PJM Interconnection's 2026 long-term load forecast does something that few forecasts in the data center era have done. It lowers expected demand for the next five years. The grid operator now projects summer peak load in 2027 about 2.3% lower than it did a year ago, roughly 3,700 MW, and 2029 about 2.7% lower, around 4,600 MW. Then the curve bends back up. By 2040 the new forecast sits 7.8% above last year's, an extra 18,600 MW or so. The 20-year annualized growth rate for summer peak rises from 2.0% to 2.4%.
That shape is not an accident or a change of heart about artificial intelligence. It is the product of new screening rules for large load adjustment requests, the utility submissions that tell PJM a data center campus is coming to a particular zone. The headline figures are still enormous. PJM expects summer peak to reach 222,106 MW in 2036, a ten-year increase of 65,733 MW, and 253,077 MW by 2046. Net energy for load is projected to grow 5.3% a year over the next decade, reaching 1,437,629 GWh in 2036, an increase of 581,554 GWh. For a system that has spent most of the past two decades with flat demand, those are numbers on the scale of adding several mid-sized states.
What changed in the method
The January presentation to PJM's Load Analysis Subcommittee sets out the new filters. PJM went through utility submissions and identified projects that had an Electric Service Agreement or a Construction Contract in place. Those were labeled "Firm." Only firm projects are allowed to affect the Reliability Pricing Model, the capacity market that sets what generators are paid to be available. Everything else was labeled "Non-Firm" and kept out of the numbers that drive procurement.
Two further adjustments follow from what PJM learned about how data centers actually come online. The first is a ramp rate. PJM imposed a minimum of 36 months for a data center to go from first energization to full demand, on the evidence that campuses fill up in phases as servers are installed. The second is a utilization rate. PJM applies 70% of a project's final contracted capacity unless the utility can support a higher figure. A campus with a 300 MW service agreement therefore shows up in the forecast as 210 MW, reached over at least three years.
The grid operator also tried to stop counting the same load twice. In zones where data centers already operate, PJM removes monthly data center consumption from the commercial sector data it gets from the Energy Information Administration, and strips hourly data center load from the historical series used to estimate its model. The accepted large load is then added back on top. Without that step, data center growth already embedded in the commercial trend would be added a second time through the adjustment.
Fourteen zones, one driver
The report lists which zones received data center adjustments: AEP, ATSI, APS, BGE, ComEd, Dayton, Duquesne, JCP&L, Met-Ed, PECO, Pepco, PPL, PSE&G and Dominion. Dominion's adjustment also covers a voltage optimization program, and PSE&G's covers port electrification, but the common factor across the list is data centers. Ten-year annualized growth rates for individual zones range from -0.2% to 6.4%, with a median of 1.6%. The spread tells you that the forecast is not a story of uniform growth. It is a story of a handful of zones absorbing most of the increase while many others barely move.
The comparison with the 2025 report is useful. A year ago PJM projected summer peak of 209,923 MW in 2035, a ten-year increase of 55,779 MW, with ten-year annual growth of 3.1%. Net energy was projected at 1,328,045 GWh in 2035. The 2026 report pushes the ten-year summer growth rate to 3.6% and the energy growth rate from 4.8% to 5.3%. So the overall trajectory is steeper, even though the next few years are lower.
Why the near-term cut matters for capacity prices
The capacity market is where the forecast turns into money. PJM's auctions procure capacity three years ahead, and the reliability requirement in each auction is built from the load forecast. Every megawatt of forecast peak that is removed from 2027 or 2028 lowers the amount of capacity PJM must buy for those delivery years. With recent auctions clearing at or near the price cap agreed with Pennsylvania, small changes in the requirement can move clearing prices substantially, because supply in the near term is close to fixed.
That is why the firmness test matters more than its technical tone suggests. A load adjustment that never materializes still raises the procurement target and the price that every customer in the region pays. By requiring a signed service agreement or construction contract before a project can influence the auction, PJM is shifting some of the risk of speculative requests back toward the developers and utilities that submit them. Non-firm projects still appear in the planning picture, but they no longer set the bill.
The longer-term increase works differently. Load in the late 2030s does not affect any auction that has been scheduled, but it does feed transmission planning and the long-term regional plans that PJM must now prepare under federal rules. A higher 2040 forecast strengthens the case for large backbone projects. It also raises the question of whether generation can be built fast enough. A forecast that adds tens of gigawatts of peak in a decade implies a build rate PJM has not achieved in recent years, given the backlog in its interconnection process and the time it takes to permit and build new plants.
The winter problem
The winter numbers deserve as much attention as the summer ones. PJM projects winter peak growing 4.0% a year over the next decade, faster than summer, reaching 204,650 MW in 2035/2036, a ten-year increase of 66,980 MW. Data centers run around the clock and through every season, so they add more or less the same load in January as in July. Heating electrification adds load in winter as well. Meanwhile the resources that matter most in a cold snap, dispatchable thermal plants with secure fuel, are the ones that have been retiring. A system that has historically planned for summer peaks is moving toward one where the hardest hours fall in winter, and where gas pipeline constraints and generator outages in extreme cold become the main reliability risk.
What to watch
Three things will show whether the new method holds up. The first is how many non-firm projects convert to firm over the next year, because conversions will drive the 2027 forecast. The second is whether the 70% utilization assumption matches what new campuses actually draw once they are running; if AI training clusters operate at higher load factors than older cloud facilities, the assumption will prove conservative. The third is the political response. State regulators and consumer advocates have pushed for tighter screening of data center requests because they flow straight into capacity costs, and the January forecast is the first to apply that discipline at scale.
The overall message is that PJM still expects data centers to reshape its system, but it now wants contractual evidence before customers pay for the capacity to serve them. That is a reasonable line to draw. The harder task, building enough generation and transmission to meet the long-run curve, is untouched by any change in forecasting method.