AI Data Centre Demand Is An Inflexibility Problem

The surge in electricity consumption driven by artificial intelligence is frequently framed as a straightforward supply challenge, yet this perspective overlooks the more critical issue of grid inflexibility. While data centres are projected to consume approximately 485 terawatt-hours globally in 2025 and potentially rise to around 950 TWh by 2030 according to International Energy Agency forecasts, these facilities will still represent only about three per cent of worldwide electricity demand at that time. The primary difficulty lies not in the total volume of power required but in how this load arrives in massive blocks within specific locations on timelines far shorter than those needed for traditional grid expansion.

Most industrial loads such as electric vehicles and air conditioning are widely distributed across millions of sites, whereas new AI campuses can introduce hundreds of megawatts through a single connection. Currently, nearly half of existing US data centre capacity is concentrated in five regional clusters, with many development projects also targeting these established areas. This concentration creates bottlenecks because technology firms plan to build computing infrastructure within two or three years while transmission lines and transformers may require four to eight years to construct. The IEA estimates that roughly 20 per cent of planned data centre projects could face delays if these sector-specific constraints are not addressed.

Although building additional generation capacity through renewables, natural gas, and nuclear power is necessary, simply meeting the theoretical maximum demand at every moment would be prohibitively expensive. Grid planners often treat data centres as fixed loads that cannot be interrupted because many computing tasks require immediate response times for search queries or financial transactions. However, other workloads such as AI training, software testing, video processing and data backups can sometimes be postponed for several hours or shifted between facilities without compromising service quality.

This distinction is becoming commercially relevant as major technology companies begin to incorporate demand response into their operations. In March 2026, Google announced agreements with several US utilities allowing it to temporarily shift selected machine-learning workloads when local grids are under stress. Previous collaborations with Indiana Michigan Power and the Tennessee Valley Authority demonstrated how such flexibility could enable new facilities to connect before all necessary grid reinforcements were complete.

The economic argument against curtailment is strong, as unused computing capacity generates no revenue and AI infrastructure is significantly more capital intensive than other industrial users like aluminium smelters. However, operators do not need to shut down entire sites; instead they can utilise spare server capacity differently or discharge onsite batteries during peak periods. Electricity market rules must evolve to reflect this reality by offering faster connections with lower tariffs for verified demand reduction rather than charging data centres as if they require continuous full contracted capacity.

Strategic planning should begin before construction commences, influencing where facilities are sited and how they are designed. A location near available generation equipped with batteries offers different system costs compared to one requiring uninterrupted maximum supply in a constrained urban area. Waste heat recovery also provides value only when incorporated into the original siting decision alongside nearby district-heating networks rather than advertised after an unsuitable location is chosen.

While AI will continue driving investment in natural gas, renewables and nuclear power across various markets, the most successful regions will be those that develop a combination of generation capacity flexible contracts and rapid connection capabilities. The competition for artificial intelligence energy resources should focus on integrating large new consumers without making the electricity system more expensive or fragile rather than simply producing additional electrons at any cost.