AI Infrastructure: The Physical Foundation of Artificial Intelligence

The Physical Dimension of AI

Artificial intelligence is primarily perceived as a digital technology. Models, algorithms, computing capacity and data availability understandably dominate much of the discussion, while the material infrastructure required for these systems to operate and continue to grow remains less visible. Behind the increase in computing capacity lies a physical chain made up of servers, electrical systems, grid connections, transformers, uninterruptible power supplies, cooling systems, any generation and storage capacity, water availability, space and ancillary infrastructure.

The growth of AI is therefore generating increasing demand for industrial systems that must be available simultaneously and provide reliability levels consistent with the required service. According to the International Energy Agency, global electricity consumption by data centers was approximately 485 TWh in 2025 and could reach around 950 TWh by 2030, effectively doubling in just five years. Over the same period, data centers more heavily focused on AI workloads are expected to grow faster than the rest of the sector. The energy figure immediately conveys the scale of the phenomenon, but on its own it does not capture the complexity of the AI infrastructure required to support its evolution.

System Capacity

When a function expands within an industrial system, the consequences normally extend to the subsystems that support it. Increasing the output of a machine may require greater power supply, additional upstream and downstream capacity, different auxiliary systems, increased cooling or a redesign of existing interfaces. The growth of computing capacity follows a similar logic: higher computational demand may be compatible with available technologies while still encountering limits in accessible electrical power, local grid capacity, connection lead times or the availability of the required components.

The International Energy Agency is already reporting constraints related to the supply of transformers, turbines, electrical equipment and IT components, as well as the time required to develop new grid infrastructure. It is therefore useful to distinguish between theoretically installable capacity and capacity that can actually be deployed. The former depends largely on the availability of the technology; the latter requires that technology to be integrated into a physical system with the necessary resources, connections and operating conditions.

Digital technologies and physical infrastructure evolve at different speeds. This timing gap creates a concrete constraint: computing capacity can expand faster than grids, facilities and supporting subsystems can be upgraded.

Interdependent Infrastructure

Energy availability is one of the most visible constraints, but its effect depends on how it interacts with other subsystems. Much of the power used by servers is converted into heat and must be dissipated, which means that increasing computing density also changes cooling requirements and may call for architectures different from those previously used. Energy supply and thermal management therefore need to be considered as parts of the same system.

Water can also play a significant role, although its relevance varies considerably. A study by Lawrence Berkeley National Laboratory found differences spanning several orders of magnitude in the water consumption associated with data center workloads, depending on server efficiency, cooling technology, climate, facility utilization and the characteristics of the electricity grid. This highlights how difficult it is to reduce resource requirements to a single parameter and how strongly overall performance depends on the chosen architecture and site conditions.

Energy, cooling, water, climate, grid capacity and space therefore combine to determine the capacity that is actually available. The compatibility between these elements affects whether new capacity can be installed, expanded over time and kept operational under the required conditions.

Location

When these dependencies are considered together, site selection also takes on greater technical importance. The theoretical availability of power does not guarantee that it can be supplied continuously; the availability of space does not resolve potential grid limitations; and the presence of water does not automatically mean that the resource is compatible with the chosen cooling system.

Location therefore becomes part of the technical architecture of the infrastructure. There is also the strong geographical concentration of data center growth: their share of global electricity consumption remains relatively limited compared with overall demand, but large loads concentrated in specific areas can have a more significant impact on local grids and available infrastructure.

The same level of computing capacity can therefore have very different implications depending on where it is installed. Site assessment needs to consider energy availability, grid capacity, climate conditions, cooling requirements, space, ancillary infrastructure and future expansion potential as part of a single evaluation.

Development Times

The complexity of physical dependencies is compounded by differences in development times. Digital technologies can evolve rapidly, while the energy infrastructure, grid connections and industrial systems required to support them involve design, procurement, permitting and construction. When computing demand grows in ways that are difficult to predict, these different timelines can become a determining factor.

Initial sizing therefore represents only one part of the issue. It is also necessary to consider how the infrastructure will respond if demand increases beyond initial forecasts. Oversizing can entail high costs for unused capacity and, at the same time, does not guarantee that the architecture will remain consistent with technological development.

Under conditions of high uncertainty, the adaptability of the infrastructure therefore takes on a central role. Design decisions need to account for the possibility that forecasts will change and that future growth may require configurations different from those originally envisaged.

Scalability

When future demand cannot be defined with sufficient precision, scalability becomes a particularly relevant design criterion. Modularity, functional separation of subsystems, provision for interfaces and the ability to expand are already used in many industrial sectors to manage configurations that are expected to evolve over time.

Applying these principles to infrastructure supporting computing capacity means assessing both the power required for the initial configuration and the possibility of subsequently adding electrical capacity, cooling, storage or additional modules without having to redesign the entire system. It is also necessary to identify which elements can be expanded relatively easily and which constitute constraints that are difficult to change once the infrastructure has been built.

The difficulty of accurately forecasting demand several years in advance makes it more important to define from the outset how the infrastructure will be able to grow. Scalability allows part of the design uncertainty to be absorbed through an architecture designed to evolve, reducing dependence on forecasts that are necessarily subject to revision.

Effects on Industry

The evolution of AI infrastructure also affects established industrial sectors. Power generation and distribution systems, transformers, storage, cooling, components and auxiliary subsystems all belong to existing industrial supply chains, but the growth of computing capacity can alter demand levels, performance requirements and the degree of integration required.

One indication of this shift emerged in discussions with a company active in the development of subsystems for energy and oil & gas applications, which is also working on solutions for powering data centers. A single case cannot be used to generalize about the direction of the market, but it does provide a concrete example of how expertise and technologies developed for other sectors can find applications in the new infrastructure required by growing computing demand.

For some companies, this may lead to the adaptation of products and expertise toward data center applications while maintaining technical continuity with systems they already know. Digital transformation therefore also affects supply chains that may appear far removed from software, because an increasing share of its evolution depends on physical components and systems.

The Evolution of AI Infrastructure

The growth of computing capacity makes it increasingly important for physical infrastructure to evolve alongside the technology it is expected to support. Energy, grid capacity, thermal management, resource availability, site characteristics and development times become elements of the same design problem and need to be assessed with future expansion in mind.

Within this framework, modularity, scalability and adaptability make it possible to address an evolution that is difficult to predict without locking the project into a rigid configuration. The growth of artificial intelligence therefore continues to depend on physical systems and industrial timelines that need to be considered from the outset as an integral part of AI infrastructure.

For businesses, this evolution raises an issue that concerns both those that will use growing computing capacity and those that manufacture the machines, components and subsystems needed to support it. The former will need to assess more carefully the compatibility between digital growth, the availability of utilities and the capacity of their own infrastructure; the latter may encounter new requirements from markets that apply familiar technologies under different operating, dimensional and performance conditions. Understanding these interdependencies early can therefore become part of industrial planning, because investments, products and infrastructure designed today may be required to operate in a scenario where computing capacity grows faster than many of the physical systems needed to support it.

Sources Cited in the Article:

• (1) International Energy Agency. (2026, 16 aprile). “Key Questions on Energy and AI.”

• (2) International Energy Agency. (2025, 10 aprile). “Energy and AI – Energy demand from AI.”

• (3) Lawrence Berkeley National Laboratory. (2025, giugno). “The water use of data center workloads: A review and assessment of key determinants.”

• (4) International Energy Agency. (2026, 16 aprile). “Data centre electricity use surged in 2025, even with tightening bottlenecks driving a scramble for solutions.”