How Much Energy Does AI Use? 2026 Facts and Math

Quick answer: A short AI chatbot text query typically uses a fraction of a watt-hour. Google measured 0.24 Wh for its median Gemini text prompt in May 2025, and OpenAI CEO Sam Altman cited an average of 0.34 Wh per ChatGPT query in June 2025. Long inputs, reasoning and AI-agent tasks can use 10 to 100 times more. Globally, AI-focused data centres used about 155 TWh in 2025, about 0.5% of world electricity, and all data centres about 485 TWh (1.5%), per the IEA, which projects roughly 945 to 950 TWh for all data centres by 2030.

Facts checked November 22, 2026.

Headlines swing between “every ChatGPT question boils a kettle” and “AI energy use is trivial.” Both can’t be right. With a few unit conversions you can check the numbers yourself, and see where AI’s energy use really is small, where it is large, and why it is growing fast.

Units First: Watts vs Watt-Hours

A watt (W) measures power, how fast energy is used. A watt-hour (Wh) measures energy: 1 watt for 1 hour. Bigger units go up by factors of 1,000:

UnitEqualsExample
1 Wh1 watt for 1 hourA few AI chatbot prompts
1 kWh1,000 WhA typical unit on an electricity bill
1 GWh1,000,000 kWhAbout a day of electricity for 60,000 EU residents
1 TWh1,000 GWhUsed for national and global totals

So 1 TWh \(= 10^{12}\) Wh. Getting these powers of 10 right is most of the math; see exponential growth and AI for working with big numbers.

Energy per AI Query

TaskEstimated energySource
Median Gemini text prompt (measured, full system)About 0.24 WhGoogle, measured May 2025
“Average” ChatGPT queryAbout 0.34 WhOpenAI CEO Sam Altman, June 2025 (methodology not published)
Typical ChatGPT queryAbout 0.3 WhEpoch AI (modeled estimate)
Query with a long input (about 7,500 words)About 2.5 WhEpoch AI
Query with a very long input (about 75,000 words)About 40 WhEpoch AI
Standard AI agent requestAbout 1.1 WhIEA, 2026
Agentic request with reasoningAbout 50 WhIEA, 2026

To feel the size: a 1,000-watt microwave uses \(1{,}000 \text{ W} \times \tfrac{1}{3600} \text{ h} \approx 0.28\) Wh per second, so a 0.24 Wh prompt is less than one second of microwave time. Note the differences in method: a measured median, a stated average and modeled estimates are not the same thing, and some figures count only the AI chips while others include cooling and the whole facility. The range in the table spans more than 200 times, so “how much energy does a query use?” depends heavily on what you ask and how hard the model thinks.

Worked Example: A Billion Prompts a Day

Suppose a service answers 1 billion short prompts a day at 0.3 Wh each:

\[ 10^9 \times 0.3 \text{ Wh} = 3 \times 10^8 \text{ Wh} = 300 \text{ MWh per day}. \]

Over a year: \(300 \text{ MWh} \times 365 \approx 110{,}000\) MWh \(\approx 0.11\) TWh. That is real, but small next to the 155 TWh used by AI-focused data centres in 2025. Most AI electricity goes to heavier work: long contexts, reasoning models, agents, image and video generation, and training new models.

How Much Electricity Does a 100 MW AI Data Centre Use?

A facility drawing a steady 100 megawatts (actual demand, not just installed capacity) uses

\[ 100 \text{ MW} \times 24 \text{ h} = 2{,}400 \text{ MWh per day}, \qquad 2{,}400 \times 365 = 876{,}000 \text{ MWh} \approx 0.88 \text{ TWh per year}. \]

At 10 cents per kWh, that is \(876{,}000{,}000 \text{ kWh} \times 0.10 \approx 87.6\) million dollars of electricity a year. The largest AI campuses being built in 2026 are planned at well over 1,000 MW.

Can AI Get More Efficient but Use More Electricity?

Yes. Suppose a query uses 1 Wh and a service gets 1 billion a day: 1 GWh per day. If efficiency improves so each query uses half as much, 0.5 Wh, but cheaper, better AI leads to 4 billion queries a day, total use becomes \(4 \times 10^9 \times 0.5 \text{ Wh} = 2\) GWh, double the original. Lower energy per query does not guarantee lower total demand.

Personal Scale: How Many Prompts Is a Day of Electricity?

Our World in Data estimates that the average person in the European Union uses about 17,000 Wh of electricity per day (people in the US use about twice as much). At 0.3 Wh per prompt:

\[ \frac{17{,}000 \text{ Wh}}{0.3 \text{ Wh}} \approx 57{,}000 \text{ short prompts}. \]

For long-input queries at 2.5 Wh, it is about 6,800. For a casual user, chatbot prompts are a tiny slice of personal electricity use; heavy agentic use is a different story.

The Global Picture

MeasureElectricityShare of world electricity
All data centres, 2025 (IEA)About 485 TWhAbout 1.5%
AI-focused data centres, 2025 (IEA)About 155 TWhAbout 0.5%
Cryptocurrency mining (estimate)About 150–200 TWhAbout 0.5–0.6%
All data centres, 2030 projection (IEA base case)About 945 TWhAbout 3%

The IEA’s 2026 summary rounds the 2030 figure to about 950 TWh; its detailed breakdown, reproduced by Our World in Data, totals 945 TWh. Check the AI share: \(155 \div 31{,}800 \approx 0.0049\), about 0.5% of the roughly 31,800 TWh of electricity the world generated in 2025. Growing from 485 to 945 TWh in five years means multiplying by about 1.95, an average growth rate of \((945/485)^{1/5} \approx 1.14\), or about 14% per year.

Why the Local Picture Matters More

Global averages hide concentration. In the United States, data centres use about 5% of electricity (AI-focused ones probably about 2%), and in Virginia, a hub for data centres, more than a quarter. In Ireland, data centres account for over 20% of electricity use. A large AI campus can need as much power as a city, which is why grid connections, power plants and electricity prices have become the real bottleneck for AI growth, the same constraint Epoch AI expects to limit training runs by 2030 (see AI scaling laws).

Electricity vs Emissions

The same amount of electricity can mean very different carbon emissions depending on whether it comes from coal, gas, nuclear, wind or solar, and when and where it is used. That is why companies report both energy and emissions, and why the location of new data centres matters so much.

Training vs Using AI

Training a frontier model is a big one-time cost: Epoch AI estimates that training xAI’s Grok 4 used about 0.31 TWh. Using the EU figure above, that is enough electricity for about

\[ \frac{0.31 \times 10^{12} \text{ Wh}}{17{,}000 \text{ Wh/day} \times 365} \approx 50{,}000 \]

people for a year. But a popular model is then used billions of times, so over its lifetime most AI energy goes to inference, answering requests, rather than training.

Common Misconceptions

“Every ChatGPT question uses as much energy as boiling a kettle.” No. Heating 1 liter of water from 20°C to boiling takes about \(4{,}186 \times 80 \approx 335{,}000\) joules, roughly 93 Wh, hundreds of times a typical short prompt. Long agentic tasks are much closer.

“AI already uses a big share of world electricity.” About 0.5% in 2025, but growing fast and very concentrated in certain regions.

“Energy per query is fixed.” It falls as chips and software improve and rises as people use longer, reasoning-heavy tasks, so totals depend on both.

Watch: AI and Electricity

If you learn better by watching, these videos are worth your time.

How much electricity does AI actually use? (International Energy Agency). The IEA explains its own estimates and projections.
Why are AI data centers using so much electricity? – Sajan Saini (TED-Ed). An animated explainer on what drives data-centre power demand.

Check Your Understanding

Five quick questions to see how much of this has stuck. Pick an answer and press Submit; you will see a short explanation either way.

Answer each question and press Submit to see your result right away.

1. A 1,000 W microwave runs for 1 second. About how much energy is that?
Found a bug? Let us know!
2. 1 billion prompts a day at 0.3 Wh each uses how much per day?
Found a bug? Let us know!
3. A data centre draws a steady 100 MW. About how much energy does it use per year?
Found a bug? Let us know!
4. AI-focused data centres used about 155 TWh of roughly 31,800 TWh in 2025. What share is that?
Found a bug? Let us know!
5. Growing from 485 to 945 TWh in 5 years is roughly what annual growth rate?
Found a bug? Let us know!

Frequently Asked Questions

How much energy does one ChatGPT query use?

Roughly 0.3 Wh for a typical short text query: OpenAI’s CEO has cited about 0.34 Wh, Epoch AI estimates about 0.3 Wh, and Google reported 0.24 Wh for a median Gemini prompt. Long inputs and reasoning agents can use 10 to 100 times more.

How much electricity does AI use worldwide?

The IEA estimates AI-focused data centres used about 155 TWh in 2025, around 0.5% of global electricity, out of about 485 TWh (1.5%) for all data centres.

Will AI electricity use keep growing?

The IEA base case projects data centres reaching about 945 TWh, roughly 3% of global electricity, by 2030, with AI-focused centres a growing share.

Does training or using AI use more energy?

Training a frontier model is a large one-time cost, around 0.31 TWh for Grok 4 by Epoch AI’s estimate, but because popular models answer billions of requests, using them (inference) usually dominates over time.

Is it bad for the environment to use ChatGPT?

A short prompt uses about as much electricity as a microwave running for one second. Impacts grow with long, reasoning-heavy and agentic use, and depend on how the electricity is generated.

Keep Exploring

Sources and Further Reading

Related to This Article

What people say about "How Much Energy Does AI Use? 2026 Facts and Math - Effortless Math"?

No one replied yet.

Leave a Reply