AI data centers are becoming a power-and-real-estate problem

The building is part of the product

AI models run on computing infrastructure. Modern data centers need servers and accelerators, but also cooling, power distribution, networking, backup systems and physical security.

The International Energy Agency estimates that data centers consumed about 415 TWh of electricity globally in 2024, around 1.5% of global electricity consumption. Its 2025 analysis projects data-center electricity use could roughly double to around 945 TWh by 2030 in its base case.

Power availability can become the bottleneck

A company can have land and capital and still be unable to open a new facility quickly if the required electricity is not available. Grid connections, transmission capacity, permitting and generation can take years to develop.

That creates an unusual situation for the technology industry: a software demand forecast can collide with infrastructure timelines that move much more slowly.

Location now has several variables

Traditional real estate questions - land, taxes, connectivity and access to workers - are joined by power availability, electricity price, water or cooling conditions and proximity to network infrastructure.

The best site for an AI data center is therefore not simply the cheapest land. It is a location where the entire infrastructure stack can work at the required scale.

Efficiency matters on both sides

Higher-efficiency chips and software can reduce the amount of compute required for a task. Better cooling and facility design can reduce overhead. Workload scheduling can also move flexible computation toward periods or locations where electricity is more available.

Efficiency does not eliminate growth in demand, but it changes how much infrastructure is required to deliver the same amount of useful AI work.

Why investors should watch the grid

The AI infrastructure story is increasingly connected to utilities, transmission, generation and industrial construction. The IEA notes that data-center electricity demand is geographically concentrated, making local grid effects more significant than the global percentage alone suggests.

The key question is therefore not only 'How many GPUs are being deployed?' It is 'Where can the electricity and infrastructure required to operate them actually come from?'

The hidden stack behind an AI model

A modern AI data center is a physical system. It needs servers and accelerators, but also substations, power distribution, cooling equipment, networking, backup systems, security and buildings that can support the required density.

That makes AI infrastructure different from a purely digital expansion. Software demand can grow quickly while electricity infrastructure and construction projects often require much longer planning cycles.

Power can determine the location

A site may look attractive because land is available, but the project can still stall if the grid cannot provide enough capacity. New connections can require studies, equipment, transmission upgrades or additional generation. The local constraint may therefore matter more than the global supply of electricity.

This creates a geographic question for AI companies: where can computing be deployed with reliable power, acceptable cost, sufficient cooling and good network connectivity?

Efficiency changes the equation

Better chips can perform more useful work per unit of energy. Software optimization, smaller models and improved scheduling can also reduce unnecessary computation. These gains matter because infrastructure efficiency can lower both operating cost and physical requirements.

Efficiency does not automatically reduce total electricity demand, however. If AI usage grows faster than efficiency improves, overall consumption can still rise. The useful question is both “How efficient is each workload?” and “How many workloads will exist?”

Why real estate is becoming technical infrastructure

Traditional real-estate analysis focuses on land cost, taxes, labor and access. AI facilities add power availability, grid timing, cooling conditions, network routes and equipment supply. A cheap site without reliable infrastructure may be less valuable than a more expensive site that can actually operate at scale.

• Check electrical capacity and connection timeline.
• Understand cooling and environmental constraints.
• Assess fiber and network connectivity.
• Model construction and equipment lead times.
• Estimate power cost under realistic workload growth.

Questions readers ask

Why does AI require so much electricity?

Training and serving models require computation, and computation requires electrical power. The amount varies by model, hardware, workload and utilization.

Will better chips solve the power problem?

They can improve efficiency, but demand can grow at the same time. Efficiency is one part of the infrastructure equation, not the whole answer.

What should investors watch?

Beyond announced GPU capacity, look at grid access, power contracts, construction schedules, cooling systems, networking and the economics of operating the facility.

The AI infrastructure cycle is becoming a bridge between software and physical economics. The winners will not simply have access to advanced models; they will also need reliable infrastructure capable of running them at sustainable cost.

Why power can matter more than the building

A data-center site is valuable only if it can support the required power, cooling, connectivity and construction schedule. A large building without dependable electricity is not ready for compute, and a site with power today may still face delays when demand grows. Grid connection timelines, transformer availability, transmission upgrades and local permitting can shape project economics as strongly as land price.

Cooling adds another layer. Different hardware densities and climate conditions change how heat must be removed, and water availability can become a local constraint. Operators need to evaluate energy efficiency alongside resilience, including backup systems, maintenance windows and the ability to keep workloads available during disruptions. These choices influence operating costs and the relationship between a facility and its surrounding community.

Local impact belongs in the project plan

Large facilities can create construction activity and demand for technical services, but they may also compete for scarce grid capacity, land or water. Communities need clear information about expected resource use, noise, traffic and infrastructure upgrades. The real effect depends on the site and the design, so broad claims that data centers are always beneficial or always harmful miss important differences.

For investors and customers, due diligence should cover more than announced capacity. Ask whether power is secured or merely planned, what the connection schedule assumes, how cooling works, whether network routes are resilient and what happens if demand forecasts change. A phased build can reduce risk when infrastructure is uncertain.

AI infrastructure is therefore both a technology and a location decision. The strongest projects align compute demand with available energy, realistic delivery timelines and responsible local planning. In many cases, the scarce asset is not the land itself but the ability to bring dependable power and connectivity to the site when they are needed.