AI demand is exposing the largest US power grid to curtailment risk

Data centers on the PJM grid—the largest in the US—may face forced shutdowns to prevent blackouts. AI infrastructure is arriving faster than transmission capacity can support it.

In plain terms: Data centers on PJM Interconnection—which serves 65 million people across 13 states—may face temporary forced power cuts to prevent grid instability. AI infrastructure buildout is outpacing grid upgrade timelines in the region with the highest data center density in the world.

The problem isn't long-term capacity; new generation is coming. It's timing. Data centers are coming online faster than transmission infrastructure can be upgraded to safely deliver power to them at peak demand. Grid operators' tool for managing this mismatch is curtailment—temporary forced shutdowns when demand exceeds safe operating limits.

For data centers, curtailment during an AI training run isn't a minor inconvenience. It corrupts checkpoints, wastes compute time, and can require full restart of multi-day runs. The reliability assumption that underlies most hyperscaler infrastructure planning doesn't hold if curtailment events become routine.

So what: Backup power and grid independence are becoming infrastructure requirements, not contingency planning. Watch for data center operators shifting capex toward on-site generation—nuclear, gas, large-scale battery—as grid reliability becomes a site selection and procurement criterion.