AI Is Doubling Data Centre Power: 415 TWh Today, 945 TWh by 2030 ⚡

Every time you ask an AI a question, something spins up in a warehouse.

That warehouse draws power. A lot of it.

In 2024, the world’s data centres used 415 terawatt-hours of electricity (International Energy Agency, 2026).

By 2030, the IEA expects that to reach about 945 terawatt-hours.

That is more electricity than Japan uses today. For one industry.

This article looks at what that number means, where it comes from, and what it does not tell you. ⚡

🧾 Key Findings at a Glance

Measure	Figure	Source
Data centre power use, 2024	415 TWh	IEA (2026)
Share of world electricity, 2024	About 1.5%	IEA (2026)
Projected use by 2030	~945 TWh	IEA (2026)
AI-specific data centre growth	More than quadruple by 2030	IEA (2026)
US share of 2024 total	45%	IEA (2026)
China share of 2024 total	25%	IEA (2026)
Europe share of 2024 total	15%	IEA (2026)

⚡ The Doubling, In Plain Terms

The headline is simple. Data centre electricity use roughly doubles by 2030.

From 415 TWh to around 945 TWh (International Energy Agency, 2026).

A terawatt-hour is a billion kilowatt-hours.

Your home probably uses a few thousand kilowatt-hours a year.

So 945 TWh is roughly the yearly power of a few hundred million homes.

Why 1.5% is the more useful number

Data centres used about 1.5% of world electricity in 2024.

That is small. Smaller than most people guess.

The concern is not the current size. It is the speed of change.

Most electricity uses grow slowly. This one is doubling in six years.

A doubling is not a crisis by itself

Plenty of things have doubled without breaking the grid.

Air conditioning did. So did electric rail.

The question is whether supply can grow at the same pace.

In some places it can. In others it cannot.

🇺🇸 Why America Feels This First

The United States used 45% of the world’s data centre electricity in 2024 (International Energy Agency, 2026).

China used 25%. Europe used 15%.

That concentration matters more than the global total.

The growth statistic that worries planners

Here is the line that gets quoted in policy meetings.

US data centres are set to account for almost half of America’s electricity demand growth to 2030 (International Energy Agency, 2026).

Not half of electricity. Half of the growth in it.

Every new power station, roughly speaking, has a data centre waiting for it.

Why they cluster

Data centres are not spread evenly. They bunch together.

They follow cheap power, cool weather and fast internet cables.

A handful of counties host an outsized share.

That means local grids feel national growth.

A county with three data centres can see demand jump like a small city arrived.

🤖 The AI Part of the Story

Not all data centre power goes to AI. Most of it still does not.

Data centres also run email, video, banking and cloud storage.

But AI is the fastest-growing slice.

The IEA expects electricity for AI-optimised data centres to more than quadruple by 2030 (International Energy Agency, 2026).

Workload	Growth to 2030
AI-optimised data centres	More than 4x
All data centres combined	Roughly 2x
Traditional cloud and storage	Slower than the average

Read those rows together.

If the total doubles while AI quadruples, AI is taking share from everything else.

Training versus asking

People often confuse two very different costs.

Training is building the model. It happens once and uses enormous power.

Inference is using the model. Each use is tiny.

But inference happens billions of times.

Over a model’s life, all those small uses can outweigh the training.

Why this distinction matters

Headlines usually quote training costs. They are dramatic and easy to picture.

The long-run bill is driven by ordinary daily use.

A model used by millions costs more to run than it did to build.

📉 The Efficiency Argument, Fairly Stated

There is a real counter-argument here, and it deserves space.

Chips get more efficient every year. So do cooling systems.

Data centres today do far more work per unit of power than a decade ago.

For years, computing grew hugely while power use stayed nearly flat.

Why that stopped working

Efficiency gains have limits.

You cannot halve energy use forever.

AI demand is now growing faster than efficiency is improving.

That is the shift the IEA figures capture.

The rebound effect

There is a twist worth knowing.

When something gets cheaper to run, people use more of it.

Cheaper AI means more AI, not the same amount for less power.

Economists call this the rebound effect. It is well documented.

Efficiency alone rarely reduces total demand.

📈 Putting 945 Terawatt-Hours in Context

Big energy numbers are hard to feel. Comparisons help.

The IEA notes 945 TWh is slightly more than Japan’s total electricity use today (International Energy Agency, 2026).

Japan is the world’s fourth-largest economy. It has about 124 million people.

By 2030, data centres alone could out-consume that entire country.

Comparison	Rough scale
Data centres, 2024	415 TWh
Data centres, 2030 forecast	~945 TWh
Japan, total electricity today	Slightly under 945 TWh
Share of world power, 2024	About 1.5%

Now the other side of the comparison.

World electricity use is somewhere near 30,000 TWh a year.

Even at 945 TWh, data centres stay a low single-digit share.

Both things are true at once. It is a lot of power. It is also a small slice.

Why both framings get used

People who want you alarmed quote the Japan comparison.

People who want you relaxed quote the 1.5% share.

Neither is lying. They are choosing which true fact to lead with.

When a statistic feels designed to make you feel something, look for its opposite framing.

The honest summary

Data centres are a small but fast-growing part of world electricity.

They are a large and fast-growing part of electricity growth.

That second sentence is where the pressure lives.

🏗️ Why Grids Cannot Simply Keep Up

Building power supply is slow. Much slower than most people assume.

Infrastructure	Typical time to build
Data centre	1–2 years
Solar or wind farm	2–4 years
Gas power station	3–5 years
High-voltage transmission line	7–12 years
Nuclear plant	10–15 years

Look at the top and bottom rows together.

A data centre can be built ten times faster than the transmission line feeding it.

That mismatch is the entire infrastructure problem in one comparison.

What operators do about it

Some are buying their own power directly.

Others sign long deals with generators before building.

A few are looking at putting plants on the same site.

When a company builds its own power station, that tells you the grid said no.

Why this shows up on your bill

Grids are shared. Upgrades get paid for by everyone connected.

If a region adds huge demand quickly, network costs rise.

Who pays is a political question, and it is being argued now in several countries.

🌱 What This Means for Emissions

Electricity use and emissions are not the same thing.

A data centre on hydro power emits far less than one on coal.

So the emissions picture depends on where the buildings sit.

Factor	Effect on emissions
Local grid mix	The single biggest factor
Time of day used	Night power is often dirtier or cleaner by region
On-site renewables	Helps, but rarely covers full load
Cooling method	Water cooling cuts power but uses water
Chip generation	Newer chips do more work per watt

Water is the quiet issue in that table.

Some cooling systems trade electricity for water use.

That is a real trade, not a free saving.

Why “100% renewable” claims need care

Many operators claim to run on renewable energy.

Often this means they buy certificates equal to their use.

The actual electrons still come from the local grid.

That is not dishonest. But it is not the same as running on solar.

Ask whether power is matched hourly or just annually. The difference is large.

💡 What Actually Changes Because of This

Data like this only matters if it affects decisions. Three groups feel it.

Governments must approve power stations and grid connections years ahead.

Businesses face rising or more volatile electricity prices.

Ordinary people may see bills shift if demand outpaces supply locally.

The grid connection queue

This is the practical bottleneck nobody expected.

In many regions, connecting a large site to the grid takes years.

Building the data centre is faster than powering it.

Electricity supply, not chips, is becoming the limit on AI growth.

What businesses can do

Most small firms cannot influence any of this.

But they can choose providers who publish real energy data.

And they can avoid running heavy AI jobs they do not need.

If you are picking tools, our guide to AI tools covers what is worth paying for.

💧 The Water Question Nobody Puts on the Chart

Electricity gets the headlines. Water rarely does.

Many large data centres use water to stay cool.

Evaporative cooling is cheap and uses far less electricity than chillers.

But it trades power for water, and water is local.

Why the trade is awkward

Electricity can be shipped across a country on wires.

Water cannot, in any practical sense.

So a site that saves power may strain a local supply.

In a dry region, that is the harder constraint.

How to read water claims

Operators report water use in different ways.

Some count only what evaporates. Some count everything drawn.

Some exclude the water used to generate their electricity.

That last exclusion is large, because thermal power plants use water too.

When you see a water figure, ask what it includes.

🧊 What Actually Uses the Power Inside the Building

People picture racks of servers. That is only part of it.

System	Rough share of site power
Computing hardware	The largest share
Cooling	Second largest, varies by climate
Power conversion losses	Small but constant
Lighting and building services	Minor

Cooling is the part that varies most.

A site in Iceland spends far less on it than one in Arizona.

That is why cold countries keep winning data centre investment.

The efficiency measure to know

The industry uses a ratio called power usage effectiveness.

It compares total site power to the power reaching the computers.

A perfect score is 1.0. Nothing wasted on cooling or losses.

Modern large sites often report figures close to that.

Why good ratios can still mislead

A very efficient building can still use enormous total power.

Efficiency is power per unit of work. It is not total power.

A site can improve its ratio every year and still double its consumption.

Both numbers matter. Reports often show only the flattering one.

🔬 How the IEA Built These Numbers

A forecast is only as good as its method. Here is how this one works.

The IEA is an intergovernmental body. It publishes energy data for member countries.

Its Energy and AI report drew on new datasets and consultation with the tech and energy industries.

Method feature	Why it matters
Government-backed body	No product to sell you
Uses national energy statistics	Built on measured data, not surveys
Publishes scenarios, not one answer	Shows a range of futures
Consults industry directly	Captures planned builds
Updated regularly	Revised as reality lands

Scenarios are not predictions

This is the most misread part of energy reporting.

The IEA publishes several scenarios with different assumptions.

News stories usually quote one and call it “the forecast”.

A scenario says “if these things happen, this follows”.

It is a conditional statement, not a prophecy.

Why the 945 figure could be wrong

It could be too low if AI use grows faster than expected.

It could be too high if efficiency improves faster, or if demand cools.

Forecasts six years out are genuinely uncertain.

Treat the direction as solid and the exact number as approximate.

🌍 How Different Countries Are Responding

The same forecast produces very different policies.

Approach	What it looks like in practice
Build more supply	Fast-track permits for generation
Slow the demand	Pause new data centre connections
Move the demand	Steer builds to regions with spare power
Make it flexible	Pay sites to reduce load at peak times
Charge the true cost	Make large users fund grid upgrades

The last two are the most interesting.

Why flexibility could change everything

Not all computing is urgent.

Training a model can often wait a few hours.

If those jobs run when power is plentiful, the grid strain drops sharply.

A flexible data centre is worth far more to a grid than a rigid one.

The IEA notes AI itself could help run energy systems better.

The demand-shifting idea

Solar power peaks at midday. Wind peaks unpredictably.

Computing that follows cheap power costs less and pollutes less.

Some operators already do this across time zones.

It is one of the few genuinely promising answers here.

🧠 What This Means If You Just Use AI

Most readers are not building data centres. So what applies to you?

Honestly, less than alarming headlines suggest.

One person’s individual AI use is a rounding error.

The scale here is industrial, not personal.

Where personal choices do matter

They matter in aggregate, and through what you pay for.

Choosing providers who publish real energy data creates pressure.

So does asking suppliers where their compute runs.

Those questions were rare five years ago. They are now normal in procurement.

The perspective worth keeping

Guilt about asking a chatbot a question is misplaced.

The decisions that matter are made by governments and large buyers.

Understanding the numbers is more useful than feeling bad about them.

That is really why this data is worth reading at all.

🚫 What These Numbers Do Not Tell You

Being clear about limits is part of using data well.

They do not tell you AI is bad for the planet. That depends on what AI replaces.

They do not measure emissions directly. Only electricity use.

They do not cover water use in detail. That is a separate concern.

They do not split AI from other computing precisely. The boundary is fuzzy.

They are global averages. Your country’s picture may look nothing like this.

The comparison nobody runs properly

If AI removes a commute, it may save more energy than it uses.

If it replaces a flight with a video call, the saving is large.

If it just produces more email, the saving is zero.

Net effect depends entirely on use, and almost nobody measures that.

🏁 The Short Version

Data centres used 415 TWh in 2024, about 1.5% of world electricity.

The IEA expects roughly 945 TWh by 2030 (International Energy Agency, 2026).

AI is the fastest-growing part, set to more than quadruple.

The United States carries 45% of today’s load and will feel the growth first.

The problem is not the size. It is the speed.

Grids take a decade to build. AI demand is arriving in three years.

That mismatch is the real story behind every big number here. ⚡

❓ Frequently Asked Questions

How much electricity do data centres use?

About 415 terawatt-hours in 2024, or roughly 1.5% of world electricity (International Energy Agency, 2026).

How much will they use by 2030?

The IEA projects around 945 TWh, slightly more than Japan’s total electricity use today.

Is AI the main cause?

AI is the fastest-growing driver, expected to more than quadruple. But most data centre power still runs ordinary cloud services.

Which country uses the most?

The United States, at about 45% of the 2024 global total, followed by China at 25% and Europe at 15%.

Does using AI harm the environment?

It depends on the local grid and on what the AI replaces. Electricity use is not the same as emissions.

Will better chips solve this?

Partly. Efficiency keeps improving, but demand is currently growing faster than efficiency gains.

What is the real bottleneck?

Grid connections. In many regions, powering a data centre takes longer than building one.

Are the 2030 figures reliable?

They are scenarios, not predictions. The direction is well supported. The exact number carries real uncertainty.

Do data centres use a lot of water?

Many use water for cooling, which saves electricity but strains local supply. Reporting methods vary widely, so always check what a water figure includes.

What does power usage effectiveness mean?

It compares total site power to the power actually reaching the computers. A score of 1.0 would mean nothing wasted on cooling or conversion losses.

Can data centres help the grid rather than strain it?

Yes, if their workloads are flexible. Jobs that can wait, such as model training, can run when power is plentiful and cheap.

Should I feel guilty about using AI?

Individual use is a rounding error at this scale. The decisions that matter are made by large buyers and governments.

Where can I read the original research?

The IEA publishes Energy and AI free online. It is linked in the references below.

📚 References

International Energy Agency. (2026). Energy and AI: Executive summary. Retrieved August 8, 2026, from https://www.iea.org/reports/energy-and-ai/executive-summary

International Energy Agency. (2026). Energy demand from AI. Retrieved August 8, 2026, from https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai

International Energy Agency. (2026). AI is set to drive surging electricity demand from data centres. Retrieved August 8, 2026, from https://www.iea.org/news/ai-is-set-to-drive-surging-electricity-demand-from-data-centres

Stanford Institute for Human-Centered Artificial Intelligence. (2026). The 2026 AI Index report. Retrieved August 8, 2026, from https://hai.stanford.edu/ai-index/2026-ai-index-report

Carbon Brief. (2026). AI: Five charts that put data-centre energy use and emissions into context. Retrieved August 8, 2026, from https://www.carbonbrief.org/ai-five-charts-that-put-data-centre-energy-use-and-emissions-into-context

Related reading on this site

For which AI tools are worth paying for, see our AI tools roundup. If you are comparing one-off software purchases against monthly fees, the lifetime deal versus subscription analysis covers the maths. Our startup software guide looks at keeping costs down while scaling.

About this analysis

All figures are drawn from the International Energy Agency’s published work and are named, dated and linked. The IEA publishes scenarios rather than single predictions, and that distinction is preserved here rather than flattened into a forecast. Figures were checked against the original publications on August 8, 2026.