AI is usually presented as something weightless.

A prompt goes into a box. Words appear. Somewhere, apparently, there is a cloud.

The cloud has buildings.

Inside those buildings are racks of computers drawing enormous amounts of electricity, producing heat and requiring cooling, networking, backup systems and actual land. The rapid growth of AI has made that physical layer much harder to ignore.

The next big limit on artificial intelligence may not be whether researchers can design a better model.

It may be whether someone can get enough power to the data centre that runs it.

The electricity request is not small anymore

Data centres already supported search, streaming, storage and business software before generative AI became popular.

AI adds particularly demanding workloads. Training large models consumes serious computing power, and serving those models to millions of users keeps hardware busy after training is finished.

The International Energy Agency expects global data-centre electricity use to rise sharply through the rest of the decade, with AI as a major driver.

That does not mean the planet’s electricity is about to disappear into chatbots. It does mean local power systems can feel the pressure much more strongly than global percentages suggest.

A new data centre does not connect to the world’s average grid. It connects somewhere specific.

That somewhere needs capacity.

Location is becoming an energy decision

Data centres have always cared about connectivity, land prices, taxes and the availability of skilled workers.

Now access to electricity is becoming an even louder part of the conversation.

A region may have cheap land but a grid that cannot support another large load quickly. Another may have abundant renewable generation but limited transmission. A utility might need years to build substations and lines before a planned campus can run at full capacity.

This is why technology companies are signing long-term energy deals and talking about nuclear power, renewables, storage and new generation in the same breath as model development.

The AI race has quietly become an infrastructure race.

Software people are discovering that electrons have planning permission.

Efficiency does not automatically reduce demand

New chips keep getting more efficient. Models are being optimised. Cooling systems improve. Data-centre operators squeeze more computing from each unit of electricity.

All of that is necessary.

It may not reduce total consumption.

When computing becomes cheaper and more efficient, we often use more of it. AI features move into email, search, coding tools, cameras, customer service and ordinary office software. A task that once called a model occasionally may run automatically for every user.

So efficiency can lower the cost per task while the number of tasks explodes.

This is a familiar technology story. Better fuel economy does not matter much if everyone starts driving twice as far.

Cooling is where the invisible heat goes

Every computation ends with heat.

A dense rack of AI hardware produces enough of it that cooling becomes a serious engineering problem. Air cooling still plays a role, but more powerful systems increasingly use liquid-based approaches to move heat away efficiently.

This creates design choices around water, pumps, chillers and local climate.

Data-centre cooling varies by design and climate. Some facilities use water directly; others rely more on air or closed-loop systems.

The useful question is how a particular facility is designed, what it draws locally and what alternatives were considered.

The electricity mix matters as much as the amount

Two identical data centres can have very different environmental effects depending on how the electricity is generated.

A facility supplied mainly by low-carbon power is not the same as one depending heavily on fossil generation. Timing matters too. A grid may have abundant clean electricity at one hour and struggle at another.

This is why companies are interested in matching workloads with power availability and building near reliable low-carbon generation. Some training jobs can also be scheduled rather than run immediately.

Real-time services are less patient, of course. Nobody wants a chatbot to say, “Excellent question, the wind is low, please return tomorrow.”

Smaller AI may become an energy feature

The industry’s obsession with the largest possible model is already weakening for practical reasons.

If a small model can perform a narrow task well, running the huge model may be wasteful. Some work can happen on phones and laptops. Some can be routed to specialised models rather than the most capable system available.

Cost pushes companies in this direction even when environmental concern does not.

Every unnecessary computation appears somewhere on an infrastructure bill.

I expect AI products to get better at deciding how much intelligence a task actually needs. A grammar correction should not require the same machinery as a difficult research problem.

That sounds obvious. Software has a long history of using expensive resources because they are easy to call.

Communities will have opinions

A data centre becomes a neighbour. It can bring construction, jobs and tax revenue, but also compete for power, land and water.

Communities are going to ask harder questions as projects become larger.

How much electricity will the site require? Who pays for grid upgrades? What happens during shortages? How much water does the cooling system use? How many permanent jobs remain after construction?

Those are reasonable questions.

Technology companies spent years describing digital services as if geography barely mattered. Data centres make geography impossible to avoid.

Your AI subscription hides the power bill

Ordinary users rarely see any of this directly.

You pay for a subscription or use a free AI feature. The provider handles the servers. Electricity is bundled into the service.

But the cost eventually appears somewhere: subscription prices, usage limits, product design or the decision to use a cheaper model for routine tasks.

Developers already think about token costs and inference costs. Energy is part of the same underlying reality.

The most sustainable AI feature may also be the one that is cheapest to operate because it does less unnecessary work.

That alignment is useful.

The cloud was never actually in the sky

AI will keep improving. Data centres will become more efficient. Power generation and grids will expand. Some processing will move closer to users.

None of that removes the physical layer.

Every impressive answer comes from chips that need electricity and cooling. Every cloud feature eventually reaches a building with cables entering it.

I don’t think this is a reason to stop using AI.

It is a reason to stop pretending software has no weight.

The next generation of AI may be decided partly by model architecture and partly by a much older question:

Where do we plug it in?