Title: How Much Energy Does AI Use? Shares, Ranges, Limits | FOUND

Description: How much energy AI uses: its share of data-center electricity now and by 2030, why the estimates disagree, and what training and inference each draw.

Canonical: https://lotsfound.com/insights/how-much-energy-does-ai-use/

# How much energy does AI use?

No meter separates AI from other computing, so every figure is a modeled estimate with its own boundary. Published estimates put AI well under half of data-center electricity today. Projections for 2030 approach half.

[Check a parcel](/parcel-check/)

In this article

1. [How much electricity AI uses today](#how-much-electricity-ai-uses-today)
2. [Why the estimates disagree](#why-the-estimates-disagree)
3. [How much electricity AI could use by 2030](#how-much-electricity-ai-could-use-by-2030)
4. [Training versus inference](#training-versus-inference)
5. [How much energy one AI request uses](#how-much-energy-one-ai-request-uses)
6. [What AI's share means for a proposed site](#what-ais-share-means-for-a-proposed-site)
7. [Key recap](#key-recap)
8. [Questions](#questions)
9. [References](#references)

> **Quick summary:** This guide explains how much electricity AI uses, why the published estimates disagree, and how training and inference differ. It is for residents, landowners, and local officials who hear AI blamed for new power demand. The key takeaway: AI is a growing minority of data-center electricity today, and each estimate counts it in a different way.

## How much electricity AI uses today

AI uses a minority of data-center electricity today, but no one measures it directly. Each estimate below counts a different thing in a different place and year. Read each row on its own.

| Estimate | What it counts | Year | Result |
|---|---|---|---|
| [International Energy Agency](https://iea.blob.core.windows.net/assets/de9dea13-b07d-42c5-a398-d1b3ae17d866/EnergyandAI.pdf) | World, accelerated servers as a stand-in for AI | 2024 | 15 percent of all data-center electricity, and 24 percent of server electricity |
| [Lawrence Berkeley National Laboratory](https://escholarship.org/content/qt32d6m0d1/qt32d6m0d1.pdf) | United States, AI servers with graphics processors | 2023 | More than 40 terawatt-hours, against 176 terawatt-hours for all data centers |
| [Electric Power Research Institute](https://restservice.epri.com/publicattachment/96784) | United States focus, AI workloads | "Currently," as of February 2026 | An estimated 15 to 25 percent of data-center electricity |

The laboratory's row counts the servers alone. It leaves out the cooling and power losses those servers cause, so it sits below a full-facility share. The research institute does not state its method or its data year.

## Why the estimates disagree

The estimates disagree because they use different boundaries, stand-ins, and places. Three choices drive most of the gap.

- **The denominator.** One estimate divides by all data-center electricity in the world. Another divides by server electricity alone, or by the United States alone. A share of one total does not compare with a share of another.
- **The stand-in.** The [International Energy Agency's Energy and AI report](https://iea.blob.core.windows.net/assets/de9dea13-b07d-42c5-a398-d1b3ae17d866/EnergyandAI.pdf) counts accelerated servers in place of AI. It notes that some AI runs on conventional servers, and some scientific computing runs on accelerated servers.
- **Missing data.** The same report states that no comprehensive data exists on the share of each workload. Colocation operators often have limited visibility into what their tenants run.

The [laboratory's 2025 update](https://escholarship.org/content/qt33m6w3x0/qt33m6w3x0.pdf) also leaves cryptocurrency mining out of its totals. Two honest numbers can differ by a wide margin for these reasons. A fair comparison holds the year, the place, and the boundary constant.

## How much electricity AI could use by 2030

AI's share of data-center electricity is projected to rise steeply by 2030. Each figure below is a scenario, not a measurement.

- **World, by facility type.** In the [International Energy Agency's 2026 update](https://iea.blob.core.windows.net/assets/3179f7f8-01f6-4dd6-bffa-c9f7b73f1dc9/KeyQuestionsonEnergyandAI.pdf), electricity for AI-focused data centers more than triples to about 465 terawatt-hours in 2030. That is the agency's Base Case, in which all data centers reach about 950 terawatt-hours. The agency runs three other cases, but that discussion gives no AI-only figure for them.
- **United States, by equipment.** In the [laboratory's 2025 update](https://escholarship.org/content/qt33m6w3x0/qt33m6w3x0.pdf), AI servers reach 55 percent of United States data-center electricity in 2030 in its Reference Case. Two of its alternative scenarios give 53 percent, with fewer chips deployed, and 59 percent, with higher inference energy.

The growth has started. The agency estimates that electricity use by AI-focused data centers grew 50 percent in 2025, against 17 percent for all data centers.

The two projections measure different things. An AI-focused facility counts a whole building, including any conventional servers inside it. AI servers count equipment, wherever it sits. Neither number converts into the other.

> **The short version:** AI is a growing minority of data-center electricity today. The share in any estimate depends on what it counts, where, and in which year. Projections for 2030 approach half, and they are scenarios.

## Training versus inference

Training builds an AI model, and inference runs the finished model to answer requests. The [laboratory's 2025 update](https://escholarship.org/content/qt33m6w3x0/qt33m6w3x0.pdf) describes training as feeding a model large datasets at high, sustained server use. It describes inference as running a trained model in response to user inputs, often where delay matters.

Inference now draws most of the power. The laboratory assumes that about 35 percent of AI server power went to training in 2024, falling to about 20 percent by 2030. The [agency's 2026 update](https://iea.blob.core.windows.net/assets/3179f7f8-01f6-4dd6-bffa-c9f7b73f1dc9/KeyQuestionsonEnergyandAI.pdf) states that AI energy use has already shifted decisively from training to inference.

The two loads behave differently. Training servers run at high use for long periods, about 80 percent in the laboratory's model. Inference servers are sized for the peak, when many users send requests at once, and run well below capacity at other times. That difference shapes how a campus draws power. [What is an AI data center?](/insights/ai-data-center/) covers how training and inference sites differ on land.

## How much energy one AI request uses

One simple AI text request uses a small amount of electricity, and heavier tasks use far more. The [agency's 2026 update](https://iea.blob.core.windows.net/assets/3179f7f8-01f6-4dd6-bffa-c9f7b73f1dc9/KeyQuestionsonEnergyandAI.pdf) reports that short text queries can use fractions of a watt-hour on modern systems. A high-resolution image can take about ten times more energy per request. Video generation can take hundreds to thousands of times more.

The agency also checks the total. If all conventional internet searches ran as simple AI text queries, they would use less than 4 terawatt-hours a year. That is less than 1 percent of data-center electricity today. The agency concludes that most planned capacity must serve other work, such as model training, video, and multistep agents.

Efficiency improves fast, but use grows faster. The agency reports that energy per AI task has dropped by at least an order of magnitude a year in recent years. The laboratory finds that the scale of computing demand still outruns those gains, so total electricity keeps rising.

## What AI's share means for a proposed site

A national share does not tell you what a proposed building will run. A data center can host AI, conventional computing, or both, and its label does not settle the question. [Are data centers bad? The objections, checked against the record](/insights/data-center-misconceptions/) shows how the workloads divide.

The local questions are the load, the ramp, and the grid connection. [Data center power requirements: what a campus needs from the grid](/insights/data-center-power-requirements/) explains what a campus asks of the utility. [Large load interconnection: how a very large customer gets power](/insights/large-load-interconnection/) covers the study steps. [How much electricity do data centers use?](/insights/data-center-energy-consumption/) gives the totals behind this page.

> **Take action:** If you own land near transmission lines in North Carolina, learn what the public record says about it before a buyer calls. Start with [Is your parcel FOUND?](/parcel-check/)

## Key recap

- No meter separates AI from other computing. Every AI share is a modeled estimate with its own boundary.
- Published estimates for recent years put AI well under half of data-center electricity.
- Global and United States estimates use different denominators. Read each one on its own terms.
- Projections for 2030 approach half of data-center electricity, and they are scenarios.
- Inference, not training, now takes most AI server power, and its share is projected to grow.

## Questions

### What percentage of data-center electricity is AI?

It depends on the estimate. The International Energy Agency puts accelerated servers at 15 percent of global data-center electricity in 2024. The Electric Power Research Institute estimates 15 to 25 percent today, with a United States focus. Each uses a different boundary.

### Does a single AI request use a lot of electricity?

A simple text request uses fractions of a watt-hour, according to the International Energy Agency. Images, video, and multistep agent tasks use many times more per request.

### Does training or inference use more power?

Inference uses more. A national laboratory assumes that training drew about 35 percent of AI server power in 2024, falling to about 20 percent by 2030.

### Will AI use half of all data-center electricity?

Possibly, by 2030. Scenarios from the International Energy Agency and a national laboratory approach or pass half, but both are projections, not measurements.

## References

Primary sources cited on this page, in APA style.

- Electric Power Research Institute. (2026, February). *EPRI Powering Intelligence 2026 FAQs*. [https://restservice.epri.com/publicattachment/96784](https://restservice.epri.com/publicattachment/96784)
- International Energy Agency. (2025, April 10). *Energy and AI*. [https://iea.blob.core.windows.net/assets/de9dea13-b07d-42c5-a398-d1b3ae17d866/EnergyandAI.pdf](https://iea.blob.core.windows.net/assets/de9dea13-b07d-42c5-a398-d1b3ae17d866/EnergyandAI.pdf)
- International Energy Agency. (2026, April 16). *Key questions on energy and AI*. [https://iea.blob.core.windows.net/assets/3179f7f8-01f6-4dd6-bffa-c9f7b73f1dc9/KeyQuestionsonEnergyandAI.pdf](https://iea.blob.core.windows.net/assets/3179f7f8-01f6-4dd6-bffa-c9f7b73f1dc9/KeyQuestionsonEnergyandAI.pdf)
- Shehabi, A., Smith, S. J., Hubbard, A., Newkirk, A., Lei, N., Siddik, M. A. B., Holecek, B., Koomey, J., Masanet, E., & Sartor, D. (2024, December). *2024 United States data center energy usage report* (LBNL-2001637). Lawrence Berkeley National Laboratory. [https://doi.org/10.71468/P1WC7Q](https://doi.org/10.71468/P1WC7Q)
- Smith, S. J., Hubbard, A., Newkirk, A., Ganeshalingam, M., Holecek, B., Sartor, D., Mills, M., & Shehabi, A. (2026, June). *United States data center energy usage report: 2025 update* (LBNL-2001758). Lawrence Berkeley National Laboratory. [https://doi.org/10.71468/P1RP4F](https://doi.org/10.71468/P1RP4F)

## Power for data-center sites

- [Power decides the site. The map only starts the question.](/what-we-check/power-and-substations/)
- [A substation next door is not power](/insights/power-proximity-is-not-capacity/)
- [Data center power requirements: what a campus needs from the grid](/insights/data-center-power-requirements/)
- [Electric cooperative vs investor-owned utility: what changes for a site](/insights/electric-cooperative-vs-investor-owned-utility/)
- [What a will-serve letter is, and what it is worth](/insights/what-is-a-will-serve-letter/)
- [What a substation is and what it tells you about land](/insights/what-is-a-substation/)
- [Utility easements: what they allow and how they shrink usable land](/insights/utility-easement/)
- [Data center redundancy: what it means and what it asks of land](/insights/data-center-redundancy/)
- [Large load interconnection: how a very large customer gets power](/insights/large-load-interconnection/)
- [Behind-the-meter data centers: on-site power and what it asks of land](/insights/behind-the-meter-data-center/)
- [How much electricity do data centers use?](/insights/data-center-energy-consumption/)
- [Do data centers raise electric bills in North Carolina?](/insights/do-data-centers-raise-electric-bills/)
- [Why are so many data centers being built?](/insights/why-are-so-many-data-centers-being-built/)
- [Data center generators: backup power and air permits](/insights/data-center-generators/)

Send the requirements. We reply with how we would run the search.

[Check a parcel](/parcel-check/)

## Tell us what the project needs

Load, acreage, counties, and timeline. We reply with how we would run the search and where we would start.

[Start a site search](/contact/)

[Try the calculators](/tools/)

Transactions are handled through a licensed North Carolina real estate brokerage.