Artificial intelligence may look like a software revolution, but its next phase will increasingly be determined by physical infrastructure. Power, cooling, networking and industrial supply chains are becoming critical constraints on how quickly AI capacity can scale.
AI growth is becoming an infrastructure problem
Artificial intelligence is usually discussed as a software story: better models, faster chips and expanding enterprise adoption. But the next phase of AI growth is increasingly being shaped by physical systems that sit far outside the model itself. Electricity, cooling equipment, network fabrics, substations, transformers, construction capacity and access to suitable sites are becoming part of the AI value chain. That changes how companies should think about digital expansion.
The International Energy Agency has made the energy dimension explicit. Its work on AI and electricity demand shows that data centres are becoming much larger actors in power systems, while AI-focused facilities are pushing rack density and local power requirements higher. The issue is not that AI will consume all available electricity; it is that AI demand is geographically concentrated, arrives quickly and often requires infrastructure that takes longer to permit, manufacture and connect.
For businesses, the implication is straightforward: the ability to acquire computing capacity is becoming partly dependent on the ability of infrastructure markets to deliver power, cooling and connectivity on time. That creates a new class of execution risk for cloud providers, data-centre operators, technology vendors and enterprises building AI-intensive services.
The chip is only one part of the bottleneck
Much of the AI investment cycle has focused on advanced accelerators. Yet a server full of high-performance processors is useful only if the surrounding systems can keep those processors supplied with electricity, data and cooling. As clusters grow, networking performance becomes more important because accelerators increasingly operate as parts of large distributed systems rather than isolated machines. NVIDIA describes AI data-centre networking in terms of scale-up, scale-out and multi-site fabrics, reflecting how tightly compute performance is now tied to interconnection.
The same logic applies to power and thermal management. McKinsey has highlighted the growing role of power and cooling equipment as data-centre demand rises. Higher-density AI hardware can require different electrical architectures, liquid cooling systems and larger mechanical capacity than conventional enterprise computing. That makes transformers, switchgear, chillers, pumps and cooling distribution units more strategically relevant than they appeared during earlier phases of cloud growth.
This is a useful reframing for investors and corporate planners. The AI infrastructure stack is not simply chips plus buildings. It is a chain of interdependent systems, and the slowest component can determine when new capacity becomes economically productive.
Power availability is becoming a location decision
Data-centre development has always been sensitive to power prices and reliability, but AI raises the importance of time-to-power. The IEA notes that data-centre electricity use is rising quickly and that grid and supply-chain constraints are becoming more visible. Large facilities can be concentrated in a relatively small number of regions, meaning national averages may hide severe local constraints.
This can alter the geography of digital investment. A location with inexpensive land but a multi-year grid connection delay may be less attractive than a more expensive site with available capacity. Regions with strong transmission infrastructure, credible generation pipelines and faster interconnection processes may gain an advantage in attracting AI-related investment. For cloud and colocation operators, power procurement and utility relationships are therefore becoming part of growth strategy rather than a back-office consideration.
For enterprise customers, this may also affect where AI workloads are hosted. Capacity may increasingly be distributed across regions according to available power and network economics, not simply user proximity. That could create more complex architecture choices around latency, resilience, data sovereignty and redundancy.
Cooling is moving from facility detail to strategic constraint
AI servers generate more heat per rack than many traditional workloads, and this is accelerating the shift toward liquid cooling and other high-density thermal designs. Cooling is often treated as an engineering detail, but at scale it can affect capital cost, site design, water use, energy efficiency and deployment timing.
The 2025 Uptime Institute Global Data Center Survey pointed to rising costs, power constraints and the difficulty of adapting existing infrastructure to AI. Legacy facilities were not necessarily designed for extreme rack densities, so operators face a choice between retrofitting existing sites and developing new capacity around modern cooling architectures.
That matters because the economics of AI infrastructure depend not only on acquiring compute but on keeping it consistently available. A cooling system that cannot support planned density can strand expensive hardware or force lower utilization. The more valuable the compute equipment becomes, the more material these supporting systems are to returns on capital.
Supply chains now extend into electrical and mechanical equipment
Another important shift is that AI-related supply risk is broadening. Semiconductor capacity remains central, but the delivery schedule for a new data centre may also depend on equipment with long manufacturing lead times. Transformers, switchgear, generators, cooling systems and specialist components can become critical-path items.
This widens the range of companies exposed to the AI capital cycle. It also makes forecasting harder. If one component arrives late, revenue associated with an otherwise completed facility can be delayed. Project developers therefore need more detailed procurement planning, supplier diversification and inventory decisions than would be required for a simpler building project.
The result is a convergence between technology strategy and industrial project management. AI infrastructure programs increasingly resemble major energy or industrial developments: they require land, permits, grid studies, equipment procurement, contractor capacity, commissioning and long-term operating discipline.
The economics of AI will be shaped by utilization
The physical infrastructure challenge also changes the financial conversation. AI equipment is capital intensive and becomes obsolete faster than many traditional infrastructure assets. That puts pressure on operators to achieve high utilization while controlling energy and cooling costs. If capacity is delayed, underused or located where power is expensive, the economics can deteriorate quickly.
The IEA energy-demand analysis shows why efficiency and workload mix matter. Data-centre electricity demand reflects not just the number of facilities, but the type of servers deployed, their utilization and the efficiency of supporting infrastructure. As AI moves from training toward large-scale inference, the demand profile may also become more continuous and geographically distributed.
For enterprises buying AI services, this may surface indirectly through cloud pricing, capacity commitments and regional availability. For providers, infrastructure efficiency can become a source of margin protection. The competitive advantage may therefore shift from simply owning scarce chips to operating a balanced system in which power, networking, cooling and compute are coordinated.
A broader definition of digital infrastructure
The current AI investment cycle is revealing a broader definition of digital infrastructure. The digital economy still depends on software, but its expansion is increasingly bounded by physical assets with very different development timelines. Electricity grids may require years of reinforcement. Large transformers and switchgear cannot always be sourced instantly. Cooling systems need engineering and water or heat-rejection solutions. Network capacity must scale alongside compute.
This mismatch between fast-moving software demand and slower physical infrastructure is likely to remain a defining feature of AI expansion. It does not imply that growth will stop. It means growth will be uneven, location-sensitive and more dependent on industrial execution.
For boards and investors, the question is shifting from “How much compute can we buy?” to “How reliably can the full infrastructure chain be delivered and operated?” That is a more complex question, but it is also a more useful one. The next generation of AI growth may be decided as much by substations, cooling loops and network fabrics as by model architecture.
Key questions
Why is AI infrastructure broader than semiconductors?
Advanced chips need reliable power, high-speed networking, cooling, buildings and grid connections. A shortage or delay in any of those systems can limit the productive use of compute capacity.
Why does location matter more for AI data centres?
Large AI facilities can require substantial power in concentrated areas. Regions with faster grid connections, reliable generation and strong network infrastructure may be able to support capacity sooner.
What is the main business risk?
The main risk is mismatch: companies may secure hardware or customer demand before the surrounding infrastructure is ready, delaying utilization and increasing capital costs.
References
• International Energy Agency - Energy and AI
• International Energy Agency - Key Questions on Energy and AI
• Uptime Institute - Global Data Center Survey 2025
• McKinsey - Beyond compute: infrastructure that powers and cools AI data centers