The race to build artificial intelligence is increasingly being measured in megawatts, not just model parameters, chip shipments or software releases. Companies can order advanced processors and erect data-center buildings relatively quickly. Securing enough reliable electricity—and the wires, substations, transformers and cooling systems required to deliver it—can take far longer.

That makes AI data center power demand a constraint on the next stage of AI expansion. The challenge is not that data centers suddenly consume all electricity everywhere. It is that very large facilities are arriving quickly in particular regions, often seeking power comparable to that used by substantial industrial operations. Local grids were not necessarily built for several such projects to connect at once.

The result is a collision between the fast-moving economics of AI infrastructure and the slower rhythms of energy planning, permitting and construction. It could affect electricity prices, grid reliability, local land use and the pace at which companies can deploy new AI services.

AI demand is growing in a system built for long lead times

Generative AI increases data-center energy consumption in two main ways. Training a large model requires running large clusters of specialized processors for extended periods. Then comes inference: the ongoing work of answering prompts, generating images, powering software features and serving business customers. A widely used model may require far more cumulative computing than its original training run.

The International Energy Agency estimated in its 2024 electricity outlook that global electricity consumption from data centers, artificial intelligence and cryptocurrency together could more than double between 2022 and 2026, rising from about 460 terawatt-hours to more than 1,000 terawatt-hours. That category includes more than AI, and projections depend on uncertain assumptions about deployment and efficiency. Still, it illustrates why utilities and grid operators are treating the sector as a major new source of load.

In the United States, a 2024 report by Lawrence Berkeley National Laboratory estimated that data centers used about 4.4% of national electricity in 2023. Under its range of scenarios, that share could reach 6.7% to 12% by 2028. AI is not the only driver: cloud computing, digital services and conventional enterprise workloads also matter. But high-density AI clusters are changing the scale and urgency of new demand.

The bottleneck is often the connection, not the building

A company may identify land, finance a campus and install servers before a utility can guarantee a connection. New data centers may require expanded substations, higher-voltage transmission lines, new generation capacity or all three. Those projects face years of engineering, land acquisition, permitting, equipment procurement and public review.

Interconnection queues are already crowded. Berkeley Lab’s annual review found that projects representing more than 2,600 gigawatts of generation capacity and nearly 1,000 gigawatts of storage capacity were seeking connection to US grids at the end of 2023. Much of that capacity will never be built, but the queue reflects a system struggling to process a surge of proposed generation and storage alongside growing demand.

Grid equipment is another practical limit. Large power transformers, switchgear and other high-voltage components have long manufacturing lead times. The US Department of Energy has identified supply-chain constraints for transformers as a grid-resilience concern. A data-center developer cannot simply bypass that problem with a faster construction schedule.

Pressure is especially visible in established data-center markets and fast-growing electricity regions, including Northern Virginia, Texas and parts of the US Southeast. These areas combine available land, fiber connectivity, tax policy and access to major population centers. They also show why location matters: national electricity supply can look adequate while a particular county or transmission zone lacks spare capacity.

Cooling turns electricity demand into a water and heat question

AI hardware produces substantial heat in a small physical footprint. Traditional air cooling remains important, but high-density racks increasingly use liquid-based approaches that move heat more efficiently. That can reduce the energy used by cooling equipment, yet it also creates new demands for plumbing, heat rejection systems and water management.

Water use is often discussed imprecisely. A facility may consume water directly through evaporative cooling, depending on its design and local climate. It also has an indirect water footprint because generating electricity can require water, particularly at thermal power plants. Neither figure is fixed: it varies by cooling technology, weather, utilization, the electricity mix and whether water is recycled or reused.

For communities facing drought or constrained water supplies, those distinctions matter. A proposal that relies heavily on potable water can raise different questions from one using reclaimed water, closed-loop cooling or a design that uses more electricity but little on-site water. Public disclosures are uneven, making it difficult for residents to compare projects cleanly.

Who pays for a grid built around large new loads?

The central policy question is not simply whether data centers should be built. It is how the costs of serving them should be allocated. A utility may need to build a substation or transmission upgrade because of a specific campus, while broader grid investments may also benefit future customers. The answer determines whether costs fall primarily on technology companies, utility shareholders, taxpayers or ordinary ratepayers.

Utilities and regulators are increasingly examining tariffs and connection agreements designed for very large customers. Such arrangements can require long-term commitments, minimum payments or contributions toward dedicated infrastructure. Supporters argue that these protections reduce the risk that households subsidize speculative projects. Critics warn that special treatment for wealthy customers can still shift risk if demand forecasts prove wrong or if new generation is built but underused.

There are local benefits: construction work, property-tax revenue and some permanent technical jobs. But data centers are not labor-intensive in the way many factories are, and communities may reasonably ask whether those benefits justify new land use, water demand, noise from backup generation and pressure on public infrastructure.

More power is not one solution

Technology companies are pursuing several routes to secure electricity. Renewable power-purchase agreements remain a major tool, although a contract for clean generation does not automatically mean a data center is physically powered by zero-carbon electricity every hour. The local grid still determines what is running when demand occurs.

Some companies are also exploring nuclear power arrangements, on-site generation and investments in grid upgrades. Gas-fired generation may provide dependable capacity more quickly in some markets, but it can increase the AI environmental impact if it extends reliance on fossil fuels. The emissions outcome depends on the regional grid, the timing of new demand and what generation is added to meet it.

Efficiency is essential but not sufficient on its own. Newer chips, improved model architectures, better utilization and smarter cooling can lower energy per computation. Yet lower costs can encourage more AI use, a familiar rebound effect. Total artificial intelligence electricity use may therefore keep rising even as individual tasks become more efficient.

• Grid modernization: build transmission, substations and flexible resources before shortages become acute.
• Demand flexibility: shift non-urgent training workloads to hours or regions with abundant power.
• Transparent pricing: ensure large new customers pay appropriately for the infrastructure they require.
• Better siting: place facilities where clean power, water and transmission capacity are available—or can be expanded responsibly.
• Efficiency standards and disclosure: make energy, water and emissions claims easier to assess.

AI’s pace will increasingly depend on public systems

The AI industry often presents computing capacity as a private competitive advantage. But the physical foundations of that capacity are deeply public: transmission corridors, utility regulation, water systems, land-use decisions and the reliability rules that keep electricity available to homes, hospitals and businesses.

That does not mean AI growth must stop. It means its expansion will be shaped by choices that cannot be solved with a faster chip cycle alone. The companies building AI infrastructure can help by making long-term commitments, paying for the capacity they need and publishing clearer information about energy and water use. Governments and regulators, meanwhile, face the harder task of accelerating needed grid investment without weakening environmental review or socializing avoidable costs.

AI’s next breakthrough may arrive in software. Its next bottleneck is more likely to be a transformer, a transmission permit or a cooling-water plan.

Image by ugoxuqu on Pixabay.