On Monday, November 2, the Load Analysis Subcommittee heard the number that reframes everything PJM is about to buy. Charles River Associates, the independent consultant PJM hired to build an hourly data-center load forecast for 2027 through 2047, presented its final results: PJM data-center demand reaches 30 GW by 2030 and 82 GW by 2047, concentrated in Dominion and AEP, and the gap between what developers want to build and what the system can power at 1-in-10 reliability reaches 11.5 GW by 2030. Absent faster supply additions, CRA said, PJM could face a shortfall of roughly that size within four years — not in some distant planning horizon, but inside the delivery years its auctions are already pricing.
Five days earlier, a different set of numbers had pointed the other way. In an October 28 note, Jefferies analysts reported that PJM utilities including PPL Electric, FirstEnergy, and AES Ohio had lowered their overall near-term large-load forecasts by 6.5 GW, citing recent filings to the same Load Analysis Subcommittee. That reduction, the analysts wrote, could cut PJM’s 2027 load forecast enough to shrink the reliability backstop procurement — conceived to fill a 6.8 GW hole — to roughly 630 MW. Two presentations to the same committee, one week apart, and the procurement at the center of PJM’s reliability plan is either barely dented or nearly an order of magnitude too large.
This is the news, and it is not the backstop refile itself. That filing, submitted October 29 and covered here last week, is a revised proposal awaiting a Federal Energy Regulatory Commission decision; nothing in it has been accepted, and no bid has been taken. What changed this week is the evidentiary ground underneath it. PJM is now running two different answers to the same question — how much large load is actually coming — into the same December auction for the 2029/30 delivery year and the same March 2027 backstop launch. Capacity procurement is only as good as the load forecast behind it, and PJM’s forecast has just been shown, in public, to be two forecasts at once.
What CRA actually found
CRA’s work deserves to be read on its own terms before it is drafted into anyone’s argument, because its headline numbers are more careful than they look in summary. The forecast reconciles two methods: a top-down estimate of PJM’s share of national AI and non-AI demand, tested against semiconductor, infrastructure, and supply-chain constraints, and a bottom-up, site-level pipeline scored on contractual stage, permitting status, and moratorium exposure. The headline trajectory — 30 GW by 2030, 82 GW by 2047 — is the “soft supply constraint” case: the reconciled demand forecast, which CRA says assumes projects deploy beyond what PJM can support at 1-in-10 reliability. The “hard supply constraint” case is the power-constrained base forecast, reflecting planned generation additions, current regulations, behind-the-meter offsets, and the 1-in-10 standard itself. The 11.5 GW figure is the distance between those two views in 2030: the load that wants to arrive, minus the load the system as currently supplied can serve without breaching the reliability criterion.
The geography matters as much as the total. Dominion and AEP comprise the largest share of demand, given pipeline size and existing deployed data centers, with PPL and ComEd as important secondary markets growing substantially in the 2030s. Near-term growth concentrates in Dominion and AEP precisely because so many projects there are already ramping, and zones with behind-the-meter projects in the pipeline hold an advantage, since that load can circumvent the headroom constraint.
CRA is equally explicit about the uncertainty wrapped around the headline. Under its scenarios, data-center demand under the soft constraint varies by 6.6 GW in 2030 and 29.3 GW in 2047. A low-adoption case converges to 19.4 GW below the base case by 2047; a constrained-power case cuts 2047 demand by 29.2 GW, to 52.2 GW. Realized load varies most in Dominion, AEP, and ComEd, with uncertainty of up to 10.4 GW in Dominion alone by 2047. The forecast’s own authors, in other words, put a band around their number that is wider than the entire backstop procurement. That is not a flaw in the work. It is the honest shape of the problem, and it is exactly what a capacity market built on point forecasts struggles to absorb.
One more CRA finding belongs in any serious reading: the pipeline is being screened, not just counted. CRA reviewed distribution company and load-serving-entity submissions and adjusted for materialization risk, removing projects that were not fully permitted, lacked a firm contract, or sat in areas under active local moratoriums. The moratorium count is striking on its own — 294 local moratoriums and bans passed between January and July 2026 — and it cuts both ways, explaining why some announced load will never arrive and why the load that does arrive concentrates where approvals are still obtainable.
Why the near-term forecast fell
The Jefferies note is not a contradiction of CRA so much as a photograph of a different horizon. CRA’s 30 GW is a 2030 demand trajectory; the 6.5 GW reduction is in utilities’ near-term large-load forecasts, the figures that feed the next delivery years and therefore the backstop’s recalculated target. Both can be true at once: the long-run pipeline can keep growing while the near-term forecast falls, because the near-term forecast is where cancellations and delays land first.
That mechanism already has a documented count. PJM’s Independent Market Monitor calculated that 4,017 MW — 42 percent — of the forecast increase in data-center load between the 2026/27 and 2028/29 delivery years had been canceled or delayed as of September 21. The monitor’s finding, filed in the backstop proceedings, is the single most important piece of evidence for anyone arguing the backstop target should shrink: nearly half the growth the July auction priced in had already evaporated before the refiled proposal was written. The utilities’ lowered forecasts that Jefferies cites are, in part, the same phenomenon reaching the official forecast through Load Analysis Subcommittee filings rather than through the monitor’s audit.
There is also a structural reason the near-term number moves faster than the long-run one. PJM’s forecast process has always depended on load-serving entities submitting large-load adjustments, with contract status distinguishing firm commitments from speculative requests. The July auction’s reliability requirement was built on adjustments that, as FERC’s September order noted, may never have fully accounted for some large loads — while simultaneously counting others that would not materialize. The registry FERC accepted in October, with tariff-backed reporting of in-service dates, ramp schedules, and delivery points for every load of 50 MW or more, is designed to firm up exactly these inputs. But the registry tool is not expected until February 2027, with aggregate public data following in March — after the December auction closes and essentially concurrent with the backstop’s launch. December and March will therefore be run on the old visibility, however improved the next cycle’s inputs may be.
The collision, in calendar order
Start with December. The auction for the 2029/30 delivery year closes December 15, capped at $337.83/MW-day, and it will clear against a demand curve fed by the revised forecasts — the lowered near-term numbers if they are adopted, the older ones if they are not. Generators bidding into December must decide whether to commit capacity for one year at a capped price or hold out for the backstop’s 15-year contracts at up to $555/MW-day, a channel the refiled proposal makes more credible by giving it a defined March window and a published results date. Thin new supply in December may therefore reflect rational waiting for March as much as physical absence.
Then March. Under the refiled schedule, the backstop bid window closes March 28, the final target is set April 28, and results are posted May 10, for capacity serving a delivery year beginning June 1, 2028. PJM has stated the target will be recalculated with updated information and reduced by megawatts procured through qualifying bilateral contracts. That recalculation clause, which this desk flagged last week as the filing’s most consequential sentence, is now the whole ballgame. If the lowered near-term forecasts hold and bilateral deals place megawatts privately, the target could land near the roughly 630 MW Jefferies sketches — a procurement an order of magnitude smaller than the 6,831 MW shortfall the July auction measured. If the updated forecasts rebound, or if bilateral placements stall, the target could remain near its original size, and the full weight of a near-$555 procurement would fall on the growth-based cost allocation the amended filing creates.
Note what that means for the cost-allocation fight that dominated the refile. The amended formula assigns backstop costs to zones based on firm large-load growth between the actual 2026 summer peak and the 2028/29 forecast. That formula is only as fair as the growth figures fed into it, and those figures are now visibly moving — down 6.5 GW in the near term by one account, up to 82 GW by 2047 by another. Zones will argue about measurement, not method, exactly as the growth-based design predicts — except now the measurement argument has two authoritative-looking exhibits pointing in opposite directions, and FERC has not yet ruled on which one the tariff should trust.
What getting it wrong costs
Aurora Energy Research, in work presented to PJM’s market-design workshops and discussed again at its Energy Transition Forum last week, has tried to put a price on this exact failure mode. Aurora’s North America managing director, Oliver Kerr, put the cost of the national forecasting gap at $79 billion to $624 billion in extra capital expenditure — before transmission and distribution costs — according to Latitude Media’s account of the forum. The range is enormous, and that is the point: it spans the difference between forecasts that overshoot modestly and ones that overshoot badly, across a system whose peak demand sits just below 760 GW. Aurora’s own framing, reported at the same forum, is that grid operators are expecting roughly 63 GW more demand by 2030 than Aurora’s baseline projections. These are Aurora’s numbers, not PJM’s, and they describe a national gap, not a PJM-only one. But they quantify what PJM’s two-forecast week makes concrete: when the forecast is wrong in either direction, someone pays — ratepayers for capacity built for load that never arrives, or everyone for reliability that falls short when load arrives faster than supply.
The design question underneath
Strip away the docket numbers and the design question is stark. A capacity market exists to convert a forecast into a purchase: it takes a peak-load projection, adds a reserve margin calibrated to 1-in-10 reliability, and buys the difference. Every element of that chain now carries a visible error bar. The peak-load projection spans 6.6 GW of scenario uncertainty in 2030 alone. The reserve margin depends on accreditation factors that will be updated before March. And the backstop’s target is explicitly provisional, to be recalculated weeks before bids are taken.
None of this makes the backstop unnecessary. CRA’s hard-constraint case is a genuine reliability warning: if demand follows the soft trajectory and supply follows the status quo, 11.5 GW of load cannot be served at 1-in-10 reliability by 2030, and the July auction’s 525 MW of new and upgraded supply is nowhere near the scale of that gap. The argument for a backstop was never that the forecast is precise. It is that the cost of being short — load shedding on a hot day in a system serving 67 million people — is worse than the cost of being long. But that argument has a limit, and the Jefferies arithmetic states it plainly: a procurement designed for 6.8 GW that buys 6.8 GW against a forecast that has already fallen 6.5 GW is not insurance. It is overbuilding, assigned by zone, locked in for 15 years, at up to $555/MW-day.
The honest resolution is not to pick a forecast. It is to design a procurement that survives being wrong. That means, at minimum, treating the April 28 final target as a genuine decision point rather than a formality — with the updated forecast, the bilateral placements, and the monitor’s cancellation accounting all on the record before it is set. It means asking, in the December auction’s aftermath, whether the two-tier price structure — $337.83 in the regular auction, $555 in the backstop — is still a transitional design or has become the market. And it means recognizing that the registry, arriving in February, is the first instrument PJM has ever had that could narrow the forecast error at its source, by replacing asserted load with reported load. The backstop buys time. The registry, if its data are complete and its reporting enforced, buys accuracy. PJM needs the second far more than the first, and the calendar gives it the first four months sooner.
What to watch, in order: whether FERC accepts the amended backstop proposal or holds the companion paper hearing open; what demand curve the December auction actually clears against when it closes December 15, and how much new supply shows up versus waits for March; what the registry’s February tool reveals about firm versus speculative load; and what number PJM posts on April 28 as the backstop’s final target. The gap between 630 MW and 6,831 MW is not a rounding error. It is the distance between two forecasts of the same system, and whichever number April produces will tell us which forecast PJM believed when it mattered.
Sources: Charles River Associates, “PJM Data Center Capacity & Peak Demand Forecast,” presentation to PJM Load Analysis Subcommittee (Nov. 2, 2026); Utility Dive, “PJM’s revised data center backstop procurement plan reduces utility risks: ICF” (Nov. 3, 2026), reporting Jefferies’ Oct. 28 note; Latitude Media, “The US power market’s forecasting problem comes at a cost to ratepayers” (Nov. 3, 2026), reporting Aurora Energy Research remarks at the Aurora Energy Transition Forum; Aurora Energy Research, “Pathways for PJM Resource Adequacy Market Evolution” (Aug./Oct. 2026); PJM, 2028/2029 Base Residual Auction results (July 14, 2026); FERC, 196 FERC ¶ 61,245, Dockets ER26-3380-000 and EL26-108-000 (Sept. 29, 2026); PJM Independent Market Monitor, analysis of canceled/delayed data-center load (Sept. 21, 2026); prior Fracmoo coverage: “PJM Filed Its Backstop Fix. The Question Now Is Whether FERC Buys the New Way of Splitting the Bill.” (Oct. 31, 2026) and “The Registry Is the Only Part of PJM’s Large-Load Plan That Survived. This Week It Gets Its Final Form.” (Nov. 2, 2026).