Live estimate, not a meter reading
AI energy consumption: electricity and water use of AI and data centers
There is no global meter for AI data centers. The counter spreads the annual figures published by the International Energy Agency (IEA) evenly across every second of the year. Below it, ten questions answer how much electricity and water AI uses, each with the origin of every number.
Electricity
167.715729TWh
Water
58.670252billion L
Facilities with large numbers of accelerator chips for training and inference. That is the IEA’s definition. Water: on-site cooling only, meaning water that does not return to its source after use. AI water is not measured separately. It is derived from the IEA water figure using AI’s share of data-center electricity. How the counter calculates
How much energy does AI use worldwide?
In short
About 155 TWh in 2025, which is 32% of the electricity of all data centers (485 TWh). By 2030 the IEA expects around 465 TWh for AI data centers.
Electricity of AI data centers, 2024 to 2030
Solid columns are published or derived values, hatched columns sit between them on even growth. The current year is outlined.
- 103
- 155
- 193
- 241
- 300
- 373
- 465
- published or derived
- interpolated
- current year
2024 is derived from the IEA’s 50% growth in 2025. For all data centers, the summary of the IEA report puts 2030 at around 950 TWh, while Table A.1 of the same report lists 945 TWh. Both describe the same scenario; the counter uses the rounded value.
All annual figures from 2023 to 2030
| Year | AI data centers | Other | All data centers | AI share | Total water |
|---|---|---|---|---|---|
| 2023 | no value | no value | no value | no value | 560 billion Lestimate |
| 2024 | 103 TWhderived | 313 TWh | 416 TWhestimate | 24.8% | 624.4 billion Linterpolated |
| 2025 | 155 TWhread from a chart | 330 TWh | 485 TWhestimate | 32% | 696.2 billion Linterpolated |
| 2026 | 193 TWhinterpolated | 362 TWh | 555 TWhinterpolated | 34.8% | 776.3 billion Linterpolated |
| 2027 | 241 TWhinterpolated | 394 TWh | 635 TWhinterpolated | 37.9% | 865.6 billion Linterpolated |
| 2028 | 300 TWhinterpolated | 426 TWh | 726 TWhinterpolated | 41.3% | 965.2 billion Linterpolated |
| 2029 | 373 TWhinterpolated | 457 TWh | 830 TWhinterpolated | 44.9% | 1,076.2 billion Linterpolated |
| 2030 | 465 TWhprojection | 485 TWh | 950 TWhprojection | 48.9% | 1,200 billion Lprojection |
How much electricity do data centers use in Germany?
In short
All German data centers used about 21.3 TWh in 2025. That is arithmetically 4.4% of worldwide data-center electricity. AI data centers alone use 7.3 times as much worldwide.
Germany 2025 against the world totals
Different surveys: Bitkom/Borderstep for Germany, IEA for the world. Their methods have not been fully reconciled.
Data centers in Germany, 2015 to 2025
All data centers and smaller IT installations, modeled by Borderstep. Bitkom estimated 2025 while the year was still running.
By the IEA rule of thumb, 100 MW of continuous power equals about 100,000 households. AI electricity in 2026 therefore matches about 22 million households. That is an order of magnitude, not a site count. The direct cooling water the counter attributes to AI fills about 21,615 Olympic pools of 3.125 million liters each in 2026.
How much energy does ChatGPT use per query?
In short
A text query uses 0.24 to 0.34 Wh according to the providers. An AI agent task with reasoning can reach 50 Wh in an IEA model calculation, about 150 times as much.
| Reference value | per item | times 100 |
|---|---|---|
| ChatGPT, 2025 average per Sam Altman (OpenAI)company figure | 0.34 Wh | 34 Wh |
| Gemini Apps, May 2025 mediancompany figure | 0.24 Wh | 24 Wh |
| IEA agent task without reasoningmodeled | 1.14 Wh | 114 Wh |
| IEA agent task with reasoningmodeled | 50 Wh | 5.00 kWh |
| ChatGPT water according to Sam Altman (0.000085 US gallons)company figure | 0.32 ml | 32.2 ml |
| Gemini Apps, direct cooling watercompany figure | 0.26 ml | 26 ml |
Example calculation: published reference value times the count entered. Not a measured consumption and not a personal footprint. Training and your own device are excluded. The values rest on different methods and do not support a ranking.
Source: OpenAI, Google Cloud, IEA 2026
How much water does AI use?
In short
All data centers consumed about 560 billion L in 2023, and 2030 is projected at 1,200 billion L. Only one quarter of that is on-site cooling, two thirds sit in the energy supply. The IEA publishes no AI-only figure.
Water consumption in 2023 and 2030 by origin
For 2030 the 2023 mix is carried forward. That is an assumption, not an IEA figure.
- Direct cooling (billion L)
- Energy supply
- Hardware manufacturing
Energy supply covers primary energy and electricity generation, including producing, processing, and transporting energy carriers. The rounded shares imply roughly 140 billion L of direct cooling and 373 billion L for energy supply in 2023. These are not separate global measurements.
Far more water is withdrawn: IEA Figure 5.27 suggests about 5,200 billion L of withdrawals in 2023, 9.3 times the consumption. That reading is uncertain. Consumption means the water does not return to its source after use. Only consumption enters the counter.
How much water does ChatGPT use per query?
In short
Sam Altman gives 0.000085 US gallons per ChatGPT query, about 0.32 ml. For a Gemini Apps text prompt, Google gives a May 2025 median of 0.26 ml of direct on-site cooling water. The methods differ, so the two figures are not directly comparable.
Milliliters of water per text query
Altman states no water boundary, Google counts on-site cooling only.
Altman’s post also gives 0.34 watt-hours of electricity as the average, but no measurement protocol and no precise water-accounting boundary. The Gemini figure is not a ChatGPT measurement.
How much will AI energy use grow by 2030?
In short
Electricity of all data centers rises from 485 TWh to around 950 TWh, so it roughly doubles. The IEA scenarios for 2030 range from 833 to 1,008 TWh. AI electricity grows by about 25% per year, much faster than the rest.
Four IEA scenarios for 2030
All data centers, meaning AI and other workloads. Possible pathways, not confidence intervals and not an AI-only range.
2025 for comparison: 485 TWh.
Annual growth between the anchor years
Average annual rate from the published years. The 50% AI increase in 2025 is an IEA figure.
How much energy do data centers use in the US and China?
In short
The US accounted for about 45% of the electricity of all data centers worldwide in 2024, China for 25% and Europe for 15%. For the US, Lawrence Berkeley National Laboratory (LBNL) expects 649 TWh in its 2030 reference case, which would be 11.8% of US electricity use.
Share of the electricity of all data centers, 2024
Worldwide. Color and flag show the region.
Data centers in the US, 2014 to 2030
All data centers. In brackets the share of US electricity use.
The range is wide. LBNL’s 2030 scenarios run from 521 to 843 TWh, the Electric Power Research Institute (EPRI) gives 380 to 790 TWh. The IEA’s world figures and the US figures rest on different models and cannot be added up directly.
How much electricity do Google, Microsoft and Meta use?
In short
Google used 43.59 TWh of electricity in 2025, 2.4 times its 2021 figure. Microsoft reached 37.03 TWh in fiscal 2025, 2.7 times. The figures cover the whole company; Amazon publishes no total.
Electricity use of the companies, 2021 and latest report year
Whole company. Microsoft’s fiscal year ends on June 30.
Water withdrawal of the companies
Withdrawal, not consumption. Scope and year are given per row.
Nuclear agreements by capacity
For existing plants and restarts the bars show supply volumes or plant capacity, for new builds project targets that could come online in the 2030s.
Meta’s values are parts of its figure of up to 6.6 GW of new and existing clean energy by 2035. The restarts of Three Mile Island Unit 1 (Crane) and Duane Arnold still need approval from the US Nuclear Regulatory Commission (NRC).
Why do other figures differ so much?
In short
Because every estimate counts something different. Epoch AI models a typical GPT-4o query with 500 output tokens at about 0.3 Wh, roughly one tenth of the older, widely quoted 3 Wh estimate.
Model and prompt
Long inputs and reasoning need more electricity than a short text answer.
Period and region
The electricity mix, cooling technology and local climate mainly change the water demand.
System boundary
Some count accelerator electricity only, others add cooling, the wider infrastructure, electricity generation or upstream water use.
Withdrawal or consumption
Withdrawal is not consumption. In 2023 about 5,200 billion liters were withdrawn against 560 billion liters consumed.
How reliable are these figures?
In short
They are estimates. The IEA models them itself, and nothing is metered worldwide. Metered values exist only for single countries such as Ireland. That is why every number on this page carries a label showing where it comes from.
Every published value with its origin
Each number this page rests on. Everything else is calculated from them.
25 of 25 values6 of 25 values5 of 25 values2 of 25 values1 of 25 values5 of 25 values3 of 25 values2 of 25 values1 of 25 values
| Value | What | Type | Source | |
|---|---|---|---|---|
| 416 TWh | estimateElectricity of all data centers in 2024 | Electricity of all data centers in 2024 | estimate | IEA, Key Questions on Energy and AI (2026), Table A.4 |
| 485 TWh | estimateElectricity of all data centers in 2025 | Electricity of all data centers in 2025 | estimate | IEA, Key Questions on Energy and AI (2026) |
| around 950 TWh | projectionElectricity of all data centers in 2030, rounded executive summary (used by the counter) | Electricity of all data centers in 2030, rounded executive summary (used by the counter) | projection | IEA, Key Questions on Energy and AI (2026) |
| 945 TWh | projectionThe same 2030 scenario as a table value, also used by Our World in Data | The same 2030 scenario as a table value, also used by Our World in Data | projection | IEA, Key Questions on Energy and AI (2026), Table A.1 |
| 833 to 1,008 TWh | projectionFour scenarios for all data centers in 2030: headwinds 833, high efficiency 868, base case 945, lift-off 1,008 TWh | Four scenarios for all data centers in 2030: headwinds 833, high efficiency 868, base case 945, lift-off 1,008 TWh | projection | IEA, Key Questions on Energy and AI (2026), Table A.1, p. 108 |
| 155 TWh | read from a chartElectricity of AI data centers in 2025, read from IEA Figure 1.5 (plus 330 TWh without an AI focus) | Electricity of AI data centers in 2025, read from IEA Figure 1.5 (plus 330 TWh without an AI focus) | read from a chart | Our World in Data (2026) |
| +50% | estimateGrowth of AI electricity in 2025. This implies about 103.3 TWh for 2024 | Growth of AI electricity in 2025. This implies about 103.3 TWh for 2024 | estimate | IEA, Key Questions on Energy and AI (2026) |
| around 465 TWh | projectionElectricity of AI data centers in 2030 | Electricity of AI data centers in 2030 | projection | IEA, Key Questions on Energy and AI (2026) |
| 560 billion L | estimateWater consumption of all data centers in 2023 | Water consumption of all data centers in 2023 | estimate | IEA, Energy and AI (2025) |
| 1,200 billion L | projectionWater consumption of all data centers in 2030 | Water consumption of all data centers in 2030 | projection | IEA, Energy and AI (2025) |
| 25 / 67 / 8% | estimateSplit in 2023: direct cooling, energy supply, hardware (one quarter, two thirds, remainder) | Split in 2023: direct cooling, energy supply, hardware (one quarter, two thirds, remainder) | estimate | IEA, Energy and AI (2025) |
| about 5,200 billion L | read from a chartWater withdrawals of all data centers in 2023, an uncertain reading of Figure 5.27 | Water withdrawals of all data centers in 2023, an uncertain reading of Figure 5.27 | read from a chart | IEA, Energy and AI (2025) |
| 100 MW ≈ 100,000 | rule of thumbHouseholds matching the electricity of a typical AI data center | Households matching the electricity of a typical AI data center | rule of thumb | IEA, Energy and AI (2025) |
| 2 million L per day | rule of thumbWater of an average 100 MW data center in the US, as much as 6,500 households, mostly indirect | Water of an average 100 MW data center in the US, as much as 6,500 households, mostly indirect | rule of thumb | IEA, Energy and AI (2025) |
| 21.3 TWh | estimateAll data centers and smaller IT installations in Germany in 2025 | All data centers and smaller IT installations in Germany in 2025 | estimate | Bitkom/Borderstep (2025) |
| 7.663 TWh | meteredAll data centers in Ireland in 2025, 23% of metered grid consumption | All data centers in Ireland in 2025, 23% of metered grid consumption | metered | CSO Ireland (2026) |
| 0.34 Wh | company figureElectricity per ChatGPT query, average | Electricity per ChatGPT query, average | company figure | Sam Altman, The Gentle Singularity |
| 0.000085 US gallons | company figureWater per ChatGPT query, about 0.32 ml, no stated accounting boundary | Water per ChatGPT query, about 0.32 ml, no stated accounting boundary | company figure | Sam Altman, The Gentle Singularity |
| 0.24 Wh | company figureElectricity per Gemini Apps text prompt, May 2025 median | Electricity per Gemini Apps text prompt, May 2025 median | company figure | Google Cloud (2025) |
| 0.26 ml | company figureDirect cooling water per Gemini Apps text prompt, May 2025 median | Direct cooling water per Gemini Apps text prompt, May 2025 median | company figure | Google Cloud (2025) |
| about 0.3 Wh | modeledTypical GPT-4o query with 500 output tokens, about one tenth of the older 3 Wh estimate | Typical GPT-4o query with 500 output tokens, about one tenth of the older 3 Wh estimate | modeled | Epoch AI (2025) |
| 1.14 Wh | modeledAgent task without reasoning, GPU electricity only | Agent task without reasoning, GPU electricity only | modeled | IEA, Key Questions on Energy and AI (2026), Figure 2.1 |
| 50 Wh | modeledAgent task with reasoning, GPU electricity only | Agent task with reasoning, GPU electricity only | modeled | IEA, Key Questions on Energy and AI (2026), Figure 2.1 |
| -23% | company figureWater use per unit of energy in Microsoft’s Phoenix data centers, fiscal 2025 | Water use per unit of energy in Microsoft’s Phoenix data centers, fiscal 2025 | company figure | Microsoft, Sustainability Report (2026) |
| 3.125 million L | standardOne Olympic pool (50 × 25 × 2.5 m) | One Olympic pool (50 × 25 × 2.5 m) | standard | World Aquatics (2026) |
Not in the register because they are calculated: derived means computed from published values (AI electricity 2024, the water split, AI water in the counter). Interpolated means computed between two anchor years with even growth (2026 to 2029 and therefore the running counter).
45 more published values on regions, Germany, training, companies and nuclear power
Every value with its type and source, grouped by topic. Company figures are not independently verified.
| Value | What | Type | Source | |
|---|---|---|---|---|
| World | ||||
| about 1.5% | estimateShare of all data centers in world electricity use in 2024 (about 415 TWh in the text, 416 TWh in the tables) | Share of all data centers in world electricity use in 2024 (about 415 TWh in the text, 416 TWh in the tables) | estimate | IEA, Energy and AI (2025) |
| about 3% | projectionShare of all data centers in world electricity use in 2030 | Share of all data centers in world electricity use in 2030 | projection | IEA, Key Questions on Energy and AI (2026) |
| less than 10% | projectionContribution of data centers to the growth of world electricity demand to 2030 | Contribution of data centers to the growth of world electricity demand to 2030 | projection | IEA, Energy and AI (2025) |
| about 1,200 TWh | projectionElectricity of all data centers in 2035, scenarios from 700 to 1,700 TWh | Electricity of all data centers in 2035, scenarios from 700 to 1,700 TWh | projection | IEA, Energy and AI (2025) |
| Regions | ||||
| 45 / 25 / 15% | estimateShares of the US, China and Europe in the electricity of all data centers in 2024 | Shares of the US, China and Europe in the electricity of all data centers in 2024 | estimate | IEA, Energy and AI (2025) |
| about +240, about +175, more than +45, about +15 TWh | projectionIncrease from 2024 to 2030 in the US, China, Europe and Japan, base case | Increase from 2024 to 2030 in the US, China, Europe and Japan, base case | projection | IEA, Energy and AI (2025) |
| United States | ||||
| 58 TWh | estimateData centers in the US in 2014 | Data centers in the US in 2014 | estimate | LBNL (2024) |
| 176 TWh | estimateData centers in the US in 2023, 4.4% of US electricity use | Data centers in the US in 2023, 4.4% of US electricity use | estimate | LBNL (2024) |
| 649 TWh | projectionData centers in the US in 2030, reference case, 11.8% of US electricity use; scenarios 521 to 843 TWh | Data centers in the US in 2030, reference case, 11.8% of US electricity use; scenarios 521 to 843 TWh | projection | LBNL, Update (2026) |
| 177 to 192 TWh | estimateData centers in the US in 2024 | Data centers in the US in 2024 | estimate | EPRI (2026) |
| 380 to 790 TWh | projectionData centers in the US in 2030, 9 to 17% of US electricity use | Data centers in the US in 2030, 9 to 17% of US electricity use | projection | EPRI (2026) |
| Germany | ||||
| 12 / 20 TWh | estimateData centers and smaller IT installations in Germany in 2015 and 2024 | Data centers and smaller IT installations in Germany in 2015 and 2024 | estimate | Bitkom/Borderstep (2025) |
| 2,980 MW | estimateIT connected load of German data centers in 2025, +9% on the year before | IT connected load of German data centers in 2025, +9% on the year before | estimate | Bitkom/Borderstep (2025) |
| more than 5,000 MW | projectionIT connected load of German data centers in 2030 | IT connected load of German data centers in 2030 | projection | Bitkom/Borderstep (2025) |
| about 15% | estimateShare of AI data centers in installed capacity in Germany in 2025 | Share of AI data centers in installed capacity in Germany in 2025 | estimate | Bitkom/Borderstep (2025) |
| 530 MW | estimateConnected load for AI in 2025, stated by Bitkom separately from the share | Connected load for AI in 2025, stated by Bitkom separately from the share | estimate | Bitkom/Borderstep (2025) |
| 2,020 MW | projectionConnected load for AI in 2030, then about 40% of capacity | Connected load for AI in 2030, then about 40% of capacity | projection | Bitkom/Borderstep (2025) |
| 37 TWh | projectionData centers in Germany in 2030 in the BEE analysis | Data centers in Germany in 2030 in the BEE analysis | projection | BEE (2025) |
| Query | ||||
| 0.03 g CO₂e | company figureEmissions per Gemini Apps text prompt, May 2025 median | Emissions per Gemini Apps text prompt, May 2025 median | company figure | Google Cloud (2025) |
| factor 33 / 44 | company figureDrop in energy and CO₂e per Gemini Apps prompt from May 2024 to May 2025 | Drop in energy and CO₂e per Gemini Apps prompt from May 2024 to May 2025 | company figure | Google Cloud (2025) |
| Training | ||||
| 5,184 / 8,930 / 72,816 t CO₂e | estimateTraining of GPT-4, Llama 3.1 405B and Grok 4 | Training of GPT-4, Llama 3.1 405B and Grok 4 | estimate | Stanford AI Index (2026) |
| about 150,000 t CO₂e | modeledTraining of Grok 4 in a second, separate estimate | Training of Grok 4 in a second, separate estimate | modeled | Epoch AI, Grok 4 (2025) |
| 20,400 t CO₂e | company figureMistral Large 2, training plus 18 months of use, plus 281,000 m³ of water | Mistral Large 2, training plus 18 months of use, plus 281,000 m³ of water | company figure | Mistral AI (2025) |
| Companies | ||||
| 18.06 / 43.59 TWh | company figureElectricity of Google, whole company, 2021 and 2025 | Electricity of Google, whole company, 2021 and 2025 | company figure | Google, Environmental Report (2026) |
| 13.62 / 37.03 TWh | company figureElectricity of Microsoft, whole company, fiscal 2021 and fiscal 2025 | Electricity of Microsoft, whole company, fiscal 2021 and fiscal 2025 | company figure | Microsoft, Data Fact Sheet (2026) |
| 9.42 / 18.42 TWh | company figureElectricity of Meta, whole company, 2021 and 2024 | Electricity of Meta, whole company, 2021 and 2024 | company figure | Meta, Environmental Data Index (2025) |
| 51.3 billion L | company figureWater withdrawal of Google, data centers, 2025 | Water withdrawal of Google, data centers, 2025 | company figure | Google, Environmental Report (2026) |
| 13.3 billion L | company figureWater withdrawal of Microsoft, whole company, fiscal 2025 | Water withdrawal of Microsoft, whole company, fiscal 2025 | company figure | Microsoft, Data Fact Sheet (2026) |
| about 9.5 billion L | company figureWater withdrawal of Amazon, data centers, 2025 | Water withdrawal of Amazon, data centers, 2025 | company figure | Amazon, Data Center Water (2025) |
| 4.1 billion L | company figureWater withdrawal of Meta, data centers, 2024 | Water withdrawal of Meta, data centers, 2024 | company figure | Meta, Environmental Data Index (2025) |
| 39.8 billion L | company figureWater consumption of Google data centers in 2025 | Water consumption of Google data centers in 2025 | company figure | Google, Environmental Report (2026) |
| 8.2 billion L | company figureWater consumption of Microsoft, whole company, fiscal 2025 | Water consumption of Microsoft, whole company, fiscal 2025 | company figure | Microsoft, Data Fact Sheet (2026) |
| 0.12 L per kWh | company figureWater intensity of Amazon data centers in 2025 | Water intensity of Amazon data centers in 2025 | company figure | Amazon, Data Center Water (2025) |
| +37% / -2% / +25% | company figureGoogle 2025: electricity demand, market-based Scope 1 and 2 emissions, supply-chain emissions | Google 2025: electricity demand, market-based Scope 1 and 2 emissions, supply-chain emissions | company figure | Google, Environmental Report (2026) |
| +25% | company figureMicrosoft, Scope 1 to 3 emissions, fiscal 2025 against the year before | Microsoft, Scope 1 to 3 emissions, fiscal 2025 against the year before | company figure | Microsoft, Data Fact Sheet (2026) |
| +16% | company figureAmazon, absolute carbon emissions in 2025 against 2024 | Amazon, absolute carbon emissions in 2025 against 2024 | company figure | Amazon, Sustainability Report (2025) |
| 5,862,614 t CO₂e | company figureMeta, location-based Scope 2 emissions of its data centers in 2024 | Meta, location-based Scope 2 emissions of its data centers in 2024 | company figure | Meta, Environmental Data Index (2025) |
| Nuclear power | ||||
| up to 2,800 MW | company figureMeta/TerraPower, new build, up to eight units | Meta/TerraPower, new build, up to eight units | company figure | Meta, Nuclear Projects (2026) |
| more than 2,100 MW | company figureMeta/Vistra, existing plant, supply deal | Meta/Vistra, existing plant, supply deal | company figure | Meta, Nuclear Projects (2026) |
| 1,920 MW | company figureAmazon/Talen, existing plant, supply deal, full contract quantity | Amazon/Talen, existing plant, supply deal, full contract quantity | company figure | Talen Energy (2025) |
| up to 1,200 MW | company figureMeta/Oklo, new build | Meta/Oklo, new build | company figure | Meta, Nuclear Projects (2026) |
| about 835 MW | company figureMicrosoft/Constellation, restart | Microsoft/Constellation, restart | company figure | Constellation (2024) |
| 615 MW | company figureGoogle/Duane Arnold, restart, plant capacity | Google/Duane Arnold, restart, plant capacity | company figure | US DOE, Duane Arnold (2026) |
| up to 500 MW | company figureGoogle/Kairos Power, new build, small reactors | Google/Kairos Power, new build, small reactors | company figure | Google, Kairos Power (2024) |
| up to 6.6 GW | company figureNew and existing clean energy that Meta’s agreements are meant to support by 2035 | New and existing clean energy that Meta’s agreements are meant to support by 2035 | company figure | Meta, Nuclear Projects (2026) |
Only national values are metered
Ireland meters grid consumption, Germany estimates. Both count all data centers, so neither gives an AI-only total.
More efficient does not mean less
Microsoft reports 23% lower water use per unit of energy for its Phoenix data centers in fiscal 2025.
That describes one site’s operation per unit of energy. It establishes neither a global water volume nor a lower total for all AI data centers.
Source: Microsoft, 2026 environmental report
How the live counter calculates
- 01
IEA annual anchors
Electricity comes from “Key Questions on Energy and AI” (April 16, 2026): around 416 TWh in 2024, 485 TWh in 2025, and around 950 TWh in 2030 for all data centers. Our World in Data reads IEA Figure 1.5 as 155 TWh for AI-focused sites in 2025. 155 plus 330 TWh of non-AI load equals the published 485 TWh. Those AI sites reach about 465 TWh by 2030. Table A.4 lists 416 TWh for 2024 and 945 TWh for the 2030 base case. For 2030, the counter follows the executive summary rounded to around 950 TWh. Water comes from “Energy and AI” (April 10, 2025): 560 billion liters consumed in 2023 and 1,200 billion liters in 2030.
- 02
Growth between anchors
Between two IEA years the annual total grows at an even annual rate. That is why 2026 AI electricity sits near 193 TWh rather than a linear midpoint. After 2030 the counter holds the last IEA value. It does not invent its own long-range forecast.
- 03
Even spread across the calendar year
The IEA publishes annual totals, not hourly loads. Inside a UTC calendar year the counter spreads that total evenly across every second, including leap years. There is no seasonality, no outage, and no peak-shaving.
- 04
Water follows the electricity share
The IEA does not split water into AI and the rest. The counter applies the current AI share of data-center electricity to the global water total. The 2023 mix stays fixed. The IEA’s rounded shares imply approximately 25% direct cooling, 67% primary energy supply and electricity generation, and 8% hardware. Hardware stays out of the live ticker. The approximately 5,200 billion liters of 2023 withdrawals are independently read from IEA Figure 5.27, with graph-reading uncertainty. The ticker uses consumption only.
Formula
annualTotal(year) × (seconds since January 1 UTC / seconds in the year)
What this is not
Not a live meter, not a satellite wattmeter, not a real-time operator feed. The IEA figures are modeled too. Other models and accounting boundaries can produce different estimates. Cooling technology, climate, and the electricity mix affect water use. Keeping the water mix fixed through 2030 can overstate or understate later consumption.
Sources
- IEA, Key Questions on Energy and AIElectricity of all data centers 2024, 2025 and 2030, AI electricity 2030, four scenarios, agent tasks, share of world electricity 2030
- IEA, Energy and AIWater consumption 2023 and 2030, split, withdrawals, rules of thumb for 100 MW, regions in 2024 and growth to 2030, outlook to 2035
- Our World in Data, How much energy do data centers and artificial intelligence use?Reading of the IEA figure: 155 TWh of AI electricity in 2025
- Google Cloud, How much energy does Google’s AI use? We did the mathGemini Apps: 0.24 Wh, 0.26 ml and 0.03 g CO₂e per text prompt (median)
- Sam Altman, The Gentle SingularityChatGPT: 0.34 Wh and 0.000085 US gallons per query
- Epoch AI, How much energy does ChatGPT use?GPT-4o: about 0.3 Wh per typical query
- Bitkom/Borderstep, Rechenzentren in Deutschland: KI treibt WachstumGermany 2015 to 2025, IT connected load and AI share
- CSO Ireland, Data Centres Metered Electricity Consumption 2025Ireland 2025: 7.663 TWh metered, 23% of grid consumption
- World Aquatics, Competition Regulations (February 2026)Dimensions of an Olympic pool (3.125 million liters)
- Microsoft, 2026 Environmental Sustainability Report (FY25)Phoenix: water use per unit of energy down 23%
- LBNL, 2024 United States Data Center Energy Usage ReportUS 2014 and 2023
- LBNL, United States Data Center Energy 2025 (update published 2026)US 2030: reference case and scenarios
- EPRI, Powering Intelligence 2026US 2024 and 2030 as ranges
- BEE, Update StrombedarfsanalyseData centers in Germany in 2030
- Google, 2026 Environmental ReportElectricity, water and emissions of Google
- Microsoft, 2026 Environmental Data Fact SheetElectricity, water and emissions of Microsoft
- Meta, 2025 Environmental Data IndexElectricity, water and emissions of Meta (2024)
- Amazon, How Amazon data centers use waterWater withdrawal and water intensity of Amazon data centers
- Amazon, 2025 Sustainability ReportAmazon emissions in 2025
- Stanford HAI, AI Index Report 2026CO₂e estimates for training GPT-4, Llama 3.1 405B and Grok 4
- Mistral AI, Our contribution to a global environmental standard for AILifecycle of Mistral Large 2
- Epoch AI, Grok 4 training resourcesCO₂e estimate for training Grok 4
- Meta, Nuclear energy projects to power American AI leadershipNuclear projects with TerraPower, Vistra and Oklo, up to 6.6 GW by 2035
- Talen Energy, expanded agreement with Amazon (SEC filing)1,920 MW for Amazon at full contract quantity
- Constellation, Crane Clean Energy CenterRestart of Three Mile Island Unit 1 for Microsoft, about 835 MW
- US Department of Energy, loan for the Duane Arnold restartRestart of Duane Arnold, 615 MW
- Google, Agreement with Kairos PowerUp to 500 MW from small reactors by Kairos Power
More questions
Changelog
New address and two new questions
- New address /en/ai-energy-consumption/, the page now also brings together the figures on AI energy use
- Two new questions: data center electricity in the US and China, and electricity, water and nuclear deals of Google, Microsoft and Meta
- Germany question with the Bitkom and Borderstep series 2015 to 2025
- More values on the world, regions, the US, Germany, queries, training, companies and nuclear power, collapsed below the value register
Rebuilt: eight questions with short answers
- Eight questions with a short answer, one chart each and the origin of every number
- Counter in TWh like every chart, calculator placed directly at the question about a single query
- Every published value in one filterable register instead of many extra conversion charts