The physical limits behind the AI economy
A proposed data centre can look complete in a presentation. The site is identified, the buildings are drawn and the computing equipment has a place to go. Then the electricity connection date appears. It can turn an apparently finished plan into a long and rather expensive waiting exercise.
AI is discussed as though it mostly happens in software. Its expansion also depends on land, substations, cables, cooling systems and people who can build and operate them. These dependencies create a different kind of opportunity from developing another application. They also require a different kind of patience.
The International Energy Agency's 2025 Energy and AI report put grid constraints near the centre of the discussion. It estimated that around a fifth of planned data centre projects could face delays if relevant risks were left unresolved, and described multi-year lead times for transmission infrastructure. These are scenario-based assessments, rather than a timetable for any particular site. They are enough to explain why access to usable power deserves careful attention.
The word usable is important. Land near a transmission line is not automatically a powered site. A connection application is not the same as an operating connection. A proposed generator still has to be financed, permitted, built and integrated into the network. Each stage involves a different set of obligations, and each can change what the completed asset is worth.
There is useful evidence on the supply side in Berkeley Lab's 2026 Queued Up report. It found that, in the regions with available data, generation projects reaching commercial operation in 2025 had spent a median of more than five years between their connection request and operation. This is a study of generation queues, not a measurement of data centre connection times. It nonetheless shows why announced generating capacity cannot simply be treated as available electricity.
Control of a genuine constraint can have economic value. A site with suitable permissions, dependable power arrangements and a credible route to operation offers something that a drawing cannot. Yet the attraction should be described carefully. The durable capability may lie in bringing all those pieces together, rather than merely owning a piece of land and waiting for the market to become impatient.
This is an area where execution is unusually concrete. Equipment must arrive. Engineering assumptions must survive detailed examination. Contractors need to coordinate. A delayed component can hold up an otherwise impressive collection of completed work. Ownership is useful when it provides enough control to solve those problems, rather than simply exposure to them.
Scarcity also attracts a response. Networks are expanded, connection rules change, customers consider other locations and computing becomes more efficient. In its June 2026 update on demand connections reform, Ofgem set out the objective of helping viable projects connect in good time while allowing strategic projects to be prioritised. It is a reminder that access arrangements develop through public policy as well as private negotiation.
That makes a simple scarcity story incomplete. A constraint that looks durable today may ease before a project reaches operation. Demand may develop elsewhere. Some workloads may become more flexible about when or where they run. The sensible question is how the asset remains useful across a range of outcomes, including a slower or less power-intensive expansion of AI.
There is a wider responsibility too. Electricity networks serve households, factories and public services. New demand can support investment in better infrastructure, but the allocation of costs matters. A project whose economics depend on other customers quietly absorbing its network costs is carrying a political risk, whether or not that risk appears in the presentation. Local concerns about water, noise and land use also deserve practical answers.
The constructive ambition is to remove constraints in a way that leaves the surrounding system stronger. That might involve bringing forward capacity, choosing a better location or designing operations that can accommodate the network's needs. There will be trade-offs, and they should be visible early.
The physical side of AI is worth understanding because it turns a broad technological forecast into specific questions about delivery. What is actually controlled? What still requires permission? When can the asset operate, and who bears the cost if it cannot? The answers are likely to be more informative than attaching an AI label to a field.