Sustainable AI infrastructure means building and running the data centers, chips, and power supply behind AI so that energy use, emissions, and local resource impacts stay as low as practical. The International Energy Agency (IEA) projects global data center electricity use will roughly double from about 485 TWh in 2025 to about 950 TWh in 2030, which makes efficiency, clean power, and grid planning central rather than optional.
Last reviewed: October 2026.
AI feels weightless on a screen, but physical servers produce every answer in physical buildings drawing real electricity. This article on sustainable AI infrastructure walks through what the credible numbers say about AI power demand, where the energy actually goes, what operators and utilities can do about it, and what remains uncertain. Above all, it is for readers who want a sober, source-based picture rather than either hype or alarm. We rely on the IEA and the U.S. Department of Energy's Lawrence Berkeley National Laboratory and flag where figures are projections.
How much electricity do data centers use
The IEA's April 2025 Energy and AI report estimated that data centers consumed about 415 TWh of electricity in 2024, roughly 1.5% of global electricity use, after growing about 12% a year over the previous five years. A later IEA publication, Key Questions on Energy and AI, reports that global data center demand grew 17% in 2025 and that consumption by AI-focused data centers grew 50%. It projects total data center consumption roughly doubling from 485 TWh in 2025 to 950 TWh in 2030, about 3% of global electricity demand.
In the United States, a Lawrence Berkeley National Laboratory report prepared for the Department of Energy found that data centers used 176 TWh in 2023, about 4.4% of U.S. electricity, and projected 325 to 580 TWh by 2028, or about 6.7% to 12%, depending on how fast the rest of the economy grows. In other words, the wide range is a reminder that these are scenarios, not forecasts carved in stone.
Where the energy goes inside a facility
The IEA notes that servers account for about 60% of data center electricity demand on average, with cooling making up most of the rest of the overhead. Cooling's share varies enormously: about 7% in efficient hyperscale facilities, but more than 30% in less efficient enterprise sites. That spread is the single clearest evidence that, for sustainable AI infrastructure, design and operations matter, not just total demand.
Meanwhile, chips are another driver. The IEA reports that the power density of AI servers rose 11 times between 2020 and 2025 and is set to rise about fourfold more by 2027, to the point where a single rack could at peak demand match the usage of around 65 households. As a result, dense racks push operators toward more advanced cooling and heavier electrical connections.
Efficiency is real, but so is rebound
The IEA says energy use per AI task has fallen by at least an order of magnitude per year in recent years, and that a simple text query uses less electricity than running a television for the same time. That is certainly the optimistic half of the story.
However, the other half is that the same report says video generation, reasoning, and agentic tasks can use hundreds to thousands of times more energy per query than a simple text request. When tasks get cheaper, people tend to run many more of them and attempt harder ones, so total demand can rise even as each unit gets more efficient. If you are curious about the shift toward systems that act on your behalf, see our explainer on AI agents.
The main levers for sustainable AI infrastructure
| Lever | What it does | Honest limits |
|---|---|---|
| More efficient chips and software | Gets more work per unit of electricity | Savings can be offset by growing usage |
| Better cooling and facility design | Cuts the overhead share of power use | Retrofitting older enterprise sites is costly |
| Cleaner electricity supply | Reduces emissions per kWh | Depends on local grids and new generation being built |
| Siting near spare grid capacity | Avoids overloading local networks | Good sites are limited and grid connections can be slow |
| Storage and flexible load | Smooths sharp swings in demand | Requires coordination with utilities |
| Right-sizing models for tasks | Uses smaller models when they are enough | Requires discipline in product design |
The grid is the bottleneck
According to the IEA, slow grid connections are pushing some U.S. developers toward onsite gas generation, with around 15 to 27 GW of onsite natural gas potentially powering data centers by 2030, mostly in the United States. It also says reliable onsite gas power for variable data center loads requires overbuilding generation capacity by 30% to 70%, and that around 20 to 25 GW of battery storage could be installed in data centers globally by 2030 to manage sharp load swings. The IEA projects data center emissions could double to around 350 million tonnes in 2035, about 2% of global electricity sector emissions.
The takeaway is that sustainable AI infrastructure is partly an energy-system question: how fast clean generation and transmission can be built relative to demand, and who pays for the upgrades.
Can AI help the energy transition too
The IEA argues that AI applications could also reduce energy use elsewhere. It says well-documented use cases could save more than 13 exajoules by 2035, about 3% of global final energy consumption, if barriers are overcome, and that AI-enabled production optimization could cut energy costs in energy-intensive industries. Of course, those are estimates conditional on adoption. Research uses such as AI in materials science illustrate how the technology might contribute, but benefits are not automatic, so measure them.
What to ask about any AI claim of sustainability
- Is the number absolute or relative? “Per query” savings can coexist with rising totals.
- Which electricity? Annual matching with purchased clean energy is not the same as hour-by-hour clean supply.
- Is water included? Cooling can use water; for example, operators and local authorities publish figures, and local conditions differ.
- Is it independently verified? Look for audited or third-party data rather than a marketing page.
- What is the scope? Training, inference, hardware manufacturing, and construction are different pieces.
What sustainable AI infrastructure means for you
- As a consumer or small business, you do not need to feel guilty about ordinary AI use. Rather, heavy generation such as long video or very large automated workflows is where the energy cost rises.
- As a buyer of AI services, ask vendors for their energy and water reporting and choose smaller models when they are good enough.
- As a citizen, local data center proposals are decided at the utility commission and zoning level. Attend hearings, read rate case summaries, and ask who bears grid upgrade costs.
- As an investor or employee, treat energy availability as a real constraint on growth, not a footnote.
Two metrics that reveal sustainable AI infrastructure: PUE and WUE
In practice, operators report a handful of standard metrics, and two of them tell you most of what you need. Learning to read them turns a marketing claim into something you can check.
Power usage effectiveness (PUE) divides total facility electricity by the electricity that reaches the IT equipment. A PUE of 1.0 would mean zero overhead for cooling, lighting, and power conversion. Large modern facilities often report figures a little above 1.1, while older enterprise rooms can sit well above 1.5. Consequently, a lower PUE means less wasted energy for the same computing work.
Water usage effectiveness (WUE) divides the water consumed on site, mostly for evaporative cooling, by IT energy, giving litres per kilowatt-hour. A low WUE can hide water used upstream at power plants, so ask whether the figure covers only the building.
Neither metric says anything about where the electricity comes from. Therefore, sustainable AI infrastructure needs a third piece of information: the carbon intensity of the grid or contracted supply, ideally measured hour by hour rather than averaged over a year.
Worked example: what a 100 MW data center means locally
Project announcements quote capacity in megawatts, which is hard to picture. Here is an illustrative calculation with rounded numbers.
- A facility with a 100 MW load running flat out for a year uses 100 MW multiplied by 8,760 hours, or about 876,000 MWh.
- If an average home in the region uses roughly 10,000 kWh a year, that is electricity for around 85,000 homes, before counting the data center’s own cooling overhead.
- With a PUE of 1.2, overhead adds another 20%, pushing the total toward 1,050,000 MWh.
- Admittedly, real facilities rarely run at full load every hour, so actual consumption is usually lower. Even so, the grid connection must be sized for the peak.
The example shows why utilities treat large projects as a planning event. Transmission lines, substations, and generation must exist before the servers switch on, and someone pays for that capacity whether or not the facility runs at full tilt. That is the practical heart of sustainable AI infrastructure.
Common misconceptions about sustainable AI infrastructure
- One query is a scandal. As the IEA figures show, a short text query uses little energy. The concern is volume and the shift to heavier tasks, not a single question.
- Renewable contracts make a facility clean. Buying clean energy certificates over a year does not guarantee that clean power is flowing at 3 a.m. in winter. Hourly matching is the stricter standard.
- Efficiency gains will solve it. Efficiency has improved enormously, yet total demand still rises because usage grows faster. Both facts are true at once.
- Water is a minor footnote. In dry regions, cooling water can matter more to the community than electricity, which is why local hearings often focus on it.
- Only hyperscalers matter. Smaller enterprise rooms are often far less efficient, so modernising them is a real, if unglamorous, lever.
A checklist for evaluating a local data center proposal
If a project is proposed near you, these questions help you form a view quickly. They are the same questions that regulators and informed investors ask.
- What is the contracted grid capacity, and who pays for any new lines or substations?
- What PUE and WUE does the operator commit to, and will it publish measured figures?
- Where will the electricity come from in the first five years, and will any of it be generated on site with gas?
- Does the design include battery storage or demand flexibility to ease peak loads?
- Is waste heat captured for nearby buildings or industry, as some European facilities do?
- What local jobs exist after construction ends, and what tax arrangements apply?
Finally, remember that the energy question connects to the privacy and security questions covered in our guide to AI data privacy; both are about making a fast-moving technology accountable.
Sources and further reading
- IEA, Energy and AI: Energy demand from AI (April 2025)
- IEA, Key Questions on Energy and AI: Executive summary (2026)
- Lawrence Berkeley National Laboratory: Report evaluates increase in electricity demand from data centers
Frequently asked questions
What is sustainable AI infrastructure?
It is the combination of efficient hardware, well-designed data centers, clean and reliable power, and responsible siting that keeps the energy, emissions, and local resource impacts of AI as low as practical.
How much electricity do data centers use?
The IEA estimates about 415 TWh in 2024 (about 1.5% of global electricity) and reports about 485 TWh in 2025. In the U.S., Berkeley Lab estimated 176 TWh in 2023, about 4.4% of national electricity.
Will AI double data center electricity demand?
The IEA projects total data center consumption roughly doubling between 2025 and 2030 to about 950 TWh. That is a projection, and the IEA also models faster and slower scenarios.
Is a single AI query a big energy user?
The IEA says a simple text query uses less electricity than running a television for the same time, but video generation, reasoning, and agentic tasks can use hundreds to thousands of times more energy per query.
Does cooling matter that much?
Yes. According to the IEA, cooling is about 7% of consumption in efficient hyperscale facilities and over 30% in less efficient enterprise facilities.
Why are grids a problem for AI growth?
Connecting large new loads to the grid can be slow, and the IEA notes this is pushing some U.S. developers toward onsite gas generation. Expanding clean generation and transmission takes time.
More questions about sustainable AI infrastructure
Can AI reduce energy use elsewhere?
The IEA says some well-documented use cases could save significant energy by 2035, but only if barriers are overcome, so the benefit is conditional rather than guaranteed.
How can I judge a company claim about green AI?
Check whether numbers are absolute or per query, whether clean power is matched hourly or annually, whether water is reported, and whether data are independently verified.
What is PUE in a data center?
Power usage effectiveness is total facility energy divided by the energy delivered to IT equipment. A value near 1.0 means little overhead, and lower is better. It does not describe where the electricity comes from.
Do data centers use a lot of water?
Some do, mainly for evaporative cooling, and the amount varies with climate and design. Operators that use closed-loop or air cooling consume less water on site, although water is also used upstream at power plants.
Can renewable energy fully power sustainable AI infrastructure?
In principle, but it requires enough clean generation, storage, and transmission to supply power every hour, not just on average over a year. The IEA notes that slow grid connections are a key constraint today.
Disclaimer: This article provides general information only and is not investment, engineering, or policy advice. Figures are projections from the cited sources as of October 2026 and may be revised. Check the original reports for details.