Why AI data centers are straining the largest power grid in the United States

PJM Interconnection, the organisation that operates the largest power grid in the United States, serving thirteen states and the District of Columbia, supplies electricity to more than 65 million people in total. The operator has said that the pace of data center construction has outstripped its ability to expand generation capacity, and it is now weighing the option of temporarily cutting power to large data centers during periods of peak demand.

That step is not a routine measure. Grid operators typically treat power cuts as a last resort, but the electricity demand from data centers built for AI training and inference has grown so quickly that PJM is now putting the option on the table as a serious possibility.

At the root of the problem is the explosion in computing power required to train and run large language models. A single major data center complex can draw as much electricity as a mid-sized city, and technology companies are bringing such facilities online one after another to keep up with demand for AI infrastructure.

There are several reasons generation capacity has failed to keep pace with that demand. New large-scale power plants — particularly natural gas plants — can take years to go from planning to operation, meaning supply is structurally lagging as demand multiplies within just a few quarters.

The situation points to an increasingly strained relationship between data center operators and grid operators. Technology companies want to add capacity as quickly as possible to avoid falling behind in the AI race, while grid operators are concerned that this rapid growth is putting the reliability of the system at risk.

If temporary cuts do go into effect, the impact would extend beyond technology companies alone to millions of households and businesses across the region. PJM's territory must manage data center load at the same time as the surge in air-conditioning demand that comes with extreme summer heat — the very season when the risk of blackouts is highest.

Some states and regulators have raised the question of whether the extra costs data centers impose on the grid should be passed on to other consumers. At the center of that debate is how much of the true cost of infrastructure expansion AI companies ought to bear themselves.

Some major technology companies are trying to engineer their own solutions to the problem: financing their own power plants, signing nuclear energy agreements, or relocating data centers to regions with more abundant electricity supply are among the strategies in play. But those solutions can also take years to come online.

Experts anticipate that short-term demand-flexibility measures — data centers agreeing to temporarily reduce their load during demand peaks — could become an increasingly common tool for preserving grid reliability. The aim is to ease consumption at the most critical moments rather than shutting data centers down entirely.

The scale of the problem illustrates how deeply AI infrastructure is now entangled with the physical world — power generation, land use, and grid planning. As AI demand continues to grow, similar strains are expected to emerge not only in PJM's territory but on other grids across the United States as well.