Building a data centre can currently be achieved in approximately 18 months. The transformer that powers this data centre, however, requires up to four years for delivery. This transformer is just one of the components whose timelines must be synchronised. Power supply, cooling, electrical equipment, permitting processes, construction, and the upgrading of utility lines all follow different timescales. These were not originally aligned.
It is possible for a single component to be available, yet the overall project cannot be realised. The next bottleneck in AI infrastructure may therefore not lie in the availability of individual resources, but in the coordination of these elements. While the physical property remains significant, the greatest challenges now lie in integrating the various systems necessary for operation.
Challenges in System Integration
The power supply must be ready exactly when the customer needs it. Cooling systems must adapt to rapidly changing rack densities. Electrical equipment may need to be procured years before actual deployment. Contractors, utility companies, and equipment manufacturers often operate on very different schedules. PJM provides a clear example of how quickly bottlenecks can shift. Their reformed interconnection process can now process new power generation projects in one to two years.
Nevertheless, projects totalling approximately 57 gigawatts have completed the PJM study process and signed or been offered interconnection agreements. Many of these projects continue to be slowed or stopped by factors outside the PJM process. A PJM analysis of the recent capacity shortage further illustrates the discrepancy. A gigawatt-scale load can materialise within 12 to 18 months, while the power generation required for it takes significantly longer. Permitting processes were responsible for 29 per cent of the changes in milestones for power generation projects tracked by PJM since 2023. Generator transformers and gas turbines can take three to four years to procure.
Consider the power supply: a project developer may control a plot of land next to transmission infrastructure and still not be able to realise a project if the utility cannot provide capacity in time. On-site self-generation offers an alternative path but brings its own requirements for generation, fuel supply, grid connection, and operation. These elements must now be coordinated with the data centre itself. Once the power problem is solved, another challenge may arise.
Complexity of Cooling and Construction Technology
AI racks sometimes require very different cooling systems than traditional computing technology. Liquid cooling involves pumps, heat exchangers, piping, controls, and water management. Each component may function individually, but the challenge lies in aligning the entire thermal system with the IT infrastructure. Construction faces similar problems. A hyperscale campus can involve thousands of workers and multiple contractors, requiring enormous quantities of electrical and mechanical equipment.
At the same time, the technology used in the building can evolve faster than the construction itself. Traditional construction management divides complexity into packages and coordinates them according to a schedule. AI infrastructure increasingly requires something more akin to system orchestration. A transformer manufacturer supplies a transformer, a chiller company supplies a cooling system, and a utility company supplies power. None of them delivers a fully functional AI infrastructure campus. The value increasingly lies in getting these systems to work together at the right capacity and at the right time.
Therefore, it is also difficult to describe AI infrastructure simply as another real estate asset class. Operators are increasingly confronted with the coordination of utilities, thermal management, control integration, commissioning, and generation strategy. These projects combine elements of energy infrastructure, industrial systems engineering, software operations, and network architecture. Real estate remains essential, but the building is only one part of the overall system.
The coordination problem also extends beyond the property boundary. A utility company may plan transmission infrastructure on one schedule, while a hyperscaler plans the deployment of computing capacity on another. Permitting authorities, water utilities, and economic development organisations may operate on yet other plans. A viable project must successfully navigate all these levels. This explains why projects that seem feasible on paper can become difficult to execute.
A site is not simply a project because it owns land. Power is not deliverable capacity just because it appears in a utility plan. Cooling systems are not a thermal solution until they work with the computing architecture. The project only functions when enough of these components are available simultaneously. This changes where a competitive advantage can arise. In the first phase of the AI infrastructure boom, companies that controlled scarce components benefited: land, power, graphics processing units, and capital.
These advantages remain important. However, another capability is becoming increasingly scarce as projects grow larger and more complex: the ability to orchestrate them. This could favour companies that can coordinate utilities, energy developers, equipment manufacturers, contractors, and technology companies around an actionable timeline. The next generation of AI infrastructure companies might resemble industrial partners as much as real estate developers. Their advantage will be in making the entire project function as a unit.
The AI infrastructure industry has spent the last few years securing land, power, equipment, and capital. Individually securing these components is no longer sufficient. The next phase will be characterised by whether companies can transform these into functioning infrastructure within the timeframe actually required by customers.
