The largest power market in the United States has spent the better part of two years trying to reconcile an immovable object with an unstoppable force. The immovable object is a transmission grid whose ability to add firm generation is constrained by long interconnection queues, supply-chain delays, and the slow arithmetic of building large machines. The unstoppable force is the artificial-intelligence build-out, which has made data centers the fastest-growing category of electricity demand the country has seen in a generation.
In late July 2026, PJM Interconnection’s board of managers advanced a package of proposals that attempts to square the two by changing what it means to be a large customer on the grid. The central idea is deceptively simple: if new load wants to connect faster than new supply can be built, that load must agree to step back when the system is under stress. It is a proposal to make the biggest new consumers of power interruptible by design, and it may prove to be one of the most consequential shifts in U.S. power planning this decade.
A New Compact Between Load and Grid
For most of the modern grid’s history, the compact between a utility and its customers has been one-directional in a specific sense: the customer pays, and the utility guarantees firm, uninterrupted service. Demand response programs have long created exceptions for large industrial users willing to trim consumption in exchange for compensation, but those arrangements were voluntary sweeteners layered on top of an implicit promise of reliability. PJM’s new framework inverts the default for a specific and rapidly expanding class of customers. Under the board’s proposal, new large loads of 50 MW or more that seek to connect without bringing their own dedicated, firm generation would do so on the explicit condition that they can be curtailed when the grid approaches emergency conditions. The reliability of the broader system, in other words, would no longer be underwritten entirely by the supply side. It would be shared with the demand side, and specifically with the hyperscale facilities driving the load surge.
This is a meaningful philosophical departure. PJM serves roughly 67 million people across thirteen states and the District of Columbia, a footprint stretching from Virginia to Illinois and including the densest concentration of data centers on the planet in Northern Virginia’s “Data Center Alley.” The operator has concluded that it cannot build firm capacity fast enough to serve the load developers want to interconnect, and that continuing to promise unconditional firm service to every new hyperscale campus would either force costly overbuilding or expose the system to unacceptable reliability risk. The proposed answer is to treat flexibility as a resource in its own right. Large-load flexibility becomes not a nice-to-have but a structural feature of how the grid keeps the lights on during the tightest hours of the year.
The Numbers Behind the Urgency
The scale of the demand shock explains why PJM felt compelled to act. The operator has forecast enormous large-load growth over the coming decade, with projections pointing toward roughly 70 GW of new large load by 2038 and interim estimates placing tens of gigawatts of that demand before the end of the decade. To put that figure in perspective, 70 GW approaches the total peak demand of several midsized ISOs combined; it is the equivalent of bolting a large new grid onto the existing one. Independent analyses tracking the AI build-out have suggested that data center electricity demand could roughly quadruple by the mid-2030s relative to today, and PJM’s own planning assumptions reflect a world in which the appetite for compute shows little sign of leveling off.
The consequences of that surge are already visible in wholesale prices. Over the past year, wholesale power prices across the PJM footprint have climbed sharply, with the market’s independent monitor attributing a substantial share of the increase to data center demand pressing against a supply stack that has not kept pace. Capacity costs, the payments generators receive for promising to be available during peak conditions, have reached levels that would have been unthinkable a few years ago. The pressure on prices and reliability margins is not a distant forecast; it is a present condition, and it is the immediate context in which the board acted. When demand growth of this magnitude collides with a generation fleet that cannot expand on the same timeline, something has to give, and PJM has decided that the something should be the firmness of new load rather than the reliability of the system as a whole.
How Curtailment Would Work
The mechanics of the proposed data center curtailment regime borrow heavily from demand response programs that grid operators have refined over decades, which helps make the approach credible rather than experimental. Facilities subject to the new obligations would receive advance notice before any reduction, with reported lead times ranging from as little as thirty minutes to as much as several days, depending on how conditions develop. During a curtailment event, an affected data center would be expected either to reduce its draw from the grid or to transition to on-site backup generation, keeping its own operations running while relieving the system of its load. Crucially, participation would be compensated, preserving the logic that has made demand response palatable to large industrial customers: flexibility has value, and those who provide it should be paid for it.
To make the system workable, PJM would build a registry of participating facilities that captures their location, capacity, and ability to ramp load up and down. Curtailment rules would be developed by utilities and transmission owners and would remain subject to approval by state and local regulators, a design choice that keeps the operator from unilaterally dictating terms to customers and preserves the states’ role in overseeing retail service. The obligation would attach to a well-defined tier of customers rather than the general population; residential and commercial ratepayers would not be asked to go dark so that a data center can keep computing. By design, the burden of flexibility falls on the loads that are creating the need for it.
The Backstop Auction and the Cost Question
The curtailment framework does not stand alone. Alongside it, the board advanced a backstop capacity procurement intended to plug a near-term supply gap, targeting about 6.8 GW of resources to address a shortfall left by the most recent capacity auction. The backstop mechanism would run in the fall of 2026, with PJM acting as administrator and counterparty and offering long-term commitments of up to fifteen years to give developers the revenue certainty they need to finance new plants. To draw out supply that the standard market has struggled to attract, the backstop would carry a higher price ceiling, reported at $555/MW-day, well above the temporary cap that has governed recent auctions. A bilateral matchmaking process, overseen by an outside firm, would attempt to pair specific large loads with specific new resources.
Behind these design details sits a politically charged question: who pays? PJM’s stated intent is that the costs of the backstop procurement and of the accommodations built for hyperscale growth should be allocated to the data centers and large industrial loads that necessitate them, rather than spread across the millions of households and small businesses on the system. That principle of cost causation is easy to state and notoriously difficult to implement, because the grid is a shared machine and the line between a cost caused by one customer and a benefit enjoyed by all is rarely crisp. Still, the direction of travel is clear. Regulators and consumer advocates have grown increasingly vocal about the risk that ordinary ratepayers will subsidize the AI boom, and PJM’s proposal is calibrated in part to answer that concern before it hardens into political opposition that could stall large-load interconnection altogether.
FERC’s Parallel Push
PJM is not acting in a vacuum. Federal regulators have been moving on a parallel track, and the two efforts are best understood as pieces of the same puzzle. In June 2026, the Federal Energy Regulatory Commission issued show-cause orders directing every organized market under its jurisdiction, PJM included, to justify or reform the rules governing how data centers, manufacturing facilities, and other large energy users connect to the grid. The commission pressed grid operators to develop efficient processes for connecting large loads, prevent cost shifts onto existing customers, address the thorny issues raised by colocation and behind-the-meter generation, and create transmission service options flexible enough to accommodate customers that can vary their consumption. FERC framed the effort as region-specific rather than one-size-fits-all, recognizing that the operational realities of the western grid differ from those of the Eastern Interconnection, but the underlying message to every market was the same: the old rules were written for a world that no longer exists.
The regulatory momentum extends to reliability standards as well. In July 2026, FERC directed the North American Electric Reliability Corporation to develop new or modified mandatory reliability standards addressing the integration of computational loads into the bulk-power system, along with registry criteria for entities that may need to come under the mandatory reliability framework. Large blocks of data center demand can drop off the system abruptly during voltage or frequency events, and engineers are still learning how that behavior interacts with the machines that stabilize the grid. Grid reliability in a data-center-heavy PJM, once a niche technical concern, has moved to the center of the reliability conversation. PJM’s curtailment proposal and FERC’s broader directives are converging on a shared conclusion: large loads must become active, cooperative participants in maintaining system stability rather than passive consumers whose only obligation is to pay their bills.
What It Means for Developers and Utilities
For data center developers, the proposal reframes the calculus of where and how to build. A campus that arrives with its own firm generation, or that pairs its load with renewables backed by long-duration storage, could earn expedited treatment and a firmer claim on the grid. A campus that shows up expecting unconditional firm service without contributing new supply will face less generous terms, trading speed of interconnection for an obligation to curtail. That trade-off will push developers to think harder about on-site generation, backup capacity, and their workloads’ operational tolerance for interruption. Some AI training workloads are more schedulable than the always-on demands of a cloud region, and the facilities best positioned to thrive under the new regime may be those that can genuinely flex.
For utilities and the states that regulate them, the proposal offers a framework for saying yes to economic development without betting system reliability on generation that may not arrive on time. It also gives them a difficult implementation task, from writing curtailment rules that are fair and enforceable to standing up the registries and settlement systems that make compensation work. The FERC filings that will carry these proposals forward are expected soon, and the paper hearings and stakeholder fights that follow will determine whether the elegant logic of the board’s design survives contact with the competing interests of generators, load, consumer advocates, and states. What is no longer in doubt is the premise. The era in which the largest new loads on the grid could expect unconditional firm service is ending, and a new compact, one in which flexibility is the price of speed, is taking its place.
Conclusion
PJM’s late-July package is, at bottom, an admission that the supply side cannot carry the entire burden of the AI era alone, followed by a constructive proposal for what to do about it. By making curtailability the condition of fast interconnection for large loads, backstopping near-term supply with long-term contracts, and insisting that the costs of accommodating hyperscale growth fall on the loads that cause them, the operator is attempting to hold reliability, affordability, and economic growth in a single frame. The design will be contested, and the details that emerge from the FERC process may look different from what the board proposed. But the direction is set, and it aligns with the federal push toward flexible large-load transmission services and reliability rules fit for a data-center-heavy grid. The question facing the industry is no longer whether large loads will be asked to bend for the grid, but how gracefully, and at what price, they will learn to do so.