The AI Boom Needs More Electricity Than Most People Realize

A data centre does not look much like artificial intelligence.

There is no floating chatbot in the corridor and no visible cloud. There are secured buildings, electrical switchgear, cables, cooling equipment, backup systems and rows of machines designed to operate hour after hour.

To the person using an AI service, almost none of that infrastructure is visible. Type a question, wait a moment and receive an answer.

But every interaction is ultimately an electrical event.

Chips perform calculations. Those chips consume electricity and produce heat. Cooling systems move that heat away. Networking equipment moves information between servers and eventually back towards the user.

As companies compete to train and operate increasingly capable AI systems, that physical layer has become impossible for the energy industry to ignore.

Data centres already consume hundreds of terawatt-hours

The International Energy Agency estimates that data centres consumed about 415 terawatt-hours of electricity worldwide in 2024.

That was roughly 1.5 per cent of global electricity consumption.

On its own, 1.5 per cent may sound modest. The more important number is the direction of travel.

In the IEA's base case, global data-centre electricity consumption rises to around 945 terawatt-hours by 2030, slightly more than double the 2024 level and just under 3 per cent of total global electricity demand.

AI is not responsible for every watt consumed inside a data centre. Cloud storage, streaming, ordinary web services, enterprise software and countless other digital services all share the same broad infrastructure category.

But the IEA identifies artificial intelligence as the most important driver of the projected increase.

Electricity consumption from accelerated servers, the high-performance systems heavily associated with modern AI workloads, is projected in the agency's base case to grow by around 30 per cent per year through 2030.

AI feels like software to the user. To the electricity grid, it looks like machines that need to stay powered.

America shows how quickly the equation is changing

The pressure is particularly visible in the United States, home to an enormous share of the world's hyperscale computing infrastructure.

In June 2026, Lawrence Berkeley National Laboratory published an updated national assessment of US data-centre electricity use.

Its reference case estimates that US data centres could consume around 649 terawatt-hours of electricity in 2030, equivalent to approximately 11.8 per cent of national electricity consumption.

Because the future of AI hardware deployment is uncertain, the researchers did not present one number as inevitable.

Their broader scenarios place the 2030 share between roughly 9.5 and 15.3 per cent of US electricity consumption, with compounded uncertainty scenarios spanning approximately 521 to 843 terawatt-hours.

The range is wide precisely because the industry is moving too quickly for an honest forecast to pretend certainty.

Why AI hardware changes the building

Data centres are not new.

Search engines, streaming platforms, cloud storage, payment systems, online shops and business software have relied on them for years.

What is changing is the power density being demanded by some modern AI workloads.

Conventional server infrastructure may spread a particular amount of computing work across a relatively large physical area. High-performance AI accelerators can pack extraordinary computational capacity into dense racks.

That is useful if the objective is to train or operate powerful models.

It is also an engineering challenge.

The IEA notes that a conventional data centre may have electrical capacity in the region of 10 to 25 megawatts, while an AI-focused hyperscale facility can exceed 100 megawatts.

Some planned campuses are dramatically larger still.

A facility drawing 100 megawatts continuously is no longer a trivial commercial customer from the perspective of a local electricity network.

Training is only half the story

Public discussion of AI energy use often centres on training a model.

Training can indeed require enormous clusters of processors running through vast quantities of data.

But once training ends, the model has not stopped consuming resources.

Every time somebody asks the deployed model to write text, analyse an image, generate software or answer a question, the system performs inference.

One request may represent a tiny fraction of the energy associated with a major training run.

Millions of users making requests every hour create a continuous workload.

The infrastructure problem is therefore not just how expensive it is to build the intelligence. It is how expensive it becomes to operate that intelligence at global scale.

Chips produce heat, and heat has to go somewhere

Electricity entering a server does not disappear.

Much of it ultimately becomes heat.

That makes cooling one of the fundamental engineering problems inside a modern high-density data centre.

Traditional facilities have relied heavily on air cooling and chilled-water systems. Increasing power density has accelerated interest in direct-to-chip liquid cooling and other approaches capable of removing heat more efficiently from tightly packed processors.

Microsoft, for example, has introduced newer data-centre designs using direct-to-chip cooling and says some of its newer facilities can avoid more than 125 million litres of cooling water per facility each year compared with previous designs.

That example is also a warning against simplistic claims about how much water "one AI question" uses.

Water consumption varies according to the facility, cooling method, climate, local electricity generation and time of operation.

A useful water figure has to describe a real system rather than pretend every prompt on Earth has one universal footprint.

The grid cannot be downloaded overnight

The computing industry can move remarkably quickly.

A company can order another generation of processors, lease a new data hall or announce another campus within a relatively short commercial cycle.

Electricity infrastructure operates on a different clock.

New substations, transmission lines and large generating facilities can take years to permit, finance and construct.

Berkeley Lab reported in 2026 that rapid growth from data centres and other very large loads was already creating connection bottlenecks across parts of the United States.

This difference in speed matters.

A server can arrive before the grid capacity required to operate it.

Electricity is becoming a technology advantage

For decades, regions competed for technology investment using familiar incentives: skilled workers, low taxes, cheap land and good telecommunications.

Available electricity is increasingly joining that list.

A cheap parcel of land has limited value to a hyperscale operator if the nearest network connection cannot support another major load for several years.

Conversely, locations with strong transmission networks, available generation and predictable planning processes can become attractive even if other costs are higher.

Computing is global in use but stubbornly local in construction.

Every apparently placeless cloud service eventually occupies land somewhere and connects to somebody's electricity system.

Technology companies are making energy deals that once looked unusual

The clearest sign that electricity has become strategic is the behaviour of the technology companies themselves.

They are no longer treating power only as a utility bill paid after a building exists.

They are signing long-term agreements intended to influence what generation exists in the first place.

Unit 1 is separate from Three Mile Island Unit 2, the reactor involved in the 1979 accident.

The commercial significance is difficult to miss: a technology company's demand for electricity is helping underpin the attempted return of a nuclear generating unit that had previously closed for economic reasons.

Nuclear is not the only answer

The renewed interest in nuclear power does not mean every future data centre will sit beside a reactor.

The IEA expects a mixture of electricity sources to meet growing demand.

In its base case, renewables supply roughly half of the increase in electricity generation associated with data-centre demand through 2035.

Natural gas also plays a major role, particularly in the United States, while nuclear becomes increasingly important later in the decade and beyond.

This creates a difficult reality for companies making aggressive climate commitments.

Electricity demand can grow faster than zero-carbon generation is built.

A company can sign renewable contracts, support nuclear power and still operate in an electricity system where fossil generation is part of the marginal supply at particular hours.

The climate argument is more complicated than electricity use alone

A terawatt-hour consumed in a low-carbon electricity system does not have the same emissions consequences as a terawatt-hour supplied predominantly by fossil fuels.

That makes location and timing important.

Companies increasingly talk about matching electricity consumption with clean generation not simply over an entire year but more closely by hour and region.

Even then, building the physical infrastructure remains a separate issue from accounting for electricity purchases.

Transmission capacity has to exist. Generators have to connect. Grid operators have to maintain reliability during periods when wind or solar output falls.

Data centres are competing with everything else that wants electrification

AI is arriving during a period when electricity demand is expected to grow for reasons that have little to do with chatbots.

Electric vehicles need charging.

Heating is moving towards electric heat pumps in many regions.

Manufacturers are considering electrification.

New homes and factories require grid connections.

The IEA projects data centres to account for less than 10 per cent of global electricity-demand growth between 2024 and 2030.

Globally, they are not swallowing the entire power system.

Locally, however, their concentration makes the challenge much sharper.

A large industrial load arriving in one specific county does not care that its share of global electricity use is small.

Who pays for the new wires?

Electricity demand from very large new customers creates another argument that is less technologically glamorous but politically important.

Who pays for the infrastructure needed to connect them?

Utilities may need to build substations, transmission upgrades or new generation in anticipation of a proposed data-centre campus.

But forecasts can be wrong.

If enormous infrastructure is built for a customer whose project is delayed, downsized or cancelled, regulators have to decide who carries the financial risk.

Berkeley Lab's August 2026 work on electricity-rate design for large loads specifically highlights concerns around insufficient supply and underused investments affecting other electricity customers.

That pushes AI infrastructure into the ordinary world of utility regulation, tariffs and cost allocation.

A data centre can be valuable without employing thousands forever

Communities hosting large facilities face a similar calculation.

Data centres can represent billions of dollars in investment. Construction can require substantial workforces, and facilities can generate significant tax revenue.

But their long-term employment profile is different from a conventional factory.

Extremely valuable computing equipment can operate with a comparatively modest permanent workforce.

That does not make the investment worthless.

It does mean governments should evaluate promised benefits against the cost of land, electricity infrastructure, tax incentives and local resource use instead of relying on one headline number.

Efficiency is improving at the same time demand is exploding

The energy story is not simply that AI models become larger and electricity demand rises forever at the same rate.

Hardware becomes more efficient.

Software improves.

Models can be specialised, compressed or routed so that relatively simple tasks do not always require the largest available system.

Google said in its 2025 environmental reporting that its Ironwood TPU was nearly 30 times more power-efficient than its first Cloud TPU from 2018.

Improvements like that matter enormously.

They do not guarantee falling total electricity use.

Making an AI request cheaper can make it economical to use AI in far more places.

A model that consumes half as much electricity per task can still drive higher overall demand if the number of tasks increases fivefold.

This is why forecasts have such large ranges

Nobody knows exactly what AI demand will look like four years from now.

It depends on how many systems are deployed, how efficiently they operate, what kind of hardware companies purchase, how heavily those machines are used and whether consumers continue adopting AI services at the current pace.

Berkeley Lab's 2026 scenarios illustrate how sensitive the outcome is to apparently technical assumptions.

Its estimates change when researchers vary the number of specialised graphics processors shipped, the operating lifetime of AI chips and how intensively servers are used.

A forecast is therefore not a promise about the future.

It is a map of what happens if particular assumptions prove approximately correct.

The strange part is how invisible all of this remains

The user sees a text box.

The answer appears in seconds.

It feels more like using a calculator than operating industrial machinery.

That interface hides an extraordinary supply chain.

Semiconductor fabrication plants manufacture the processors. Factories produce servers and cooling equipment. Fibre networks connect facilities. Utilities reinforce substations. Power stations produce electricity. Engineers operate buildings that may consume the output of a small power plant.

The simplicity of the product conceals the complexity of the system.

AI may help the energy system too

There is an important counterpoint.

Artificial intelligence is not only a new electricity consumer.

The same technology can be used to improve weather forecasting, optimise transmission networks, predict equipment failures, manage industrial processes and accelerate research into batteries and other energy technologies.

The IEA has identified substantial potential for AI to improve the efficiency of parts of the energy sector.

That does not cancel the electricity consumed by data centres.

It means the relationship runs in both directions.

The next AI bottleneck may not be the model

The technology industry spent the first phase of the generative-AI boom talking about model size, chip shortages and access to specialised processors.

Those constraints still matter.

Electricity is now joining them.

A company may have the money for servers and still lack a place where enough power can be delivered quickly.

A region may want the investment and still lack transmission capacity.

A utility may be capable of supplying the energy eventually while being unable to connect the load on the technology company's preferred schedule.

The IEA puts the contrast neatly in structural terms: a data centre can be operational within two or three years, while major energy infrastructure frequently requires much longer planning and construction periods.

The cloud is becoming an industrial sector

None of this means the AI boom has to stop.

It means the public description of it needs to catch up with its physical reality.

AI companies are software companies, but increasingly they are also enormous infrastructure customers.

Their future depends on processors, but also on transformers.

It depends on algorithms, but also on substations.

It depends on model architecture, but also on cooling equipment, transmission lines, water availability and construction schedules.

The public sees a chatbot.

Behind it sits an industrial system.

And the faster artificial intelligence grows, the harder that system will be to keep invisible.

International Correspondent

Edward Lee

Edward Lee covers international affairs, technology infrastructure, energy, economics and history for Crazy News. His reporting focuses on explaining how large systems work, where their incentives come from and how apparently technical decisions eventually affect countries, communities and ordinary people.

Reporter profile