The artificial intelligence boom is often described as a software revolution, but the economics of the next phase may be shaped as much by substations, transmission lines, transformers, cooling systems and construction schedules as by algorithms. The newest generation of AI systems requires enormous amounts of computing capacity, and that capacity has to exist somewhere physical. As investment accelerates, the constraint is increasingly not whether companies can imagine new AI products, but whether they can secure the power, equipment, sites and grid connections needed to run them at scale.
That shift matters well beyond the technology sector. Data centres are becoming a meaningful source of electricity-demand growth in several major economies, while utilities, governments and infrastructure providers are being pushed to rethink how fast power systems can expand. For investors and corporate planners, this changes the way digital growth should be understood: the software layer can scale in seconds, but the supporting physical infrastructure may take years.
AI demand is moving into the physical economy
The scale of the shift is visible in the latest energy forecasts. The International Energy Agency's Key Questions on Energy and AI notes that data-centre electricity use surged in 2025 and that AI-focused facilities are increasing their power consumption even faster. The agency says the capital expenditure of five large technology companies exceeded $400 billion in 2025 and was expected to rise sharply again in 2026, illustrating how rapidly the sector is moving from experimentation to infrastructure build-out.
The important point is not simply that AI uses electricity. Digital services have always depended on servers, networks and cooling. What is different is the concentration and speed of new demand. Large AI clusters can require power at a scale that is difficult to accommodate within existing local grids, and high-density computing places additional pressure on electrical equipment and cooling design inside the data centre itself. The result is a mismatch between the speed of digital investment and the speed at which physical systems can be permitted, financed, manufactured and connected.
The IEA's Electricity 2026 describes a broader 'Age of Electricity' in which global power demand is rising quickly because of electrification, advanced manufacturing, data centres and AI. It estimates that more than 2,500 gigawatts of generation, storage and large-load projects are stalled in grid connection queues worldwide. That backlog is a reminder that electricity networks are becoming a competitive factor for digital development, rather than a background utility that can be assumed to expand automatically.
Grid access is becoming a strategic asset
For years, companies building digital infrastructure competed mainly on access to fibre, land, customers and tax incentives. Increasingly, power availability is joining that list near the top. A site with good connectivity but no realistic path to a grid connection may be far less valuable than one with lower-profile geography but a credible route to additional megawatts.
The IEA's dedicated analysis of grid bottlenecks says record connection queues are emerging as a critical constraint on new supply, storage and demand. The agency argues that meeting forecast electricity demand through 2030 will require a major increase in grid investment as well as larger supply chains for transformers, cables and other equipment. For data-centre operators, this turns grid planning into a core part of capacity planning.
The commercial consequences are substantial. If power cannot be secured on schedule, a completed building may sit underused, expensive computing equipment may arrive before the supporting infrastructure is ready, or projects may be redirected to different regions. This gives utilities, transmission operators and local regulators a greater role in determining the geography of digital growth.
It may also alter the negotiating balance between technology companies and energy providers. Large data-centre customers can bring long-term demand and investment, but they can also require unusually concentrated capacity. Utilities therefore have to consider how new connections affect system reliability, network upgrades and costs for other customers. The economics of AI infrastructure are becoming inseparable from questions about who pays for new generation and grid capacity, and how quickly those assets can be delivered.
Efficiency is improving, but total demand can still rise
One reason the debate is complicated is that computing efficiency and electricity demand can move in opposite directions. Chips, models and software can become more efficient even while overall power use rises, because lower costs and better performance encourage more usage. This rebound effect is familiar in other technologies: efficiency reduces the resource required per unit, but demand for the service itself can grow even faster.
The IEA's 2026 update says power consumption per AI task is declining rapidly, yet the number and complexity of AI uses continue to expand. Its earlier Energy and AI report projected global data-centre electricity consumption could more than double by 2030. This is why efficiency improvements, while important, do not remove the infrastructure challenge. They may simply allow more computing to be performed for the same amount of power before demand expands again.
The United States illustrates the scale of the issue
A 2026 update from the U.S. Department of Energy's Lawrence Berkeley National Laboratory, the United States Data Center Energy Usage Report: 2025 Update, estimates that data centres could account for about 11.8% of total U.S. electricity use by 2030 in its central estimate, with a wide scenario range around that figure. The estimate is based on a bottom-up model using expected IT equipment shipments, power characteristics, cooling performance and facility types.
The exact number matters less than the direction of travel. In regions where power demand had been relatively flat for years, utilities are now planning for faster growth from data centres, manufacturing, transport electrification and other large loads at the same time. That increases competition for equipment, skilled labour and capital across the power system.
It also complicates assumptions about location. Historically, some data-centre markets became dominant because of network connectivity and established clusters. If electricity becomes harder to obtain in those locations, growth may move toward areas with stronger grids, available generation or faster permitting. Digital infrastructure could become more geographically dispersed even as computing services remain globally accessible.
Data-centre design is being forced to change
The latest Uptime Institute Global Data Center Survey 2026 says operators continue to expand and invest but face intensifying limits involving power availability, grid reliability, costs, supply chains and staffing. These pressures are arriving while AI workloads are pushing facilities toward higher-density computing environments.
That combination makes design flexibility more valuable. Operators may need to support different rack densities, cooling architectures and power configurations over the life of a facility. Traditional assumptions about how much power a building will need per square metre are being challenged by clusters of accelerators that can draw far more electricity than conventional enterprise servers.
Cooling is part of the same equation. Higher-density computing can require liquid-cooling systems or other approaches that change capital costs, water considerations and maintenance requirements. This means that data-centre economics can no longer be assessed only through the price of servers and real estate. Power quality, cooling capacity, local climate, network resilience and access to technical staff increasingly influence whether a site can support the next generation of workloads.
Regulation is moving closer to the infrastructure layer
Governments are also paying closer attention to the physical footprint of digital infrastructure. The European Commission data-centre energy performance framework requires monitoring and reporting of energy performance for significant data centres and is developing an EU-wide rating approach. The Commission has also signalled plans for minimum performance standards and greater transparency around energy and water use.
This does not mean policymakers are trying to stop digital growth. In many jurisdictions, the opposite is true: governments want more AI capacity, cloud infrastructure and domestic computing capability. But the policy objective is increasingly two-sided. Authorities want investment while also protecting grid reliability, affordability and environmental resources. That creates a more complex approval environment in which developers may need to demonstrate how projects will integrate with local energy systems rather than simply how much capital they will bring.
A new class of infrastructure winners may emerge
The transition has implications across sectors. Grid equipment manufacturers, engineering firms, power developers, energy-storage providers, cooling specialists, fibre operators and construction companies all sit beneath the AI stack. Their capacity can determine how quickly technology companies convert capital budgets into usable compute.
This broadens the economic footprint of AI. The most visible companies may be those developing models and chips, but the enabling layer includes businesses that have traditionally been treated as industrial or utility infrastructure. If demand continues to outpace supply in critical components such as transformers or specialist electrical equipment, lead times and procurement discipline may become as important as processor availability.
For corporate users of AI, the lesson is similar. Cloud capacity is often presented as elastic, but elasticity ultimately depends on someone having built physical capacity in advance. Companies that assume computing resources will always be available at stable prices may discover that the economics of AI services are influenced by power markets and infrastructure constraints several layers below their software provider.
The investment case needs a longer time horizon
Physical infrastructure changes the timing profile of the AI cycle. Software can be launched quickly and upgraded continuously, while grids, substations and generation projects often require long planning and construction periods. This introduces a risk of timing mismatches: computing demand may grow faster than infrastructure for several years, followed by periods when new capacity arrives in large blocks.
That possibility argues for caution around simple extrapolation. Power constraints could delay projects or raise costs, but aggressive infrastructure build-outs could later relieve bottlenecks. Technology itself may also change the equation through more efficient chips, better model architectures, workload scheduling and energy management. The infrastructure story is therefore not a one-way prediction of permanent scarcity. It is a question of how quickly supply can adapt relative to demand.
What businesses should watch next
Several indicators will help show whether the physical bottleneck is easing or intensifying. Grid connection times are one. Delivery schedules for transformers and other power equipment are another. Data-centre vacancy rates, power pricing, local permitting decisions and the pace of generation additions will also matter. On the technology side, improvements in compute efficiency and the spread of smaller or more specialised models could reduce the amount of electricity required for some workloads.
The most important shift is conceptual. AI is no longer only a technology-sector story. It is becoming a capital-intensive infrastructure story that links digital services to electricity markets, industrial supply chains, construction capacity and regulation. That makes the next phase of the AI cycle potentially slower, more geographically uneven and more dependent on physical execution than the software narrative suggests.
For Global Banking & Finance Review readers, the significance is broader than data centres themselves. The financing of grids, power generation and supporting infrastructure will influence where digital investment is possible and at what cost. Banks, investors, corporates and policymakers will increasingly encounter the same question from different angles: can the physical economy expand fast enough to support the ambitions of the digital one?
Key Questions
Why is AI becoming an infrastructure issue?
Because large-scale AI depends on data centres, electricity, cooling, grid connections and specialised equipment. Growth in software demand can therefore be limited by the speed at which physical capacity is built.
Will efficiency improvements solve the power problem?
They can reduce the electricity needed per task, but total demand may still rise if AI usage grows faster than efficiency gains.
Why do grid connection queues matter?
A data centre cannot operate at planned capacity without reliable power. Long connection delays can postpone projects, change their economics or push development to other regions.
Which sectors could benefit from the infrastructure build-out?
Potential beneficiaries include grid equipment, engineering, power generation, energy storage, cooling, construction and network infrastructure. Outcomes will vary by company and region.
What is the main risk to the outlook?
Demand, regulation and technology can change quickly. Overbuilding is possible if AI demand develops more slowly than expected or if computing efficiency improves faster than infrastructure planners assume.
References
• International Energy Agency - Key Questions on Energy and AI (2026)
• International Energy Agency - Electricity 2026
• International Energy Agency - Grids chapter, Electricity 2026
• International Energy Agency - Energy and AI (2025)
• Lawrence Berkeley National Laboratory - United States Data Center Energy Usage Report: 2025 Update