The AI Boom Has a Power Problem: How Data Centers Are Reshaping Global Electricity Demand

Artificial intelligence may live on screens, but the infrastructure behind it is intensely physical.

Every AI-generated image, chatbot response, coding assistant and increasingly sophisticated reasoning system ultimately depends on processors operating inside data centers. Those processors require electricity. They generate heat that must be removed. They depend on networks, substations, transformers and power plants capable of supplying electricity reliably around the clock.

As AI adoption accelerates, that physical infrastructure is becoming one of the defining constraints of the technology boom.

The numbers illustrate the scale of the change.

Data centers consumed roughly 485 terawatt-hours (TWh) of electricity worldwide in 2025, according to the International Energy Agency’s latest assessment. By 2030, the agency’s central projection puts consumption at approximately 950 TWh — nearly double today’s level and around 3% of global electricity demand. Electricity consumption specifically from AI-focused data centers is expected to grow considerably faster, roughly tripling over the same period. IEA

That does not mean AI is about to consume the world’s electricity supply. Three percent of global electricity demand remains a relatively small share.

But the global percentage hides a much more complicated problem.

Data centers are concentrated geographically, their power requirements can be enormous, and new facilities can sometimes be built considerably faster than the electricity infrastructure required to serve them.

The AI race is therefore becoming an energy race as well.

Why does AI require so much electricity?

Traditional data centers already operate millions of servers supporting websites, cloud computing, streaming, online banking, business applications and countless other digital services.

AI adds another layer of computational intensity.

Training large artificial-intelligence models requires huge numbers of calculations performed across specialized processors. Once those models are deployed, they continue consuming computing resources every time users interact with them.

And not every AI request is equal.

Generating a short piece of text, producing an image, creating video or asking an advanced reasoning system to work through a complex task can impose very different computational requirements.

The important issue is therefore not simply how many people use AI.

It is also what they ask AI systems to do.

As AI moves from basic text generation toward multimodal models, autonomous agents, scientific computing, software development and increasingly complex reasoning, the amount of computation associated with some tasks can rise substantially.

At the same time, efficiency is improving. Chips are becoming more capable, models can be optimized, and data-center operators continue finding ways to obtain more computing output from each unit of electricity.

The future electricity footprint of AI will consequently depend on a race between two forces: rapidly expanding demand for computation and rapidly improving efficiency.

The IEA explicitly acknowledges this uncertainty. Its projections vary according to AI adoption, hardware and software efficiency, server deployment and potential energy-system bottlenecks. IEA

The world’s data centers are approaching a new scale

The expansion is already visible.

Global electricity consumption by data centers increased by approximately 17% in 2025, representing an additional 70 TWh of consumption, according to the IEA’s 2026 Global Energy Review. IEA

And the investment behind that growth is extraordinary.

The IEA reported in April that capital expenditure by five large technology companies exceeded $400 billion in 2025 and was expected to rise another 75% in 2026, driven by the expansion of AI and data-center infrastructure. IEA

This is turning what was once primarily a technology-sector expansion into a major infrastructure buildout.

A modern AI facility does not simply need racks of expensive processors. It can require new transmission connections, substations, transformers, cooling equipment, backup systems and sometimes entirely new sources of electricity generation.

That creates a fundamental mismatch in development times.

A data center can potentially become operational within a few years. Major transmission infrastructure and new generating capacity can take considerably longer to plan, approve and construct. The IEA identifies this difference in lead times as one of the uncertainties surrounding future data-center expansion. IEA

Computing capacity can therefore grow faster than the infrastructure needed to power it.

The United States is at the center of the electricity challenge

The effect is particularly striking in the United States.

Data centers consumed approximately 180 TWh of electricity in the United States in 2024, according to IEA estimates. IEA

Their influence has continued to grow.

In 2025, rapidly expanding data-center loads accounted for roughly half of the increase in US electricity consumption, according to the agency’s Global Energy Review. IEA

Looking toward 2030, the IEA expects data centers to account for around half of US electricity-demand growth over the forecast period. IEA

That represents a significant change for an electricity system that spent years experiencing comparatively modest demand growth.

The issue is not necessarily whether the United States can generate enough electricity in aggregate. Electricity infrastructure is regional, and data centers tend to cluster around locations offering connectivity, land, favorable economics and access to power.

That means pressure can emerge locally long before it becomes obvious in national statistics.

A large concentration of facilities may require additional transmission capacity, new substations or generation projects in a particular region.

Increasingly, the question facing technology companies is therefore not simply:

Where can we build a data center?

It is:

Where can we obtain enough reliable electricity to operate it?

AI is arriving during a much bigger electricity boom

AI is also not the only technology increasing electricity demand.

Electric vehicles are expanding. Air-conditioning use is growing. Heat pumps are spreading. Manufacturing is becoming more electrified. Emerging economies continue adding appliances, industrial capacity and infrastructure.

The IEA expects worldwide electricity demand to grow at an average annual rate of about 3.6% between 2026 and 2030, with annual demand growth over that period averaging about 50% more than during the previous decade. IEA

The agency’s July 2026 update forecasts global electricity demand increasing 3.6% in 2026 and 3.8% in 2027, compared with 3% growth in 2025. IEA

China remains the largest source of overall electricity-demand growth, while India and Southeast Asia are also expanding rapidly.

The IEA forecasts average electricity-demand growth through 2030 of approximately 6.4% annually in India and 5.3% in Southeast Asia, compared with 4.9% in China. IEA

That gives the AI infrastructure race an international dimension.

Future data centers will compete for electricity within a world that already needs substantially more of it.

Where will all the additional power come from?

There will probably be no single answer.

Renewables are expected to supply a substantial part of the increase.

The IEA projects renewable electricity generation supplying data centers to grow by more than 450 TWh through 2035, while natural gas and nuclear power also contribute significantly. IEA

Solar and wind have advantages including comparatively short development timelines in many markets and increasingly competitive costs. But data centers operate continuously, meaning operators also need dependable power when renewable generation is low.

That increases the importance of grids, energy storage and dispatchable electricity sources.

Natural gas is consequently expected to remain part of the equation in several markets.

Nuclear energy is attracting renewed attention as well.

Microsoft, for example, has entered a power purchase agreement supporting the planned restart of the former Three Mile Island Unit 1 — now called the Crane Clean Energy Center — as a source of round-the-clock carbon-free electricity for its data-center needs. Microsoft

Meanwhile, technology companies are exploring emerging nuclear technologies and other forms of reliable low-carbon generation.

The IEA expects the first small modular reactors in its data-center energy outlook to begin contributing around 2030. IEA

The result may be an unusually diverse energy mix: solar, wind, batteries, natural gas, existing nuclear plants, new nuclear technologies, geothermal resources and expanded transmission networks all contributing in different locations.

Big Tech is becoming an energy player

One of the most consequential changes may be happening inside the technology industry itself.

For decades, electricity was largely something technology companies purchased.

Increasingly, securing electricity is becoming part of their strategic infrastructure planning.

Google reported that it signed agreements for more than 12 GW of new clean-energy capacity during 2025, even as the company’s electricity demand continued to grow. blog.google

Microsoft’s nuclear agreement illustrates another approach.

These developments blur the traditional boundary between technology infrastructure and energy infrastructure.

The companies competing to build the world’s most powerful AI systems increasingly have to think about processors and models alongside substations, generation capacity, cooling technology and grid availability.

The AI industry’s competitive advantage may therefore depend partly on something decidedly old-fashioned:

access to power.

Could electricity become a bottleneck for AI?

It already can be — particularly at the local level.

The problem is less about the world literally running out of electricity than about delivering enough power to the right place at the right time.

New generation must connect to transmission networks. Transformers and other grid equipment have manufacturing lead times. Projects require planning and construction. Electricity systems must also remain reliable while serving homes, businesses, factories and transportation.

The IEA’s updated 2026 assessment says bottlenecks throughout the data-center value chain are reducing the likelihood of the most aggressive near-term expansion scenarios, even though investment and project pipelines remain strong. IEA

This creates an unusual constraint on AI development.

For years, discussions about the future of artificial intelligence concentrated on algorithms, chips, data and talent.

The next phase increasingly includes another scarce resource:

megawatts.

Efficiency could change the equation

There is an important counterweight to all of these projections.

Technology improves.

AI processors can become more efficient. Models can require less computation. Cooling systems can improve. Workloads can be shifted geographically or scheduled for periods when electricity is more readily available.

Software optimization can also dramatically reduce the computing resources necessary for particular tasks.

This is why forecasts of AI electricity consumption should not be interpreted as predetermined outcomes.

The IEA uses multiple scenarios precisely because future demand depends heavily on efficiency improvements, AI adoption and infrastructure constraints. IEA

A breakthrough that makes AI inference significantly more efficient could reduce electricity requirements per task.

But efficiency can produce another effect: when something becomes cheaper to use, people often use more of it.

If efficient AI makes it economically practical to embed sophisticated models in billions of everyday interactions, total computation could continue increasing even as individual calculations become less energy intensive.

AI could also help the electricity system

There is another side to the relationship.

Artificial intelligence is not merely an electricity consumer.

It could become a tool for improving electricity systems themselves.

AI systems can potentially help forecast demand, predict renewable generation, optimize industrial energy use, identify equipment problems, manage buildings and improve operation of increasingly complex grids.

That creates an unusual feedback loop.

Electricity enables AI.

AI could, in turn, help make electricity systems more efficient.

Whether those efficiency gains offset a meaningful portion of AI’s own energy footprint will depend on how quickly useful applications move from experiments into real-world deployment.

The next AI race may be fought on the power grid

The popular image of artificial intelligence is almost entirely digital: a chatbot window, an image generator or software quietly producing an answer.

Behind that interface is an expanding industrial system.

Factories manufacture advanced chips. Data centers house them. Cooling systems remove their heat. Transmission lines deliver electricity. Power plants and renewable projects generate it.

The physical footprint of AI is becoming impossible to separate from the digital revolution it enables.

By 2030, global data-center electricity consumption could be roughly twice its 2025 level. Yet the more important story may not be the global percentage.

It is the speed at which a new source of electricity demand is appearing, the locations where it is concentrated, and the infrastructure race now developing around it.

The countries and companies that dominate the next era of artificial intelligence may need more than the best algorithms and fastest processors.

They may also need something much more fundamental:

enough electricity to turn them on.