How Much Energy Does AI Use?

AI energy use is easy to get wrong. Some numbers make one chatbot reply sound like running a dishwasher. Others make it sound like nothing at all.

The honest answer sits in the middle, and it depends on what you measure. A single query is tiny. The data centers behind millions of queries are not.

This guide gives you the current 2025-2026 figures from primary sources. It covers electricity, water, and how both quietly shape what you pay for AI.

How much energy does AI use? The headline numbers

A typical AI chatbot query uses roughly 0.3 watt-hours of electricity, based on 2025 analysis by Epoch AI. That is about what an LED bulb uses in a few minutes.

Training a large model is a different scale. Training GPT-4 is estimated to have used 50,000 to 62,000 megawatt-hours, enough to power a small city for days.

So AI energy use spans two very different worlds. One query is small. Building the model, and running it for billions of people, adds up fast.

The key point most headlines miss: per-query figures usually leave out the huge one-time training cost. Both numbers matter, and they are rarely shown together.

• Per query: about 0.3 watt-hours for a standard text answer (Epoch AI, 2025).
• Google reported about 0.24 watt-hours per median Gemini text query.
• A long query with 100,000 input tokens can use around 40 watt-hours, over 100 times more.
• Training one frontier model draws 20 to 25 megawatts of power for about three months.

Why AI uses so much power

AI uses a lot of power because it runs on specialized chips that never rest. These chips, mostly GPUs from Nvidia, do trillions of math steps per answer.

A modern AI GPU can draw 700 to 1,000 watts on its own. A single AI server rack can pull as much power as dozens of homes.

The chips are only half the story. They run hot, so data centers spend more energy on cooling and power delivery to keep them alive.

This is why AI feels different from older computing. It concentrates enormous, steady power demand in one building, running around the clock.

• One high-end AI GPU can draw 700 to 1,000 watts under load.
• Racks of these GPUs run 24/7, not in short bursts.
• Cooling and power conversion add overhead on top of the chips themselves.
• Demand is steady, which strains the local power grid more than spiky loads.

How much energy does an AI data center use?

AI data centers use enough electricity to rival whole countries when you add them up. The IEA projects global data center demand will more than double to about 945 terawatt-hours by 2030.

That would be near 3% of all electricity used on Earth. It is slightly more than Japan uses in a year today.

The United States carries the biggest share. Data centers there are set to drive nearly half of the country's electricity demand growth through 2030.

In the US, Lawrence Berkeley National Lab found data centers used about 4.4% of national electricity in 2023. That could reach 6.7% to 12% by 2028.

• Global data center electricity: about 415 TWh in 2024, projected near 945 TWh by 2030 (IEA).
• AI-focused data center use is expected to roughly triple over that period.
• US data center use climbed from 58 TWh in 2014 to 176 TWh in 2023 (Berkeley Lab).
• US projection for 2028: 325 to 580 TWh, a wide range that reflects real uncertainty.

The water side: how much water does AI use?

AI also uses water, mainly to cool the servers and to generate the electricity they run on. This part gets far less attention than power.

Researchers at UC Riverside estimate that a short set of 10 to 50 AI responses can use about half a liter of fresh water. That counts both cooling and power-plant water.

Training is thirstier still. The same team estimated GPT-3 training in Microsoft US data centers used about 700,000 liters of fresh water.

A non-obvious detail: indirect water, used to make the electricity, can be far larger than the water used directly for cooling. Location and grid mix matter a lot.

• About 500 mL of fresh water per 10 to 50 medium AI responses (Li and Ren, UC Riverside).
• GPT-3 training: roughly 700,000 liters, similar to making about 320 electric cars.
• Global AI water demand projected at 4.2 to 6.6 billion cubic meters per year by 2027.
• Indirect water for electricity can outweigh direct cooling water many times over.

Energy per query vs training: the honest ranges

A short AI answer and a giant training run are worlds apart, so treat any single number with care. The right figure depends on the task.

Epoch AI puts a standard text query near 0.3 watt-hours. Google reported about 0.24 watt-hours for a median Gemini text answer, and one estimate from OpenAI leadership landed near 0.34 watt-hours.

But query cost is not flat. A very long prompt with 100,000 input tokens can use around 40 watt-hours, more than 100 times a short one.

Training is a large upfront cost, but spread across billions of queries it becomes tiny per answer. Day to day, running the model (inference) is what dominates AI's ongoing energy use.

• Short text query: about 0.24 to 0.34 watt-hours across recent estimates.
• Medium query (10,000 input tokens): about 2.5 watt-hours.
• Long query (100,000 input tokens): about 40 watt-hours.
• Image and video generation use much more than text, though exact figures vary widely.

How AI energy shows up in what you pay

Energy is a core cost of every AI service, and it flows into the price you pay per token. More compute means more power, which means a higher bill.

When you buy AI through an API, you pay per token. Long inputs cost more partly because they burn more electricity, as the 40-watt-hour long-query figure shows.

Providers also pay to build and power the data centers. Those capital and energy costs get baked into subscription and usage pricing over time.

This is why prompt length and model choice affect both your footprint and your invoice. Leaner prompts and right-sized models cut energy and cost at the same time.

• AI pricing is per token, and token count tracks roughly with compute and energy.
• Long context windows raise energy use and price together.
• Bigger models cost more per token because they do more computation per answer.
• Efficiency work that lowers your energy footprint usually lowers your bill too.

What businesses can do about AI's footprint

You cannot rewire a data center, but you can control how much AI compute you buy and waste. Smart usage is the biggest lever most teams have.

Start by matching the model to the job. Many tasks run fine on a smaller, cheaper, lower-energy model instead of the largest one.

Then trim your prompts and context. Shorter inputs cut both energy and cost, with no loss of quality when done well.

Finally, choose providers that publish real sustainability data and run on cleaner grids. For deeper tactics, see our Sustainable AI guide and why data centers are turning to nuclear.

• Right-size the model: use a smaller model when it meets the need.
• Trim prompts and context length to cut tokens, energy, and cost.
• Batch and cache repeat work instead of re-running the same queries.
• Prefer vendors with transparent energy and water reporting and clean-grid siting.

AI energy use in perspective: a quick comparison

It helps to compare AI energy use against familiar activities. The table below puts common AI tasks next to everyday energy uses.

These are approximate figures drawn from the sources cited throughout this guide. Real numbers vary by model, hardware, and data center.

The takeaway is consistency, not precision. A short chat is trivial, a long document analysis is not, and training is in a class of its own.