Where the Cloud Touches the Ground

In July, Britain's energy regulator found itself dealing with an unusual problem. Too many data centres wanted electricity.

Applications for large demand connections to the electricity network had risen from 41 gigawatts to 125 gigawatts in less than a year. Ofgem said at least 80GW of that queue came from proposed data-centre projects. It was sufficiently concerned that some developers might be reserving scarce grid capacity without ultimately building anything that it proposed a new commitment fee, refundable when a project actually connects and potentially forfeited if it drops out.

There is something revealing about an artificial-intelligence boom eventually becoming a queue for electrical substations.

For most of us, computing has become almost aggressively invisible. We open an application, upload a photograph, ask a question or summon a paragraph of text. The physical work involved rarely enters the experience. Even the language we use encourages the abstraction. Our photographs are in "the cloud". Companies move their systems "to the cloud". AI models appear inside websites as though intelligence has simply become another utility available through a browser.

Cloud computing was never actually in the cloud, of course. It was always someone else's building.

Artificial intelligence is making that increasingly difficult to ignore.

The International Energy Agency estimates that data centres consumed around 485 terawatt-hours of electricity globally in 2025. Its central projection has that figure reaching roughly 950TWh by 2030. AI-focused facilities are expected to grow considerably faster than data centres overall. These numbers remain uncertain — the IEA publishes substantially different scenarios depending on efficiency, adoption and infrastructure constraints but the direction is clear enough: more digital activity increasingly requires more physical capacity somewhere.

Britain currently offers a particularly useful glimpse of what that means.

The government has created AI Growth Zones intended to accelerate construction of the computing infrastructure it believes the country will need. The criteria are telling. A potential site is not assessed only for telecommunications links or proximity to universities. It must demonstrate access to power, land, planning permission and water. Applicants are expected to show that at least 500MW of electricity can be available by 2030, alongside sufficient water and discharge capacity to support the infrastructure.

Five hundred megawatts is not an abstract software requirement. It is the language of industrial infrastructure.

The British government is now explicitly trying to shorten the time between asking for power and obtaining it. Its AI Growth Zone programme proposes faster grid connections, planning reforms and financial incentives to locate large computing facilities where the electricity system can accommodate them more easily. The government says these measures could reduce "time to power" by as much as five years for qualifying projects.

This is where the cloud touches the ground.

The interesting story is not simply that AI uses electricity. Every modern industry uses electricity. Nor is it necessarily that data centres are about to consume all available power. They are not: even the IEA's central projection puts data centres at around 3 per cent of global electricity consumption in 2030.

The deeper change is that computing is becoming large enough, concentrated enough and strategically important enough to influence decisions about physical infrastructure.

For decades, one of software's greatest economic advantages was that it could scale without every additional customer requiring another shop, railway carriage or factory line. A digital product could acquire millions of users while appearing almost detached from geography.

AI weakens that illusion.

Additional demand still travels through software, but eventually it lands somewhere: in processors, cooling systems, backup equipment, transformers and transmission networks. When enough computing is concentrated in one place, location begins to matter again.

This creates a strange collision between two industries that operate on very different clocks.

Technology companies can release a model, attract millions of users and dramatically change computing requirements within a few years. Electricity infrastructure moves differently. Transmission lines can take years to plan and construct. New generation requires permits, capital and supply chains. Transformers and grid equipment can themselves have long procurement periods.

The IEA makes precisely this distinction: a data centre can sometimes be built in two or three years, while the wider energy infrastructure needed to serve it often requires substantially longer planning and construction periods.

Software can scale at internet speed. Power grids cannot.

That mismatch helps explain Britain's extraordinary connection queue. Developers have an incentive to secure electrical capacity early because without electricity their future project is worthless. But when many developers behave rationally in the same way, the collective result can become irrational: grid capacity is reserved for projects that may never exist, while viable projects wait behind them.

Ofgem's proposed commitment fee is therefore not really about punishing data centres. It is an attempt to put a price on certainty.

The regulator wants developers to demonstrate that their plans are real enough to justify occupying a scarce place in the system.

This is a useful reminder that infrastructure scarcity is not always caused by physical shortage alone. Sometimes it is created by expectations.

If businesses believe AI demand will be enormous, they apply for electricity connections. Those applications make future demand appear enormous. Grid planners then have to decide how much infrastructure to build around a future that is partly composed of companies anticipating one another.

There is no easy answer. Underbuild, and economically valuable projects may spend years waiting for electricity. Overbuild, and consumers may ultimately help pay for assets that were constructed for demand that never arrived.

The difficulty is not unique to AI. Railways, airports, power stations and telecommunications networks have always required societies to build ahead of demand. Infrastructure is almost definitionally a wager on future use.

What makes AI interesting is the speed at which expectations can change.

The physical requirements also mean that geography starts reasserting itself.

The British government's Growth Zone strategy explicitly encourages data centres to locate in places where additional electricity demand may actually help the grid. Scotland, for example, can sometimes generate more wind power than the transmission system can economically move south. Under current proposals, qualifying data centres in suitable locations could receive reduced electricity-system charges because consuming power closer to where it is generated may reduce wider network costs.

That reverses the usual way we imagine digital geography.

We tend to assume infrastructure follows technology companies. Here, computing may increasingly follow infrastructure.

A future data centre may be built somewhere not because executives particularly want to locate there, but because that is where electricity is plentiful, grid congestion is manageable, land is available and planning permission can be obtained.

Water introduces another layer. Different cooling systems have very different water requirements, so it would be misleading to treat every data centre as equally water intensive. But the fact that Britain's Growth Zone application process explicitly asks authorities to demonstrate sufficient water supply and discharge capacity is itself significant.

A conversation ostensibly about artificial intelligence has reached reservoirs, pipes and wastewater.

The European Union is moving in a similar direction. Large data centres are already subject to energy-performance reporting requirements, and in September the European Commission proposed a common rating scheme intended to make energy and water performance easier to compare. The EU also says it wants to triple its data-centre capacity by 2035.

This combination build much more infrastructure while simultaneously measuring its physical impact more closely is likely to become increasingly common.

That matters for people who have no particular interest in artificial intelligence.

Physical infrastructure creates local consequences in a way software often does not. A new application does not usually require planning permission. A large data centre does. It occupies land. It connects to roads and electricity infrastructure. It can create construction work and tax revenue. It can also prompt arguments about energy, water, visual impact and what else scarce infrastructure might have supported.

Those are political and economic choices rather than engineering details.

Imagine a region with limited grid capacity. A housing development needs electricity. An advanced manufacturer wants to expand. Transport is becoming electrified. Residents are installing heat pumps. A data-centre developer wants hundreds of megawatts.

The useful question is no longer simply whether the data centre is "good" or "bad".

It is what kind of infrastructure should be built, who should pay for it, how quickly it can expand and which uses generate enough value to justify their demands on the system.

Those debates can become distorted because the benefits and costs are often measured differently.

A government may value a data centre as strategic computing capacity. A developer may value its commercial returns. A local authority may care about investment and employment. An electricity-system operator cares about when and where demand occurs. A resident may care about whether local infrastructure improves or becomes more constrained.

None of these perspectives is necessarily wrong. They are looking at different parts of the same system.

There is another complication. Data centres can sometimes help electricity networks rather than simply burden them.

A sufficiently flexible facility might shift some computing to periods when electricity is plentiful, use local generation, participate in demand-response programmes or locate near renewable supply that would otherwise be curtailed. The European Commission explicitly notes that flexible data centres could improve grid stability and help integrate renewable energy.

This is why simple claims about AI "using too much electricity" are not particularly useful.

The important variables are where the demand appears, when it occurs, what generation serves it, how flexible it is, what infrastructure it requires and what economic activity it enables in return.

Britain's own recent energy research offers another reason for caution. A government-commissioned study of digital services found that digital alternatives can sometimes reduce overall energy consumption compared with the physical activity they replace. AI-assisted work, for example, may use additional computing while reducing the amount of energy consumed per completed task if productivity rises enough.

Energy use is therefore not a simple tally of server consumption.

A technology can consume more electricity directly while making another process use less energy elsewhere. Equally, efficiency improvements can lower the cost of computing and encourage people to use much more of it. The system moves.

What appears more certain is that electricity itself is becoming strategically important again.

For many advanced economies, demand had been relatively stagnant for years. That is changing. The IEA expects electricity consumption in advanced economies to rise through 2030, driven not only by data centres but also electric vehicles, heat pumps, air conditioning and industrial activity.

AI is therefore arriving at the grid at the same time as transport, heating and parts of industry are being asked to electrify.

That is the larger operating question.

Digital expansion is no longer happening in an isolated digital economy. It is becoming one participant in a much broader competition to reorganise the physical economy around electricity.

Governments increasingly understand computing infrastructure in those terms. Britain designated data centres as Critical National Infrastructure in 2024, putting them in the same broad category of national importance as systems such as energy and water. The language is revealing. Data centres have moved from being anonymous industrial buildings to infrastructure that governments believe requires explicit resilience, planning and security.

The cloud has become part of the country's machinery.

That does not mean every proposed AI campus should be built, or that every concern about its resource use is justified. Some of today's investment expectations may prove excessive. Computing hardware may become more efficient. AI demand may grow differently from current forecasts. Ofgem itself is worried that some of the projects occupying Britain's grid queue may never proceed.

But that uncertainty is part of the story rather than a reason to ignore it.

Electricity networks have to make decisions before anyone knows exactly how important today's AI boom will ultimately become. Land has to be allocated. Connections have to be planned. Generation has to be financed. Developers have to decide where to build. Governments have to decide how much support to provide.

The physical world cannot wait for the technological argument to be settled.

Which brings us back to that strange 125GW connection queue.

At first glance it looks like an energy-policy problem: too many proposed data centres asking for too much electricity.

Look at it differently and it becomes a picture of something larger.

For years we have talked about artificial intelligence as though its primary constraints were algorithms, chips, data and talent. Increasingly, another set of constraints is entering the conversation: substations, transmission lines, cooling systems, reservoirs, planning departments and pieces of land.

They are mundane compared with a new AI model.

They may also determine how much of the AI economy can actually be built.

The cloud was never weightless. We are simply reaching the point where its weight is becoming difficult not to see.

References

Ofgem, Ofgem acts to free up grid capacity by tackling speculative data centre projects, 29 July 2026.

International Energy Agency, Key Questions on Energy and AI, 16 April 2026.

International Energy Agency, Electricity 2026.

UK Department for Science, Innovation and Technology, Delivering AI Growth Zones, November 2025.

UK Government, AI Growth Zones: open for applications, June 2025.

UK Department for Energy Security and Net Zero, Impact of growth of data centres on energy consumption, updated July 2026.

European Commission, Energy performance of data centres, including proposed EU rating scheme, September 2026.

UK Department for Science, Innovation and Technology, Data centres to be given massive boost and protections from cyber criminals and IT blackouts, September 2024.