# AI energy statistics (updated November 2026)

> AI energy statistics, checked on primary sources: data centre electricity, energy per prompt, images, video, training, water, carbon and hardware trends.

Source: https://greenai.guide/articles/ai-energy-statistics/  
Updated: 2026-11-20

## In short

These AI energy statistics are 40 of the most useful figures on AI's electricity, water and carbon, each checked on its primary source and dated. In short, data centres used about 415 TWh of electricity in 2024 (around 1.5% of the world's total), the IEA expects around 950 TWh by 2030, and a typical text prompt uses roughly 0.24–0.34 Wh by providers' own figures, while images, video, long inputs and reasoning use far more.

## Key points

- Data centres used about 415 TWh of electricity in 2024, around 1.5% of the global total, and 485 TWh in 2025, according to the IEA.
- The IEA projects around 950 TWh by 2030, about 3% of global electricity demand; electricity use by AI-focused data centres grew 50% in 2025.
- Providers report about 0.24 Wh (Google, median Gemini text prompt) and 0.34 Wh (OpenAI, average ChatGPT query) per text prompt.
- Reasoning models used on average 30 times more energy than standard ones in Hugging Face's 2025 tests.
- Training GPT-3 used an estimated 1,287 MWh and emitted about 552 tonnes of CO₂e.
- Data centres used 23% of Ireland's metered electricity in 2025.

This page collects the AI energy statistics people most often look for, each checked on its primary source, with
the year it applies to. Use it to find a figure quickly, and follow the source before quoting it in anything
important.

## What are the key AI energy statistics?

**Data centres use about 1.5% of the world's electricity, that is set to roughly double by 2030, and AI is the main
driver.** A typical text prompt uses a fraction of a watt-hour, while images, video, very long inputs and reasoning
use far more. The tables below give 40 figures in ten groups. Where sources disagree or a figure is a range, the
range is given.

## Data centre electricity

**Data centres used about 415 TWh in 2024 and 485 TWh in 2025, according to the IEA.** A terawatt-hour (TWh) is a
billion kilowatt-hours.

| # | Statistic | Source | Year |
|---|---|---|---|
| 1 | Data centres used about 415 TWh of electricity in 2024, around 1.5% of global consumption. | International Energy Agency (IEA) | 2025 |
| 2 | Data-centre electricity use grew 17% in 2025, to 485 TWh. | IEA | 2026 |
| 3 | The IEA projects around 950 TWh by 2030, about 3% of global electricity demand. | IEA | 2026 |
| 4 | Projections for 2035 range from 700 to 1,700 TWh across the IEA's scenarios. | IEA | 2025 |
| 5 | Data-centre electricity use has grown by around 12% a year since 2017. | IEA | 2025 |
| 6 | Data centres will account for around one-tenth of global electricity demand growth to 2030. | IEA | 2025 |

## Where demand is concentrated

**A few countries and regions carry most of the load.** See [AI and data centre electricity demand](/articles/ai-data-centre-electricity/)
for what this means for grids.

| # | Statistic | Source | Year |
|---|---|---|---|
| 7 | The United States and China account for nearly 80% of global growth in data-centre electricity to 2030. | IEA | 2025 |
| 8 | US data centres used 176 TWh in 2023, about 4.4% of US electricity. | Lawrence Berkeley National Laboratory (LBNL) | 2024 |
| 9 | US data centres are projected to use 325–580 TWh by 2028, 6.7–12% of US electricity. | LBNL | 2024 |
| 10 | Data centres account for nearly half of US electricity demand growth to 2030. | IEA | 2025 |
| 11 | Data centres used 23% of Ireland's metered electricity in 2025, up from 5% in 2015. | Central Statistics Office, Ireland | 2026 |
| 12 | Data centres used nearly a quarter of Virginia's electricity in 2023. | Electric Power Research Institute | 2024 |

## AI's share

**There is no exact public figure for AI's share of data-centre electricity, but AI is the fastest-growing part.**

| # | Statistic | Source | Year |
|---|---|---|---|
| 13 | Electricity use by AI-focused data centres grew 50% in 2025. | IEA | 2026 |
| 14 | Electricity use by accelerated servers, mainly used for AI, is projected to grow 30% a year in the IEA's base case. | IEA | 2025 |
| 15 | A typical AI-focused data centre uses as much electricity as 100,000 households. | IEA | 2025 |

## Energy per prompt

**Providers' own figures put a typical text prompt at around 0.24–0.34 Wh.** These are self-reported, use
different methods and cover text only. See [measurement is the bottleneck](/articles/measurement-is-the-bottleneck/).

| # | Statistic | Source | Year |
|---|---|---|---|
| 16 | A median Gemini Apps text prompt used 0.24 Wh, emitted 0.03 g CO₂e and consumed 0.26 mL of water. | Google (Elsworth et al.) | 2025 |
| 17 | Counting only active AI chips in the most efficient data centres gives 0.10 Wh for the same prompt. | Google (Elsworth et al.) | 2025 |
| 18 | An average ChatGPT query uses about 0.34 Wh and about 0.000085 gallons of water. | Sam Altman, OpenAI | 2025 |
| 19 | A typical GPT-4o query uses roughly 0.3 Wh, rising to about 2.5 Wh with a 10,000-token input and almost 40 Wh with a 100,000-token input. | Epoch AI | 2025 |
| 20 | A 400-token response from Mistral Large 2 emits 1.14 g CO₂e and uses 45 mL of water. | Mistral AI | 2025 |
| 21 | Reasoning models used on average 30 times more energy than standard models; one model used 150–700 times more with reasoning on. | Hugging Face AI Energy Score | 2025 |

## Images and video

**Generating images and video uses far more energy than text tasks.**

| # | Statistic | Source | Year |
|---|---|---|---|
| 22 | Image generation used a mean of about 2.9 kWh per 1,000 images, against 0.002 kWh per 1,000 text classifications. | Luccioni, Jernite and Strubell | 2024 |
| 23 | The least efficient image model tested used about half a smartphone charge per image. | Luccioni, Jernite and Strubell | 2024 |
| 24 | Open text-to-video models used from 0.14 Wh to over 415 Wh per short clip, with energy growing quadratically with resolution and length. | Delavande, Pierrard and Luccioni | 2025 |

## Training runs

**Training a large model can use gigawatt-hours, but most developers of the newest models don't disclose it.**

| # | Statistic | Source | Year |
|---|---|---|---|
| 25 | Training GPT-3 used an estimated 1,287 MWh and emitted about 552 tonnes of CO₂e. | Patterson et al., Google | 2021 |
| 26 | Training BLOOM emitted about 24.7 tonnes of CO₂e from electricity, or 50.5 tonnes including manufacturing and idle hardware. | Luccioni, Viguier and Ligozat | 2022 |
| 27 | Training the Llama 3.1 models took 39.3 million GPU hours and 11,390 tonnes of CO₂e (location-based). | Meta | 2024 |
| 28 | Training Mistral Large 2 and 18 months of its use produced 20.4 kilotonnes of CO₂e and consumed 281,000 m³ of water. | Mistral AI | 2025 |

## Water

**AI uses water mainly to cool data centres and, indirectly, to generate electricity.** See
[how to estimate the water footprint of AI](/how-do-i/estimate-the-water-footprint-of-ai/).

| # | Statistic | Source | Year |
|---|---|---|---|
| 29 | Training GPT-3 in Microsoft's US data centres could directly evaporate 700,000 litres of fresh water. | Li, Yang, Islam and Ren | 2023 (revised 2025) |
| 30 | Global AI demand could account for 4.2–6.6 billion m³ of water withdrawal in 2027. | Li, Yang, Islam and Ren | 2023 (revised 2025) |
| 31 | Google consumed 41 billion litres (10.9 billion gallons) of water across its data centres and offices in 2025, and replenished about 78% of its freshwater consumption. | Google | 2026 |

## Carbon and company emissions

**The largest AI providers' emissions are rising, despite large purchases of clean energy.**

| # | Statistic | Source | Year |
|---|---|---|---|
| 32 | Google's ambition-based emissions were about 14.5 million tonnes of CO₂e in 2025, up 18% on 2024 and 81% on 2019. | Google | 2026 |
| 33 | Microsoft's total emissions were about 20.3 million tonnes of CO₂e in its 2025 financial year, against about 12.9 million in 2020. | Microsoft | 2026 |
| 34 | Amazon's absolute emissions rose 16% from 2024 to 2025, to 80.85 million tonnes of CO₂e. | Amazon | 2026 |
| 35 | Generative AI could add 1.2–5.0 million tons of electronic waste in total over 2020–2030. | Wang et al., Nature Computational Science | 2024 |

## Hardware and efficiency trends

**Hardware and software are getting much more efficient, but demand is growing faster.**

| # | Statistic | Source | Year |
|---|---|---|---|
| 36 | Leading machine learning hardware becomes about 40% more energy-efficient each year. | Epoch AI | 2024 |
| 37 | The most energy-efficient supercomputer on the November 2026 Green500 list, KAIROS in France, delivers 73.28 gigaflops per watt. | TOP500 | 2026 |
| 38 | Computations per kilowatt-hour doubled about every 1.57 years from 1946 to 2009 ("Koomey's law"). | Koomey et al. | 2011 |

## Efficiency gains reported by providers and researchers

**Reported gains are large, but most are self-reported or measured against a chosen baseline.**

| # | Statistic | Source | Year |
|---|---|---|---|
| 39 | Google reports a 33-fold fall in energy and a 44-fold fall in carbon per median Gemini text prompt over 12 months. | Google (Elsworth et al.) | 2025 |
| 40 | Small task-specific models can cut energy use by up to 90%, and shorter prompts and responses by over 50%. | UNESCO and University College London | 2025 |

Two further figures are often quoted. NVIDIA said in 2024 that its Blackwell platform reduces cost and energy
consumption by up to 25 times against its previous generation for large language model inference; this is a vendor
claim for a specific workload. And Google researchers estimated in 2021 that the choice of model, data centre and
processor can reduce the carbon footprint of training by up to around 100 to 1,000 times. For how to get such gains
in practice, see [energy-efficient AI](/articles/energy-efficient-ai/).

## Public attitudes

**Concern about AI's environmental impact is rising, at least in the US.** Survey evidence elsewhere is thinner.

- In 2026, 53% of Americans were extremely or very concerned about AI's environmental impact, up from 41% in 2025
  (AP-NORC and the Energy Policy Institute at the University of Chicago).
- In 2025, 72% of Americans were at least somewhat concerned (same pollsters).
- In the UK, the government's 2024 public attitudes tracker found 54% expected AI to help monitor climate change,
  and noted that this optimism was not matched by awareness of AI's negative environmental impacts.

## How we choose figures

**We use primary sources, give the year, and prefer ranges where figures are contested.**

- **Primary sources only.** Each figure is checked on the publisher's own report, paper or page, not on news
  coverage. Where we could not find a primary source, the figure is left out.
- **Stated as published.** We don't round figures into something the source doesn't say. Where we calculate
  anything ourselves, we say so.
- **Boundaries matter.** Per-prompt figures change with what is counted: Google's own figure falls from 0.24 Wh to
  0.10 Wh if only active chips are counted. Company figures are self-reported and not independently audited.
- **Ranges where contested.** Projections differ by scenario, and we show the range rather than pick one number.
- **Older figures flagged.** A widely quoted 2019 estimate for one training run with neural architecture search
  (about 626,000 lb of CO₂e) was later argued by Google researchers to overstate the real cost; we have not included
  it in the tables.

To check a figure you have seen elsewhere, see [how to check a claim about AI energy use](/how-do-i/check-a-claim-about-ai-energy-use/).

## Last checked

**All figures on this page were last checked on their sources on 20 November 2026.** We update the page when the
IEA, national statistics offices or major providers publish new data. If you spot a newer figure or an error,
please [contact us](/contact/).

## What to do next

- Link to this page or a specific table when you cite AI energy figures, and name the original source.
- Put personal use in context with [how to put my AI energy use in perspective](/how-do-i/put-my-ai-energy-use-in-perspective/).
- Measure your own AI's energy with [how to measure the energy and carbon of AI](/practice/measure-ai-energy/).
- Ask suppliers for their figures: see [how to ask an AI company for energy figures](/how-do-i/ask-an-ai-company-for-energy-figures/).
- Cut your own footprint with [energy-efficient AI](/articles/energy-efficient-ai/).

## Frequently asked questions

### How much energy does AI use?

There is no single figure for AI. The IEA estimates all data centres used about 415 TWh in 2024, around 1.5% of global electricity, with AI-focused data centres the fastest-growing part: their electricity use rose 50% in 2025.

### How much energy does a ChatGPT query use?

OpenAI's chief executive gave about 0.34 Wh for an average query in 2025, and Epoch AI independently estimated about 0.3 Wh for a typical GPT-4o query. Very long inputs raise this sharply: Epoch AI estimates almost 40 Wh with a 100,000-token input.

### How much energy does it take to generate an AI image?

In a 2024 study by Luccioni, Jernite and Strubell, image generation averaged about 2.9 kWh per 1,000 images, about 2.9 Wh each. The least efficient model used roughly half a smartphone charge per image.

### How much water does AI use?

Google reports 0.26 mL of water for a median Gemini text prompt, and Mistral 45 mL for a 400-token response from Mistral Large 2. Researchers estimate global AI demand could account for 4.2–6.6 billion cubic metres of water withdrawal in 2027.

### How much energy did it take to train GPT-3?

Google researchers estimated 1,287 MWh of electricity and about 552 tonnes of CO₂e for GPT-3's training, published in 2021. Newer, larger models' training energy is mostly not disclosed.

### What percentage of electricity do data centres use?

About 1.5% globally in 2024 (IEA), 4.4% in the United States in 2023 (Lawrence Berkeley National Laboratory) and 23% of metered electricity in Ireland in 2025 (Central Statistics Office).

### Is AI getting more energy efficient?

Yes, per task. Google reports the energy of its median Gemini prompt fell 33-fold in a year, and Epoch AI finds leading AI chips becoming about 40% more energy-efficient each year. Total use is still rising because demand grows faster.

### Are people worried about AI's environmental impact?

Increasingly, in the US at least: an AP-NORC and University of Chicago poll found 53% of Americans extremely or very concerned about AI's environmental impact in 2026, up from 41% in 2025.

## Sources

- [Energy and AI: executive summary](https://www.iea.org/reports/energy-and-ai/executive-summary), International Energy Agency (2025)
- [Key questions on energy and AI](https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary), International Energy Agency (2026)
- [DOE releases new report evaluating increase in electricity demand from data centers (2024 United States Data Center Energy Usage Report, LBNL)](https://www.energy.gov/articles/doe-releases-new-report-evaluating-increase-electricity-demand-data-centers), US Department of Energy and Lawrence Berkeley National Laboratory (2024)
- [Data centres metered electricity consumption 2025: key findings](https://www.cso.ie/en/releasesandpublications/ep/p-dcmec/datacentresmeteredelectricityconsumption2025/keyfindings/), Central Statistics Office, Ireland (2026)
- [Powering data centers: U.S. energy system and emissions impacts of growing loads](https://restservice.epri.com/publicdownload/000000003002031198/0/Product), Electric Power Research Institute (2024)
- [Measuring the environmental impact of delivering AI at Google scale](https://arxiv.org/abs/2508.15734), Elsworth et al., Google (2025)
- [The gentle singularity](https://blog.samaltman.com/the-gentle-singularity), Sam Altman (OpenAI) (2025)
- [How much energy does ChatGPT use?](https://epoch.ai/gradient-updates/how-much-energy-does-chatgpt-use), Epoch AI (2025)
- [Our contribution to a global environmental standard for AI](https://mistral.ai/news/our-contribution-to-a-global-environmental-standard-for-ai), Mistral AI (2025)
- [AI Energy Score v2: refreshed leaderboard, now with reasoning](https://huggingface.co/blog/sasha/ai-energy-score-v2), Hugging Face (Luccioni and Gamazaychikov) (2025)
- [Power hungry processing: watts driving the cost of AI deployment?](https://arxiv.org/abs/2311.16863), Luccioni, Jernite and Strubell (FAccT 2024) (2024)
- [Video killed the energy budget: characterizing the latency and power regimes of open text-to-video models](https://arxiv.org/abs/2509.19222), Delavande, Pierrard and Luccioni (2025)
- [Carbon emissions and large neural network training](https://arxiv.org/abs/2104.10350), Patterson et al., Google (2021)
- [Energy and policy considerations for deep learning in NLP](https://arxiv.org/abs/1906.02243), Strubell, Ganesh and McCallum (2019)
- [Estimating the carbon footprint of BLOOM, a 176B parameter language model](https://arxiv.org/abs/2211.02001), Luccioni, Viguier and Ligozat (2022)
- [Llama 3.1 model card](https://huggingface.co/meta-llama/Llama-3.1-8B), Meta (2024)
- [Making AI less 'thirsty': uncovering and addressing the secret water footprint of AI models](https://arxiv.org/abs/2304.03271), Li, Yang, Islam and Ren (2025)
- [Google 2026 environmental report](https://sustainability.google/google-2026-environmental-report/), Google (2026)
- [2026 environmental data fact sheet](https://cdn-dynmedia-1.microsoft.com/is/content/microsoftcorp/microsoft/msc/documents/presentations/CSR/2026-Microsoft-Environmental-Data-Fact-Sheet-PDF.pdf), Microsoft (2026)
- [2025 Amazon sustainability report](https://sustainability.aboutamazon.com/2025-amazon-sustainability-report.pdf), Amazon (2026)
- [E-waste challenges of generative artificial intelligence](https://doi.org/10.1038/s43588-024-00712-6), Wang, Zhang, Tzachor and Chen, Nature Computational Science (2024)
- [Leading ML hardware becomes 40% more energy-efficient each year](https://epoch.ai/data-insights/ml-hardware-energy-efficiency), Epoch AI (2024)
- [Implications of historical trends in the electrical efficiency of computing](https://doi.org/10.1109/MAHC.2010.28), Koomey, Berard, Sanchez and Wong, IEEE Annals of the History of Computing (2011)
- [Green500, November 2026](https://top500.org/lists/green500/2026/11/), TOP500 (2026)
- [NVIDIA Blackwell platform arrives to power a new era of computing](https://nvidianews.nvidia.com/news/nvidia-blackwell-platform-arrives-to-power-a-new-era-of-computing), NVIDIA (2024)
- [AI large language models: new report shows small changes can reduce energy use by 90%](https://www.unesco.org/en/articles/ai-large-language-models-new-report-shows-small-changes-can-reduce-energy-use-90), UNESCO and University College London (2025)
- [2026 poll: Americans' views on the impact of AI and data centers](https://epic.uchicago.edu/insights/2026-poll-americans-views-on-the-impact-of-ai-and-data-centers/), Energy Policy Institute at the University of Chicago and AP-NORC (2026)
- [2025 poll: Americans' views on the environmental impact of AI](https://epic.uchicago.edu/insights/2025-poll-americans-views-on-the-environmental-impact-of-ai), Energy Policy Institute at the University of Chicago and AP-NORC (2025)
- [Public attitudes to data and AI: tracker survey (wave 4)](https://www.gov.uk/government/publications/public-attitudes-to-data-and-ai-tracker-survey-wave-4/public-attitudes-to-data-and-ai-tracker-survey-wave-4-report), Department for Science, Innovation and Technology (2024)
