The next constraint on AI may not be intelligence. It may be infrastructure.
By The Wider Lens Desk
When people talk about the AI race, the conversation usually starts with the same things: better models, faster chips, bigger data centres and companies like OpenAI, Google, Anthropic and Nvidia.
But there is another part of the story that is much easier to overlook.
All of that computing needs somewhere to run. And that somewhere needs electricity.
A new report from the International Energy Agency suggests that the physical infrastructure behind AI is becoming a constraint in its own right. The question is no longer just how powerful AI models can become. It is also how quickly the world can build the electricity systems, data centres, cooling equipment and grid connections needed to run them.
The Hidden Physical Cost of the AI Boom
The money flowing into AI infrastructure is already enormous.
The IEA's Key Questions on Energy and AI, published in April 2026, says capital expenditure by the largest technology companies exceeded $400 billion in 2025 and is expected to rise another 75% in 2026. Spending by just five technology companies is now larger than global investment in oil and natural gas production.
And much of that money is going into physical infrastructure.
The IEA's satellite-based tracking shows that AI-focused "factories" — specialised data centres built for advanced AI workloads — have more than tripled in capacity over the past 18 months.
That is an important shift.
The AI story is often told as if it lives entirely inside software. In reality, every new model that reaches millions of users has to sit somewhere. It needs servers, buildings, cooling systems, electrical equipment, land and a reliable connection to the power system.
The more AI grows, the more visible that physical layer becomes.
The Electricity Number Behind the AI Race
The numbers are already moving quickly.
Global electricity demand from data centres increased 17% in 2025. Electricity consumption from AI-focused data centres grew even faster, rising 50% over the same year.
The IEA expects total data-centre electricity consumption worldwide to roughly double from 485 terawatt-hours in 2025 to around 950 terawatt-hours by 2030. By then, data centres could account for about 3% of global electricity demand.
AI-focused facilities are expected to grow much faster than data centres overall, with their electricity consumption projected to triple during the same period.
There is, however, an important detail that can get lost in the headline numbers.
It would be misleading to assume that every AI query is becoming dramatically more power-hungry. In fact, the IEA says energy use per individual AI task has fallen sharply. A simple text query can now use less electricity than running a television for the same amount of time.
The bigger issue is scale.
More people are using AI, while newer applications are also becoming much more demanding. Video generation, complex reasoning and agentic systems can require hundreds or even thousands of times more energy per query than simple text generation.
So the problem isn't necessarily that one AI request is becoming impossibly expensive in energy terms.
It is that billions of requests — plus increasingly demanding workloads — are being added to a system that still has to be physically built.
Why Location Matters More Than the Global Total
There is a difference between having enough electricity somewhere in the world and having enough reliable electricity at the exact place where a huge data centre is being built.
That distinction matters.
A new AI facility can be announced in months. Building the transmission infrastructure, substations, transformers, generation capacity and grid connections needed to support it can take much longer.
The IEA describes growing pressure across the AI supply chain, including electricity, grid access, manufacturing capacity, chips and capital. It also says bottlenecks in energy supply chains and advanced chip manufacturing have tightened since its previous assessment.
In the United States, some developers are responding to slow grid connections by looking at onsite natural-gas generation. The IEA estimates that roughly 15 to 27 gigawatts of onsite gas generation could be serving data centres by 2030, mostly in the United States.
But onsite generation isn't a simple solution either.
Because AI data-centre loads can change rapidly, providing reliable power can require substantially more generation capacity than the facility's average demand. The IEA estimates an overbuild of around 30% to 70% relative to demand in some onsite gas configurations.
And then there is the hardware itself.
An advanced AI server rack is roughly the size of a large refrigerator. By 2027, the IEA estimates that one rack could have peak power demand equivalent to about 65 households.
The change has been remarkably fast. AI server power density increased elevenfold between 2020 and 2025, with another fourfold increase expected by 2027.
More computing power in the same physical space also means more heat.
The IEA compares the heat load of an advanced rack with roughly 30 natural-gas boilers. The equipment can also weigh more than a pickup truck.
At that point, AI stops looking like a purely digital technology.
It starts looking like heavy infrastructure.
A Race for More Than Compute
This changes the way the AI race should be understood.
The obvious competition is still between models, chips and companies. But underneath that competition is another race — for electricity, land, cooling, grid access, construction capacity and the equipment needed to connect new facilities to the power system.
There is also a financial dimension.
The IEA says data-centre investment has become too large to rely on company balance sheets alone. Capital markets will play an increasingly important role in funding the expansion.
That means AI infrastructure is not only affected by technological progress. It is also exposed to interest rates, investor expectations, financing conditions and questions about whether the enormous investments being made today will generate sufficient returns.
A powerful AI model can be demonstrated in a research environment relatively quickly.
A power plant, transmission line or major data centre cannot.
That difference may become increasingly important as the industry scales.
The Governance Question Nobody's Asking Yet
Most discussions about AI regulation focus on what the software should be allowed to do.
What content can an AI system generate?
What decisions can it make?
What safeguards should apply?
Those questions matter.
But the infrastructure behind AI creates another set of questions that are much more physical.
Where should large AI data centres be built?
What happens when a new data centre needs electricity capacity that is also needed by homes and businesses?
Who should pay for the transmission lines, substations and other grid upgrades required to connect it?
And what happens to the surrounding community when a project changes local land use, energy demand or environmental conditions?
The IEA's policy recommendations include better management of data-centre project pipelines, improvements to permitting and grid-connection processes, and mechanisms for allocating the cost of grid upgrades more fairly.
These issues are already appearing in communities where data-centre projects are being proposed, with affordability and environmental impacts among the concerns being raised.
That makes the infrastructure side of AI more than an engineering problem.
It is also becoming an economic and public-policy issue.
What This Means for the Next Era of AI
None of this means AI development is about to stop.
The IEA itself leaves room for electricity demand to exceed its central projections if investment in energy equipment and chip manufacturing removes some of today's bottlenecks and if energy-intensive AI applications grow faster than expected.
But the direction is becoming clearer.
The next phase of AI will not be measured only by benchmark scores, model parameters or the number of users a service can attract.
It will also depend on much less glamorous things: how quickly a transformer can be delivered, how long a grid connection takes, how much land is available, where new generation can be built, and whether the financing exists to pay for it.
That is the part of the AI revolution most users never see.
We experience AI as a box on a screen. Behind that box is a very physical system of servers, buildings, cooling equipment, power lines and electricity.
The smarter AI becomes, the more important that physical system becomes.
The AI race, in other words, is no longer just about building better machines.
It is also about building the world those machines need in order to work.
Source: International Energy Agency (IEA), Key Questions on Energy and AI (2026), IEA, Paris. Licensed under CC BY 4.0.
Do you think physical infrastructure — rather than model capability — could become one of the biggest constraints on AI's growth? Share your view in the comments.
Comments
Post a Comment