For two years, the defining number in American grid planning has been the coming data center wave — tens of gigawatts of new load, arriving faster than power plants can be built. Last month, PJM’s independent market monitor put a number on the other side of that ledger: the load that was forecast, planned for, then never showed up. In a September 25 filing with the Federal Energy Regulatory Commission, Monitoring Analytics calculated that 4,016.9 megawatts of forecast data center load — 42 percent of the forecasted increase between the 2026/27 and 2028/29 delivery years — had been canceled or delayed as of September 21, 2026.
That is not a rounding error. It is roughly four large nuclear reactors’ worth of demand, erased after PJM had built its reliability case around it. And it lands in the middle of the most consequential capacity fight in PJM’s history, over a one-time emergency procurement meant to close a 6,831-megawatt shortfall for the 2028/29 delivery year. The monitor’s estimate is that correcting for the canceled and delayed load would cut that shortfall to about 3,055 megawatts — a 55 percent reduction. The gap would still be real — just less than half the size PJM claimed when seeking approval for an extraordinary buy.
Two cautions belong up front, because they shape everything that follows. First, these are the monitor’s figures, not PJM’s. Monitoring Analytics is an advocate in these dockets, arguing against PJM’s procurement, and PJM has stood by its forecast of steep continued growth. Second, the monitor’s headline number combines two different things — load that was canceled and load that was merely delayed — and the filing, as described in coverage of it, does not break out the split. A delayed data center is not a dead one; some of that load may still arrive, on a later schedule. What the number measures is not the death of data center demand. It is the unreliability of the forecast at the moment it was used to justify emergency spending.
The arithmetic of a smaller gap
Start with the shortfall itself. PJM’s 2028/29 Base Residual Auction, announced July 14, 2026, procured 138,318 megawatts of unforced capacity, with another 10,864 megawatts from Fixed Resource Requirement resources. The auction cleared at $325 per megawatt-day, the FERC-approved price cap, for about $16.4 billion in total capacity commitments — and still came up 6,831 megawatts short of PJM’s reliability requirement. It was the second consecutive auction in which the entire PJM region missed its reliability requirement; the prior auction, for 2027/28, fell short by 6,623 megawatts. Only 525 megawatts of new generation and uprates cleared.
PJM’s answer to the miss was the Reliability Backstop Procurement: a one-time, out-of-market purchase of 6,831.3 megawatts of new capacity, with offers capped at $555 per megawatt-day on a megawatt-weighted average levelized basis and contracts up to 15 delivery years. At the offer cap, the arithmetic is stark: 6,831.3 megawatts times $555 times 365 days comes to roughly $1.38 billion per delivery year. Apply the same arithmetic to the monitor’s corrected shortfall of 3,055 megawatts and the figure falls to about $619 million per year. That is simple multiplication at the cap, not a filed cost estimate — but it shows why the size of the gap matters so much.
The monitor’s argument is that the gap was inflated by stale information. PJM set the 2028/29 reliability requirement using a load forecast, and between then and September, a large chunk of the data center load in that forecast was canceled or delayed. Without an updated forecast, the monitor argues, ratepayers would pay for capacity to serve load that no longer exists on the assumed schedule. The monitor notes that forecast data center load for 2028/29 exceeded 21 gigawatts, nearly three times the shortfall, and argues there would have been no shortfall without it.
How a forecast becomes phantom load
To understand how 4,000 megawatts of load can vanish, return to the mechanics this desk has covered since the beginning: the interconnection queue is not a list of projects that will be built — only of projects that asked for a place in line. Developers routinely file multiple requests for alternative sites, some purely speculative, filed to hold a position while the developer decides whether to proceed. And because every request must be studied against the loads ahead of it, a queue full of speculative projects distorts everything downstream, including the load forecast.
That is the phantom-load problem in its purest form. A forecast built on queue positions treats a reservation as a commitment. But a reservation, as TVA’s new tariff implicitly recognized, is cheap to make and cheap to abandon. In PJM, developers could claim thousands of megawatts without posting the financial commitment that would separate serious projects from options. When projects slipped or evaporated, the forecast kept carrying them as if they were still coming.
The monitor’s September filing is, in effect, a reconciliation of the forecast against reality: comparing what PJM projected with what projects actually did. The gap — 4,016.9 megawatts — is the first hard, public measurement of how much of PJM’s projected data center growth was never real. The monitor’s own table of measured July peak data center load shows real growth: from 5,036.9 megawatts in 2023 to 6,135.6 megawatts in 2024 to 8,167.2 megawatts in 2025, a 62 percent increase in two years. What it says is that the forecast increments — roughly 30 gigawatts of data centers in peak-load growth between 2024 and 2030, and about 70 gigawatts of new large load by 2038 — rest on reservations melting away faster than planning ever accounted for.
The backstop, suspended
The monitor’s numbers arrived in a docket that was already moving against PJM’s plan. PJM filed the backstop tariff on July 31, 2026, in docket ER26-3380. The monitor protested on August 21; its September 25 answer contained the 4,016.9-megawatt calculation. On September 29, FERC issued an order — 196 FERC ¶ 61,245 — accepting the backstop filing but suspending it for five months, to February 28, 2027, and establishing a paper hearing plus a companion Section 206 investigation. The auction that was supposed to open September 30 never opened.
FERC’s objections were not about the existence of a shortfall. They were about the design: cost allocation, transmission-owner exit rules, and collateral requirements for load-serving entities. The commission found PJM had not demonstrated those elements were just and reasonable. The concurrences pointed in the same direction as the monitor’s critique: Commissioner Lindsay See said PJM had not adequately ensured that those driving the need for the procurement would bear its costs, and called for better forecasts of large-load project progress; Commissioner David Rosner discussed assigning costs to the zones experiencing data center growth. The Department of Energy filed a statement of position urging the same cost-allocation reforms.
PJM intends to file a revised backstop proposal by October 29, after a special Members Committee meeting on October 22 to consult members on its substance. The grid operator may also reduce the procurement target to account for new supply expected on the system. Whether the revised filing uses an updated load forecast — one that reckons with the canceled and delayed load — is now the central substantive question, and the paper hearing is where the size of the gap gets litigated.
The registry survives; the curtailment plan does not
A second FERC order, issued October 9 in docket ER26-3515, split the difference on PJM’s other big data center package. FERC accepted PJM’s definition of a Large Load — end-use customer load with a cumulative peak of 50 megawatts or more at a single electrical site, counting all affiliated facilities within a one-mile radius as one site — and approved a Large Load Registry, effective October 12. It rejected the rest: the Interim Resource Adequacy Service, which would have curtailed new large loads without their own new capacity first in emergencies; the Bring Your Own New Capacity rules; and the plan to exclude new large loads from the capacity demand curve.
FERC found the curtailment provisions appeared to place obligations directly on retail customers and to dictate which ones would be curtailed — retail rate regulation, outside FERC’s authority. Commissioner Rosner, concurring, said states must set the curtailment order. The registry requires load-serving entities to report, for every large load they serve, the location, in-service dates, peak load, ramp schedule, telemetry, and backup generation. PJM will post only aggregated data publicly, but it will finally have a current picture of which large loads are real, which are slipping, and which exist only on paper. If the phantom-load problem is fundamentally an information problem — a forecast built on reservations nobody verified — then a registry that tracks actual project status is the beginning of a cure. The one-mile rule closes an obvious dodge: splitting a campus into sub-50-megawatt buildings no longer escapes Large Load status.
The other phantom: load that vanishes in real time
There is a second kind of disappearing load in PJM, worth distinguishing from the forecast kind, because the two problems point in opposite directions. On July 22, 2026, nearly 4,000 megawatts of data center load in northern Virginia unexpectedly disconnected during a transmission disturbance and switched to on-site backup generation. PJM and Dominion operators had to manage the resulting imbalances and spikes in transmission voltage and system frequency. It was the third such event in two years. PJM’s executive director of system operations, David Souder, called that kind of operation unsustainable.
The forecast phantom is load that is counted but never arrives. The disconnection phantom is load that is there and then suddenly isn’t. The first causes over-procurement; the second causes real-time instability. PJM’s response — proposed ride-through standards requiring large loads to stay connected through normally cleared transmission disturbances, with a FERC filing expected in November — is moving on a separate track. But they share a root cause: the grid was designed for predictable loads, and data centers, at scale, are not. Any serious reckoning has to address both.
What the reckoning means for each side of the table
For utilities and grid planners, the monitor’s filing is a warning about forecast discipline: a planning process that treats every queued reservation as a firm commitment will systematically overbuild — and overbuilding is not free. It shows up as higher capacity prices for everyone, and as transmission lines built for campuses that never break ground. The lesson is not to ignore the queue but to discount it: weight reservations by financial commitment, site control, and construction milestones, and update forecasts as projects slip or die. A registry only helps if the forecast actually uses the data it collects.
For developers, the message cuts the other way: the era of the free option is ending from several directions at once. TVA’s tariff, which took effect October 1, charges new and expanded data center load above five megawatts roughly $1.5 million per megawatt to reserve capacity — the tariff version of making the reservation costly. Texas’s Senate Bill 6 requires large loads to demonstrate site control, self-fund upgrades, disclose duplicate requests, and accept curtailment during grid stress. PJM’s rejected curtailment plan would have penalized loads that show up without their own capacity. Even with IRAS rejected, the direction is unmistakable: a reservation must carry a cost or a commitment.
For the hyperscalers and their financiers, the monitor’s number introduces a new diligence question. If 42 percent of forecast growth can evaporate, the queue positions, interconnection agreements, and load letters underwriting a site’s value are less solid than they appeared. It also cuts against the scarcity narrative behind premium pricing for capacity and prime sites. Scarcity is still real — a 3,055-megawatt gap is no surplus — but it is a different market when the gap is half the advertised size.
What to watch
Four things will determine whether this reckoning changes anything durable. The first is PJM’s October 29 refile: whether the revised backstop uses an updated forecast, and what procurement target it proposes. A refile that simply re-argues the 6,831-megawatt gap will invite the monitor to re-argue the 4,016.9 megawatts. The second is the October 22 Members Committee meeting, where PJM will test its revised approach with members.
The third is the registry itself. The first rounds of reported data will either validate the monitor’s critique or show that the September snapshot overstated the erosion. The fourth is the December capacity auction for 2029/30, the first with the registry in place and the backstop suspended. Its clearing price and shortfall, if any, will show whether the market absorbed the lesson.
The deeper question is whether the industry builds a planning system that measures commitments instead of counting reservations. The phantom-load problem was never a mystery: everyone knew the queue held projects that would never be built. What was missing was a number — a credible, dated, filed measure of how much of the forecast was fiction. The monitor has now supplied one. It is contested, combined, and a snapshot rather than a trend. But it changes the burden of proof. From here on, anyone who wants to procure capacity, build transmission, or raise rates on a data center forecast will have to answer a simple question: how much of that load is real?