layne / McFowland et al.
Ada Lovelace Institute & Alan Turing Institute
University of Melbourne & KPMG · Gillespie, Lockey and colleagues
Andrew A. Chien · Stanford Energy Seminar
MIT Energy Initiative · Emma Strubell and symposium participants
AI Now Institute
International Energy Agency
U.S. Energy Information Administration
U.S. Geological Survey
European Environment Agency
United Nations Statistics Division
International Energy Agency
Sony Interactive Entertainment
Ofcom · commissioned research by Carnstone
LinkedIn Economic Graph
Cui, Demirer, Jaffe, Musolff, Peng & Salz · Microsoft Research
Penn Wharton Budget Model · University of Pennsylvania
Nearing et al. · Google Research and collaborators · Nature
EMBL European Bioinformatics Institute
International Energy Agency
No matching sources. Try another word or topic.