The next AI bottleneck isn't the GPU.
It's the plug.
AI may live in the cloud, but its constraints are becoming remarkably physical.
Land.
Power.
Cooling.
Concrete.
Grid connections.
Permits.
People.
You cannot API your way around any of them.
For the past few years, the AI infrastructure conversation has been dominated by compute.
Who has the GPUs?
Who has access to the newest accelerators?
Who can build the largest clusters?
Who can train the biggest models?
Those questions still matter.
But another constraint is rapidly becoming just as important.
Can you actually supply enough electricity to switch all of it on?
The International Energy Agency expects global data-centre electricity consumption to more than double to around 945 TWh by 2030, with AI the biggest driver of that growth. In the United States, data centres are projected to account for almost half of electricity-demand growth between now and 2030.
That changes the AI conversation considerably.
Because software scales differently from electricity.
You can deploy another software instance in seconds.
You cannot deploy another gigawatt in seconds.
The Cloud Hid the Physical World
One of cloud computing's greatest achievements was abstraction.
For years, CIOs didn't really need to think about where computing happened.
We selected a cloud region.
Provisioned capacity.
Negotiated contracts.
Built applications.
Somewhere behind that abstraction were enormous data centres, substations, generators, cooling systems, fibre routes and power contracts.
They weren't our problem.
Or at least they didn't feel like our problem.
AI is beginning to break that abstraction.
The sheer density of modern compute means electricity availability is becoming an architectural constraint.
And that creates an uncomfortable reality:
For years, the cloud allowed CIOs to forget where computing happened. AI is forcing us to remember.
You Can't Scale Electricity Like Software
The technology industry has become accustomed to exponential curves.
More compute.
More storage.
More bandwidth.
More users.
Lower unit costs.
Energy infrastructure doesn't behave that way.
Power stations need to be built.
Transmission capacity needs to be created.
Grid connections need approval.
Transformers need manufacturing.
Land needs acquiring.
Data centres need cooling.
People need hiring.
Planning permission needs granting.
And every one of those things operates on physical-world timescales.
That creates a strange mismatch.
AI development moves in months.
Energy infrastructure can move in years.
You can have the capital.
You can have the GPUs.
You can have the land.
You can even have the customer.
And still be waiting for the electricity.
Europe Has a Particular Problem
This is where the conversation becomes uncomfortable for Europe.
Europe wants to be competitive in AI.
It wants European AI champions.
It wants sovereign technology.
It wants more data-centre capacity.
The European Commission itself says the EU aims to triple its data-centre capacity by 2035.
But AI doesn't run on ambition.
It runs on electricity.
And electricity economics matter.
Germany remains one of Europe's more expensive markets for non-household electricity. Eurostat recorded German non-household prices of €22.64 per 100 kWh in the second half of 2025, compared with an EU average of €18.37.
Britain has struggled with the same competitiveness problem. UK government statistics have repeatedly shown unusually high industrial electricity prices by international standards; official comparisons for the second half of 2024 placed UK industrial electricity prices above the EU14 countries compared. The government has subsequently introduced additional support specifically to reduce that disadvantage.
That matters because AI infrastructure is extraordinarily sensitive to power economics.
A small difference in electricity cost becomes a very large difference when multiplied across hundreds of megawatts running continuously.
And then comes regulation.
The EU already requires qualifying data centres to report energy-performance and water-footprint information under the Energy Efficiency Directive. It is now preparing further measures including an EU-wide data-centre rating scheme and minimum performance standards.
Individually, many of these objectives are understandable.
Energy efficiency matters.
Water consumption matters.
Environmental impact matters.
Transparency matters.
But competitiveness is cumulative.
Every regulation may be defensible in isolation while the combined environment becomes increasingly difficult to invest in.
That's the part policymakers too often miss.
Regulation Has an Opportunity Cost
Europe has become extremely good at asking:
"How should this technology be regulated?"
It needs to become equally good at asking:
"What does this technology require in order to exist here?"
Those are not the same question.
AI requires compute.
Compute requires data centres.
Data centres require electricity.
Electricity requires generation, transmission and grid capacity.
All of it requires capital, land, infrastructure and permission.
If each layer becomes slower, more expensive or more administratively difficult, investment doesn't necessarily disappear.
It moves.
That's particularly dangerous with AI infrastructure because location is becoming increasingly flexible.
If a workload doesn't need to be physically close to the customer, capital can increasingly follow available power.
The winning geography may not have the best tax incentive or the largest technology cluster.
It may simply be the place where someone can say:
"Yes, we can give you 500 megawatts."
Data-Centre Architecture Is About to Change
Historically, data centres were built around connectivity and population.
Put compute close to users.
Put it near fibre.
Put it near established technology hubs.
Power was important, but largely assumed.
That assumption is reversing.
For the largest AI workloads, we may increasingly put compute where the energy exists and connect everything else to it.
That could mean more infrastructure built away from major cities.
It could mean colocating generation and compute.
It could mean greater use of nuclear, gas, solar, wind and long-duration storage depending on geography.
It could mean hyperscalers becoming increasingly involved in energy generation rather than simply purchasing electricity from a grid.
There won't be one answer.
And that's important.
Solar isn't the answer.
Nuclear isn't the answer.
Gas isn't the answer.
Wind isn't the answer.
The answer is likely to be an energy mix designed around availability, economics, resilience and geography.
This is less ideologically satisfying than declaring one technology the winner.
It's also considerably more pragmatic.
The AI Infrastructure Reality Check
For CIOs and boards approving significant AI investment, I think five questions increasingly need to sit alongside the technology strategy.
1. Compute
What computing capacity do we actually need?
Not what can we buy.
What creates measurable business value?
The easiest infrastructure problem to solve is the capacity you never needed in the first place.
2. Power
Where does the electricity come from?
Not simply today.
At the scale you expect to require three or five years from now.
Availability, reliability and price all matter.
3. Location
Does the workload need to be where we traditionally put computing?
Latency, sovereignty, data residency and connectivity matter.
But so does energy availability.
The optimal location for AI infrastructure may be very different from the optimal location for yesterday's enterprise data centre.
4. Regulation
What could prevent this capacity from being built or operated?
Planning.
Grid connections.
Environmental requirements.
Data sovereignty.
AI regulation.
Reporting obligations.
Regulation isn't simply a compliance question anymore.
It's part of infrastructure architecture.
5. Economics
Finally:
Does the business case still work when all of those constraints are included?
GPU cost is only one line.
Energy.
Cooling.
Networking.
Land.
Financing.
Regulatory compliance.
Resilience.
People.
All ultimately determine the economics of AI.
And that means the cheapest model may not produce the cheapest outcome.
This Becomes a CIO Problem
Most enterprise CIOs aren't about to start building power stations.
Nor should they.
But we do need to understand that the economics underneath our technology platforms are changing.
Cloud providers will feel these constraints.
SaaS providers will feel them.
AI vendors will feel them.
Eventually their customers will feel them too.
Through pricing.
Through capacity constraints.
Through regional availability.
Through architecture choices.
And potentially through decisions about where new digital infrastructure gets built.
Power therefore stops being somebody else's infrastructure problem.
It becomes part of technology strategy.
The Pragmatic View
We've spent much of the AI era discussing intelligence.
Models.
Agents.
Automation.
Productivity.
But underneath all of it sits something considerably less glamorous.
Electricity.
The next phase of AI won't be determined purely by who develops the smartest model or purchases the most GPUs.
It will also be shaped by which countries, companies and regions can build the physical infrastructure required to operate them.
That should be particularly sobering for Europe.
You cannot simultaneously demand AI leadership, affordable energy, rapid decarbonisation, stronger sovereignty, more regulation, greater infrastructure scrutiny and slower physical development without eventually confronting the trade-offs between them.
Some compromises will have to be made.
Because ultimately AI doesn't care about policy objectives.
It doesn't care about political ambition.
And it doesn't care how sophisticated the regulation surrounding it becomes.
It needs electricity.
And increasingly, whoever can provide enough of it—reliably, affordably and quickly—will have an advantage that no algorithm can easily overcome.
If electricity became the limiting factor in your AI strategy tomorrow, would you even know where the dependency sits?
