Data Centre Energy & Water Statistics 2026
How much electricity data centres consume, how much water they use directly and indirectly, what a single AI query actually costs, and where the published figures disagree. Every statistic carries a named source and a publication date, and the calculations we perform ourselves state their inputs.
The water figure almost everyone quotes is the smaller one
US data centres consumed 66 billion litres of water directly in 2023. The water consumed on their behalf at power stations was roughly twelve times that. Both numbers come from the same Berkeley Lab report, but reporting overwhelmingly quotes the first and omits the second.
Original analysis by Paid Hosting · August 2026 · calculated from Berkeley Lab figures · methodology · full source list
📋 What’s in this page
Key Statistics at a Glance
The headline figures on data centre energy and water in 2026. Every number links to a named report in the sources list.
Global Data Centre Electricity Consumption
The International Energy Agency is the authoritative global source. Its most recent assessment was published in April 2026 and updates the landmark report it issued a year earlier.
- Global data centre electricity consumption was 485 TWh in 2025 and is projected to roughly double to around 950 TWh by 2030, reaching about 3% of global electricity demand.Source: IEA, Key Questions on Energy and AI, 16 April 2026
- Global data centre electricity demand grew 17% during 2025, in line with the IEA’s earlier projections.Source: IEA, 16 April 2026
- Electricity consumption by AI-focused data centres specifically grew 50% in 2025 — roughly three times the rate of data centres overall — and is projected to triple between 2025 and 2030.Source: IEA, 16 April 2026
- Capacity of “AI factories,” data centres purpose-built for AI, more than tripled in the 18 months to April 2026, measured by the IEA’s satellite-based tracking.Source: IEA, 16 April 2026
- Capital expenditure by the largest technology companies exceeded USD 400 billion in 2025 and is expected to rise a further 75% in 2026. The combined capex of five technology companies now exceeds global investment in oil and natural gas production.Source: IEA, 16 April 2026
- Major AI model providers reported a threefold increase in active users and a fivefold increase in revenue over the year to April 2026.Source: IEA, 16 April 2026
- In the IEA’s April 2025 assessment, data centres accounted for around 1.5% of world electricity consumption in 2024, or 415 TWh.Source: IEA, Energy and AI, 10 April 2025
- The United States accounted for 45% of global data centre electricity consumption in 2024, followed by China at 25% and Europe at 15%.Source: IEA, 10 April 2025
- Global data centre electricity consumption has grown around 12% per year since 2017 — more than four times faster than total electricity consumption.Source: IEA, 10 April 2025
- A typical AI-focused data centre consumes as much electricity as 100,000 households; the largest under construction will consume twenty times that.Source: IEA, 10 April 2025
- The IEA’s base case sees global data centre consumption reaching around 1,200 TWh by 2035, with a range across scenarios of 700 to 1,700 TWh.Source: IEA, 10 April 2025
US Data Centre Electricity Consumption
The definitive US figures come from Lawrence Berkeley National Laboratory, produced for Congress under the Energy Act of 2020 and published in December 2024.
US data centre electricity consumption, 2014–2028
Terawatt-hours per year. The 2028 figure is a scenario range, not a single forecast.
Source: Lawrence Berkeley National Laboratory, 2024 United States Data Center Energy Usage Report, 19 December 2024.
Chart may be reproduced with attribution to paidhosting.com- US data centres consumed 176 TWh in 2023, representing 4.4% of total US electricity consumption.Source: Berkeley Lab, 19 December 2024
- Consumption is projected at 325 to 580 TWh by 2028, or 6.7% to 12.0% of forecast US electricity consumption.Source: Berkeley Lab, 19 December 2024
- At an assumed 50% average capacity utilisation, that 2028 range translates to a total power demand of 74 to 132 GW.Source: Berkeley Lab, 19 December 2024
- US data centre electricity use was essentially flat at about 60 TWh between 2014 and 2016, continuing a trend of minimal growth observed since around 2010.Source: Berkeley Lab, 19 December 2024
- The turning point was 2017, when the server installed base began growing and GPU-accelerated servers for AI became a significant share of the data centre server stock. By 2018 consumption had reached 76 TWh, or 1.9% of US electricity.Source: Berkeley Lab, 19 December 2024
- US data centre electricity consumption tripled between 2014 and 2023, rising 3.03 times from 58 TWh to 176 TWh.Source: Paid Hosting calculation from Berkeley Lab figures (176 ÷ 58)
- Nearly half of US data centre capacity sits in five regional clusters, meaning local grid impacts are far more pronounced than the national share suggests.Source: IEA, 10 April 2025
- By 2030 the United States is projected to consume more electricity for data centres than for producing aluminium, steel, cement, chemicals and all other energy-intensive goods combined.Source: IEA, 10 April 2025
Berkeley Lab presents 2028 as a scenario range rather than a point estimate because three variables remain genuinely uncertain: how many GPUs ship each year, how intensively AI hardware in the installed base is actually used, and which cooling systems operators choose. The low and high ends differ by a factor of 1.8. Any article that quotes “580 TWh by 2028” as a prediction is quoting the top of a scenario range as though it were a central estimate.
Water Consumption: Direct and Indirect
Data centres consume water twice: directly, in cooling systems on site, and indirectly, at the power stations generating their electricity. The second figure is far larger than the first and is reported far less often.
US data centre water consumption, 2023
Billions of litres per year. Direct is on-site cooling; indirect is water consumed generating the electricity used.
Source: Lawrence Berkeley National Laboratory, 19 December 2024. Chart compiled by Paid Hosting.
Chart may be reproduced with attribution to paidhosting.com- US data centres directly consumed 66 billion litres of water in 2023, up from 21.2 billion litres in 2014 — a tripling over nine years.Source: Berkeley Lab, 19 December 2024
- The indirect water footprint of US data centres in 2023 was nearly 800 billion litres, consumed at power stations generating their electricity.Source: Berkeley Lab, 19 December 2024
- Indirect water consumption is therefore roughly 12.1 times the direct figure, and the combined total for 2023 is about 866 billion litres.Source: Paid Hosting calculation from Berkeley Lab figures (800 ÷ 66; 800 + 66)
- In metric-free terms, that is approximately 17.4 billion US gallons directly and 211.3 billion US gallons indirectly.Source: Paid Hosting conversion of Berkeley Lab figures
- Hyperscale and colocation facilities accounted for 84% of direct water consumption in 2023, while internal corporate data centres fell to 12%, driven by water efficiency improvements.Source: Berkeley Lab, 19 December 2024
- In 2014 the picture was reversed: internal data centres accounted for 64% of direct water consumption. By 2028 they are projected to fall to just 2% of the total.Source: Berkeley Lab, 19 December 2024
- Hyperscale data centres alone are projected to consume between 60 and 124 billion litres of water directly in 2028.Source: Berkeley Lab, 19 December 2024
- Electricity supplied to US data centres carries an average indirect water intensity of 4.52 litres per kWh, against a US average of 4.35 L/kWh for all electricity — meaning data centre grid mixes are about 3.9% more water-intensive than the national average.Source: Berkeley Lab, 19 December 2024; percentage calculated by Paid Hosting
The distinction matters geographically as much as arithmetically. Direct water is consumed where the data centre sits, so it competes with local supply and is what residents and local officials care about. Indirect water is consumed wherever the electricity is generated, which may be hundreds of miles away and in an entirely different watershed. A facility that cuts on-site cooling water to near zero by switching to air cooling will typically use more electricity doing so, which raises its indirect footprint. Reporting that treats the two as a single “water use” number obscures a real trade-off. Berkeley Lab also notes its indirect estimates exclude power purchase agreements and behind-the-meter generation, which could shift the figures significantly for individual facilities.
Energy and Water Per AI Query
Per-query figures are the most quoted and the most misused statistics in this field. The numbers below are the only ones published by a major provider with a stated methodology.
In August 2025 Google released measurements for the median Gemini Apps text prompt under two methodologies. Its comprehensive methodology, which accounts for all elements of serving AI globally, gives 0.24 Wh of energy, 0.03 gCO2e and 0.26 mL of water. A narrower methodology counting only active TPU and GPU consumption gives 0.10 Wh, 0.02 gCO2e and 0.12 mL. That is a 2.4-fold difference in energy and a 2.2-fold difference in water for an identical prompt, arising purely from what the measurement includes. Both are published by Google in the same document. When you see a per-query figure quoted anywhere, the first question is which boundary it used — and most reporting does not say.
- A median Gemini Apps text prompt uses 0.24 Wh of energy, emits 0.03 gCO2e and consumes 0.26 mL of water under Google’s comprehensive methodology.Source: Google, “Measuring the environmental impact of AI inference,” 21 August 2025
- Under a narrower methodology counting only active TPU and GPU consumption, the same prompt uses 0.10 Wh, emits 0.02 gCO2e and consumes 0.12 mL of water.Source: Google, 21 August 2025
- At 0.24 Wh per prompt, roughly 4,167 median text prompts consume one kilowatt-hour.Source: Paid Hosting calculation from Google’s figure (1,000 ÷ 0.24)
- Google’s own comparisons put a median text prompt at less energy than nine seconds of television and about five drops of water.Source: Google, 21 August 2025
- Over a recent twelve-month period, the energy footprint of the median Gemini text prompt fell 33-fold and its total carbon footprint fell 44-fold.Source: Google, 21 August 2025
- Energy use per AI task has been dropping by at least an order of magnitude annually in recent years — a rate of efficiency improvement the IEA describes as unprecedented in energy history.Source: IEA, 16 April 2026
- If every conventional internet search were performed as a simple AI text query, it would consume under 4 TWh annually — less than 1% of current total data centre consumption.Source: IEA, 16 April 2026
- Video generation, reasoning and agentic tasks can consume hundreds or thousands of times more energy per query than simple text generation.Source: IEA, 16 April 2026
Per-query energy is falling fast, and total AI energy consumption is rising fast. Both are correct, and quoting either alone produces a misleading picture. The IEA identifies three trends operating simultaneously: efficiency improvements, surging uptake, and changing model capabilities that unlock far more energy-intensive uses. A simple text query is now genuinely cheap; a video generation or agentic task is not, and usage is shifting toward the expensive end even as each individual task gets cheaper.
Power Density and Grid Impact
AI hardware concentrates far more power into the same physical space than conventional servers, which changes what data centres demand from the grid and from cooling systems.
- The power density of AI servers increased elevenfold between 2020 and 2025, and is set to rise a further fourfold by 2027.Source: IEA, 16 April 2026
- By 2027 a single server rack the size of a large refrigerator could have a peak power demand equivalent to 65 households.Source: IEA, 16 April 2026
- AI data centres experience repeated swings in server load exceeding 50% of rated capacity within a single second, which is why energy storage is becoming critical to their operation.Source: IEA, 16 April 2026
- AI data centres must remove heat equivalent to 30 natural gas boilers per server rack, and a single rack can weigh more than a pick-up truck.Source: IEA, 16 April 2026
- Around 20 to 25 GW of battery storage could be installed in data centres globally by 2030, potentially making them a grid asset rather than purely a load.Source: IEA, 16 April 2026
- Roughly 15 to 27 GW of onsite natural gas generation may power data centres by 2030, mostly in the United States. The US pipeline of such projects totals 47 GW, of which about one-fifth has begun land clearing or construction.Source: IEA, 16 April 2026
- Supplying reliable onsite gas-fired electricity to meet variable data centre load requires overbuilding generation infrastructure by 30% to 70% relative to demand.Source: IEA, 16 April 2026
- Global gas turbine orders surged 70% in 2025, highlighting chokepoints in energy technology supply chains.Source: IEA, 16 April 2026
- A shortage of high-bandwidth memory, integral to AI chip production, developed during 2026 and is anticipated to persist through at least the end of 2027.Source: IEA, 16 April 2026
Emissions
- Greenhouse gas emissions associated with US data centre electricity use totalled 61 billion kilograms of CO2 equivalent in 2023.Source: Berkeley Lab, 19 December 2024
- Electricity supplied to US data centres carries an average emission intensity of 0.34 kg CO2e per kWh, marginally below the US average of 0.35 kg/kWh for all electricity.Source: Berkeley Lab, 19 December 2024
- Emissions associated with data centres double in IEA projections, reaching around 350 million tonnes by 2035 — but still only about 2% of global electricity sector emissions by that date.Source: IEA, 16 April 2026
- Fossil-fuel electricity generators account for more than 99% of emissions associated with electricity generation.Source: US Energy Information Administration, cited in Berkeley Lab, 19 December 2024
- Well-documented AI use cases have the potential to save over 13 exajoules of energy by 2035, equivalent to 3% of global final energy consumption, if barriers to uptake are overcome.Source: IEA, 16 April 2026
- An AI-driven boost to economic growth could raise global energy demand by 1–4% in 2035 compared with trends without it.Source: IEA, 16 April 2026
How Reliable Are the Forecasts?
AI energy projections are often described as wildly unstable. On the evidence of the two most authoritative global assessments, published a year apart, they have been more stable than that reputation suggests.
In April 2025 the IEA projected global data centre electricity consumption of around 945 TWh by 2030. In April 2026, after a year of surging investment, revised datasets and a new satellite-tracking programme, it projected around 950 TWh — a change of roughly half a percent. The IEA states explicitly that its central projection “remains close to the trajectory set out in the IEA’s 2025 report.” What did move was the baseline: the 2024 estimate of 415 TWh became a 2025 estimate of 485 TWh, reflecting a year of actual 17% growth. The direction of travel was correctly anticipated; what remains genuinely uncertain is the period after 2030, where the IEA flags possible upside, and the scenario range for 2035 spans 700 to 1,700 TWh.
- The IEA’s projection for 2030 global data centre consumption changed from around 945 TWh (April 2025) to around 950 TWh (April 2026), a revision of about 0.5%.Source: Paid Hosting comparison of IEA reports dated 10 April 2025 and 16 April 2026
- The IEA states that data centre demand growth of 17% in 2025 was in line with its projections.Source: IEA, 16 April 2026
- Bottlenecks across energy equipment and chip manufacturing are reducing the likelihood of more aggressive near-term scenarios, despite booming investment and surging project pipelines.Source: IEA, 16 April 2026
- Data centre investment has grown too large to be funded from company balance sheets alone, making the pace of growth sensitive to market sentiment and financing conditions.Source: IEA, 16 April 2026
- Only 10% of global electricity consumption is covered by open electricity data policies, one reason independent verification of data centre energy claims remains difficult.Source: IEA, 16 April 2026
- The IEA calls for more systematic energy consumption disclosures from the technology sector to improve the robustness of AI energy demand forecasts.Source: IEA, 16 April 2026
Cite this page
https://www.paidhosting.com/data-center-energy-statistics/
Figures, charts and calculations on this page may be reproduced with attribution and a link to this page. Where a statistic is attributed to a third party, please cite that source directly. Press and data enquiries: [email protected]
Sources & Methodology
Every statistic on this page was read directly from the source named beside it, and carries that source’s publication date. Figures marked as Paid Hosting calculations are arithmetic performed on published source data, with the inputs stated in the citation so the working can be checked. Where two authoritative sources give different figures for the same quantity, we present both and explain the methodological reason rather than choosing one. Projections are labelled as projections and scenario ranges are presented as ranges, not as point forecasts. This page is updated as new data is published; the IEA and Berkeley Lab both update on annual cycles, and the next material revisions are expected in spring 2027 and whenever Berkeley Lab issues its next Congressional report.
- International Energy Agency — Key Questions on Energy and AI, published 16 April 2026 (CC BY 4.0). Current IEA assessment; supersedes the 2025 report for baseline and projection figures.
- International Energy Agency — Energy and AI, published 10 April 2025 (CC BY 4.0). Used for figures not restated in the 2026 report, and for the projection comparison.
- Lawrence Berkeley National Laboratory — 2024 United States Data Center Energy Usage Report, 19 December 2024. DOI 10.71468/P1WC7Q. Authors: Shehabi, Smith, Hubbard, Newkirk, Lei, Siddik, Holecek, Koomey, Masanet and Sartor. Produced for the US Department of Energy under the Energy Act of 2020.
- US Department of Energy — announcement of the Berkeley Lab report, 20 December 2024.
- Google — “Measuring the environmental impact of AI inference,” Google Cloud blog, 21 August 2025.
- US Energy Information Administration and US Environmental Protection Agency — power plant generation, water consumption and emissions datasets, as used within the Berkeley Lab analysis.
A note on the Google figures: they are self-reported by a company with a commercial interest in AI being perceived as efficient, and they cover Google’s own infrastructure rather than the industry. We include them because they are the only per-query measurements published by a major provider with a documented methodology, and because Google published both a narrow and a comprehensive number, which is more transparency than any comparable disclosure offers. They should be read as one provider’s figures, not as an industry average.