AI infrastructure is no longer constrained only by algorithms or access to GPUs. It is being constrained by land, by the price and availability of power, and by the basic realities of the grid. Increasingly, electricity is what shapes the economics of compute.

For the past three years, much of the industry has behaved as though GPU access is the main determinant of scale. That assumption has underpinned hundreds of billions of dollars of investment. It is now starting to look shaky.

Infrastructure rarely breaks at a single point. It breaks when land, power, cooling, compute and permitting fail to line up in the same place at the same time. That mismatch is already visible across the US, the UK and Europe. Power is not available where it is needed. Grid connections are taking too long. Planning and permitting cannot move fast enough. Cooling systems cannot always handle the densities modern hardware demands. The problem is not any one bottleneck. It is a fact that the whole system has to work together.

That is difficult because each part of the stack moves at a different speed. GPU platforms change on a roughly annual cycle. Grid upgrades, site development and permitting can take years. Even where governments are trying to accelerate delivery, the pace of physical infrastructure still lags behind the pace of hardware development. The gap between where capital is going and what the real world can support is getting wider.

The cost base reflects that shift. Infrastructure is now driving most of the cost in AI deployment, especially power, networking and other energy-related inputs. GPU pricing matters, of course, but often less than expected. At the same time, many organisations remain highly confident in their ability to scale. That confidence looks increasingly out of step with conditions on the ground.

Greater efficiency does not solve this. In many cases, it makes the problem worse. As systems become cheaper or easier to run, they get used more, and total demand rises. The hard limits do not disappear – they remain tied to physical resources that do not scale at anything like software speed.

This is already reshaping where AI gets built. Site selection is becoming less about proximity to end users and more about proximity to available power. Regions with spare grid capacity and developable land are attracting investment, while established hubs are running into congestion, delay and saturation.

In reality, projects rarely unfold in the clean sequence imagined in planning decks. Timelines slip when grid access falls behind. Sites have to be reconsidered when permitting stalls. Backup plans become central plans once external constraints start to bite.

Under those conditions, AI infrastructure is more than a procurement challenge. It is a systems challenge.

The most common mistake is getting the sequence wrong. Hardware is secured before power is locked in. Sites are chosen before grid access is confirmed. Cooling is treated as a later-stage issue, after density assumptions have already been made. By the time the infrastructure constraints are fully understood, many of the decisions that matter most have already been taken.

The operators that scale well tend to run counter to that logic. They think in megawatts and years, not GPUs and quarters. They secure power first. They choose sites based on what can actually be delivered, not what looks best on paper. They build infrastructure that can adapt to successive hardware generations rather than a single deployment cycle. That also means planning for the broader supply chain, where critical equipment often has long lead times.

The real differentiator is not just access to money or chips. It is the ability to integrate land, power, and compute into a single system.

This report examines how that alignment is breaking down across the market and what it now takes to build within the limits that define AI infrastructure.