AI’s Electricity Demand Is Not the Real Problem. Its Inflexibility Is

The electricity consumption caused by artificial intelligence is typically discussed as a straightforward supply issue. Large data centers consume massive amounts of electricity, prompting utilities to build more power plants. According to the International Energy Agency, AI data centers could see their electricity usage triple by 2030, growing from 485 terawatt-hours in 2025 to roughly 950 TWh.

Some AI campuses require several gigawatts of capacity, surpassing what many cities need. However, the global figures mask the actual problem. Data centers are projected to make up only about 3% of worldwide electricity demand by 2030. While their impact is significant, it's not large enough to overwhelm the global electricity system. The challenge lies in AI's inflexibility.

The IEA anticipates data-center electricity usage to roughly double between 2025 and 2030, though it would still represent only a fraction of the total global demand growth. Industries such as industrial motors, air conditioning, electric vehicles, and broader electrification will contribute even more. Yet, these loads are distributed widely.

Over half of the data centers currently under construction in the U.S. are being built in established clusters, creating a market imbalance. Technology firms are ready to invest billions in computing infrastructure that can be built within two to three years, but the transmission lines needed to support these facilities may take four to eight years to construct.

This disparity in construction timelines has led to delays for approximately 20% of planned data-center projects.

While building more generation capacity is part of the solution, it's not the most cost-effective approach. Not every AI workload is equally urgent. Data centers are designed to be highly reliable and continuously available, but some computing tasks genuinely require this level of service, such as search queries, financial transactions, cloud applications, and many AI inference services.

However, tasks like AI training, software testing, video processing, and data backups can sometimes be postponed for several hours or moved between facilities. Grid operators cannot simply shut off these time-critical operations, but cooling systems and uninterruptible power supplies can provide some flexibility.

In March 2026, Google announced it had incorporated 1 GW of data-center demand response into agreements with several U.S. utilities. This flexibility can help new facilities connect before all the long-term generation and grid reinforcements are completed, potentially more valuable than a power-purchase agreement. By making data centers partially controllable industrial loads, rather than passive consumers, the system can avoid or postpone investments in generation and network capacity that would otherwise be used only occasionally.