Layer 1: The Reportable Facts

PJM Interconnection, the regional grid operator serving about 67 million people across 13 states and Washington, D.C., is advancing a framework called Interim Resource Adequacy Service, or IRAS. The plan targets new large loads of at least 50 MW, including AI data centers, that do not bring or otherwise secure enough capacity to support their demand.

Under the framework described by PJM and independent energy trade reports, qualifying new large loads would be tracked through a Large Load Registry. That registry would identify major sites, service territories, load levels, and whether the customer is bringing new supply or capacity arrangements. PJM says the registry is meant to improve forecasting, inform load-reduction priorities, and help states evaluate cost allocation for new large loads.

The operating consequence is direct: starting June 1, 2027, new large loads that lack qualifying capacity support could be reduced during capacity shortages before PJM deploys broader Pre-Emergency Load Management actions. Existing facilities are reported as outside the immediate scope of the new-load framework, and PJM itself has emphasized that state and utility coordination is essential because it does not directly control retail curtailment of individual customers.

The proposal sits beside PJM’s Reliability Backstop Procurement. PJM’s July 31, 2026 Inside Lines post said the board advanced a parallel path to secure new generation against a 6,831 MW shortfall from the 2028/2029 Delivery Year capacity auction, with procurement terms of up to 15 years and a maximum willingness to pay based on a MW-weighted average of $555/MW-day. PJM also said qualifying bilateral contracts or new resources would reduce the backstop procurement target to avoid double-counting supply.

Layer 2: The System Read

The verified fact is that PJM is trying to make large-load growth visible, attributable, and interruptible when it is not backed by capacity. The inference is that this turns AI compute into a grid-reliability product. A data center is no longer only a capital project with servers, transformers, fiber, and cooling; it is also a claim on future capacity, and PJM is trying to make that claim conditional.

This is a major change in the AI industrial flywheel. The old assumption was that hyperscalers could site where land, fiber, tax treatment, and utility interconnection looked attractive, while the regional market would procure enough supply over time. PJM’s framework says that assumption breaks when new demand arrives faster than capacity markets, interconnection queues, and state-regulated utility planning can absorb it.

The proposal also shifts the political burden. If ordinary ratepayers pay higher capacity costs because data centers arrive without matching supply, governors and state regulators own the backlash. If data centers must bring capacity or accept curtailment, developers own more of the reliability stack. That pushes hyperscalers and colocation developers toward bilateral capacity deals, dedicated generation, storage, virtual power plants, and flexible-load contracts that can perform during peak stress, not just look good in annual energy accounting.

The Pattern Nexus read: PJM is not anti-data-center; it is anti-unpriced-firmness. AI infrastructure is still welcome if it brings the grid product it consumes. The constraint is no longer simply megawatts consumed per year. It is whether those megawatts are firm, deliverable, dispatchable, and paid for by the party causing the new reliability need.

Layer 3: What To Watch Next

First, watch FERC. The practical impact depends on what federal regulators approve, reject, or modify. A narrower order could preserve the Large Load Registry while softening curtailment priority; a broader approval could give PJM a template for treating AI-scale load as conditional service across future capacity years.

Second, watch the states inside PJM. The proposal relies on state regulators and utilities to decide who gets service, who pays for new infrastructure, and which customer classes bear new capacity costs. Large-load tariffs, special contracts, and retail cost-allocation proceedings may become as important to AI data center deployment as power purchase agreements or tax incentives.

Third, watch capacity contracting. If IRAS survives, the scarce asset is not only land near transmission. It is credible capacity that can be counted during peak conditions. Developers that can pair campuses with new generation, batteries, demand-response resources, or utility-approved bilateral contracts will have a faster path to firm service than projects that arrive with load forecasts but no reliability backing.

Fourth, watch whether other grid operators copy the pattern. Data Center Knowledge reported that ERCOT, MISO, and SPP are considering similar policies. Even if each market designs its own mechanism, the direction is converging: AI load must become more forecastable, more flexible, or more responsible for the capacity it requires.

Conclusion

PJM’s proposal is a warning shot to the AI infrastructure market: firm electricity is not an entitlement at hyperscale. If a new 50 MW-plus data center wants to be treated like essential load during capacity shortages, it may need to bring capacity, contract for capacity, or provide flexibility. The next AI siting race will not be won only by whoever finds cheap land and fast permits. It will be won by whoever can turn compute demand into a reliability plan the grid can actually carry.