PJM’s independent market monitor has put the most precise price tag yet on the data center boom — and then supplied the reason to doubt the forecast behind it. In the first half of 2026, the inclusion of data center load in PJM’s capacity market added $11.11 per megawatt-hour, or 9.7 percent, to the total cost of wholesale power in the 13-state region, according to Monitoring Analytics’ quarterly State of the Market report. Across the last four capacity auctions, the monitor calculates, data center load produced a cumulative, irreversible increase in capacity costs of $29.4 billion. Then, in a September filing at the Federal Energy Regulatory Commission, the same monitor calculated that 4,016.9 MW of forecast data center load — 42 percent of the growth priced into the two most recent auctions — had already been canceled or delayed.
The monitor’s arithmetic
Start with the headline figures, which come from the monitor’s own publications rather than anyone’s summary of them. In the first six months of 2026, the total cost of wholesale power in PJM rose 50.3 percent year over year, from $76.16 to $114.50 per MWh, according to the monitor’s press release accompanying its January-through-June State of the Market report. Energy accounted for 63.5 percent of that total, transmission 17.4 percent, and capacity 17.1 percent. The capacity component alone jumped 207.1 percent, from $6.39 to $19.61 per MWh.
The data center share of that increase is isolated in the report: including data center load in the capacity market raised the total wholesale cost by $11.11 per MWh, from $103.40 to $114.50. That figure covers the capacity market only; it excludes any data center effect on energy prices and transmission costs. A seven-month update, presented to PJM’s Members Committee on August 24 and reported by Utility Dive, put the data center capacity-market adder at $10.48 per MWh, or 9 percent of a $116.53 per MWh total wholesale cost through July.
The monitor’s verdict on the capacity auctions was blunt. It concluded that the results of the capacity auctions for the 2025/2026, 2026/2027, and 2027/2028 delivery years were not competitive, primarily because of the inclusion of forecast demand for data centers. Its recommendation: exclude forecast data center load from the capacity market, and require data centers to bring their own new generation.
How a forecast becomes a bill
To understand how a load forecast turns into money on a retail bill, recall the mechanics from this desk’s earlier explainers. Load-serving entities must buy capacity equal to their customers’ peak demand plus a reserve margin, through PJM’s Reliability Pricing Model auctions, typically three years ahead of the delivery year. The auction’s demand curve — the Variable Resource Requirement curve — sets the price. When the system is short, the price rises to the curve’s cap. That is what has happened three auctions running. The 2028/2029 Base Residual Auction, held in July 2026, cleared at the FERC-approved cap of $325 per MW-day and still procured only 138,318 MW of unforced capacity — 6,831 MW short of the reliability requirement, for a total commitment of $16.4 billion, according to PJM. The prior auction, for 2027/2028, was 6,516.6 MW short of its required reserve level of 23,494.0 MW.
The price cap itself is a political artifact, not a market outcome. Under an agreement between Pennsylvania Governor Josh Shapiro and PJM, incorporated into a FERC filing, the 2026/2027 and 2027/2028 auctions ran on a restricted demand curve with a defined maximum and minimum price. The monitor calculates that the agreement cut 2026/2027 auction revenues by about $3.17 billion and 2027/2028 revenues by about $9.91 billion, relative to the unrestricted curve. Without the collar, the bills would have been larger still.
From the auction, the costs flow downhill. PJM collects capacity revenue from load-serving entities, which recover it from customers through the allocation method each uses for capacity costs — allocated on customers’ contributions to peak load, not spread evenly per kilowatt-hour. That is the first reason the wholesale-to-retail translation is uneven: a customer whose usage coincides with system peaks carries more of the capacity cost than one who uses the same energy off-peak. The second reason is timing. As this desk has explained, hedges, rate cases, and adjustment clauses delay and smooth wholesale costs before they reach retail bills — which means the capacity price shock of the last three auctions is still working its way through retail rates.
For scale: the monitor’s $11.11 per MWh wholesale adder works out to roughly $11 per 1,000 kWh at the wholesale level. What any given household actually sees depends on its peak contribution, its supplier’s hedging, and its state’s retail structure. The wholesale number is the starting point of the translation, not the end of it.
Four auctions, $29.4 billion
The cumulative figure requires care, because it is easy to misread. The monitor’s State of the Market report puts the increase in capacity market revenues attributable to data center load, actual and forecast, at $23.1 billion across the 2025/2026, 2026/2027, and 2027/2028 auctions. With the completion of the 2028/2029 auction, the report says, the total reached $29.4 billion — an irreversible increase in costs.
Irreversible is the operative word. Each auction buys capacity for a single delivery year. Once an auction clears, the commitments are struck and the payments flow, whether or not the load that justified them ever materializes. The $29.4 billion is the sum of the incremental revenues across four one-year delivery years, not an annual run rate.
The monitor is also explicit that the problem is not generic load growth. In its words, the current conditions are almost entirely the result of large load additions from data centers, both actual historical and forecast, and it is misleading to describe the auction results as a simple reflection of supply and demand. Its own data show the scale of the shift: data center load included in the capacity auctions rose from 7,927 MW in the 2025/2026 auction to 11,993 MW, then 17,071 MW, then 21,474 MW in the 2028/2029 auction — most of it forecast rather than connected.
The other number: 4,017 MW that may never arrive
Now the second number. In a September 25 answer filed in the docket over PJM’s proposed Reliability Backstop Procurement, the monitor calculated that 4,016.9 MW of forecast data center load had been canceled or delayed as of September 21, 2026. That is 42 percent of the 9,481 MW increase in data center load between the 2026/2027 and 2028/2029 auctions. Dominion’s territory accounted for 51 percent of the canceled or delayed load and AEP-Ohio’s for 29 percent.
The calculation’s provenance matters. The monitor based it on project-status tracking by Industrial Info Resources, corroborated by publicly available data, and counted only projects originally scheduled to be in service by 2029. It also disclosed a significant limitation: it does not have access to the project-level data behind the large load adjustments that electric distribution companies submit to PJM. Its count is therefore an independent check built from outside the forecast process.
The underlying forecast is also older than the auction it priced. The data center load forecast for the 2028/2029 auction rested on large load adjustments PJM received before September 5, 2025 — more than ten months before the auction cleared. PJM’s backstop filing contains no provision to revise the load forecast or the procurement target before the procurement runs. The forecast is a snapshot that ages while the market keeps paying on it.
What the smaller gap would mean
The monitor translated the canceled and delayed load into the auction’s own terms. Start from the 2028/2029 auction’s actual result: a 6,831.3 MW shortfall against the reliability requirement. If data center load had stayed at the level forecast for the 2026/2027 auction — 11,993 MW instead of 21,474 MW — PJM would have had a 1,096.5 MW surplus rather than a shortfall. If only half the forecast growth had materialized, the shortfall would have been 2,374.7 MW. And in the scenario that treats the 4,016.9 MW of canceled or delayed load as never arriving by the 2028/2029 delivery year, the shortfall falls to 3,055.0 MW — a 55 percent reduction.
Two caveats belong on that scenario. First, delayed is not canceled: some of the 4,016.9 MW may arrive after the 2028/2029 delivery year, in which case the capacity is needed later rather than never. Second, the scenario assumes the identified load does not come online by the delivery year at all — a clean assumption for arithmetic, not a prediction about any specific project. What the table demonstrates is not that the shortfall is 3,055 MW. It is that the shortfall is highly sensitive to a forecast that has already moved by more than 4,000 MW.
The money follows the same sensitivity. PJM’s backstop proposal would procure the 6,831.3 MW gap at a maximum weighted-average levelized offer of $555 per MW-day, for contracts of up to 15 delivery years. At that cap, the arithmetic — our calculation, not the monitor’s — runs to about $1.38 billion per delivery year for the full gap, versus about $620 million for the 3,055 MW gap in the monitor’s third scenario. Those are illustrative figures, not expected charges: the backstop is suspended, and the cap is a ceiling on levelized offers, not a clearing price. But they show why the forecast dispute is worth billions.
Why the market prices the forecast, not just the load
This is the structural point, and the one that connects the monitor’s two numbers. A capacity market procures for the future on the basis of a forecast, and it does not true up when the forecast is wrong. When PJM includes speculative, developer-supplied load in the demand curve three years ahead, the auction clears at a higher price, and every load-serving entity pays that price on its full obligation — including the share attributable to load that never connects. There is no refund mechanism for a forecast that shrinks. The $29.4 billion is, in that sense, the cost of pricing the forecast rather than the load.
That is also why the backstop procurement is so consequential. It would convert a one-year auction shortfall into 15-year commitments, locking in payments for capacity sized to a gap that the monitor says is already more than half overstated. FERC, on September 29, accepted the backstop filing but suspended it until February 28, 2027, and set a paper hearing on cost allocation, collateral requirements, and transmission-owner exit rules. In a separate order on October 9, FERC accepted PJM’s definition of a Large Load — end-use load of 50 MW or more at a single site — and a Large Load Registry effective October 12, but rejected the companion Interim Resource Adequacy Service proposal, finding that it intruded on retail rate regulation reserved to the states.
The monitor’s alternative, laid out across its protest and answer, is to stop pricing the forecast altogether: remove data center load not interconnected by June 1, 2026, from the capacity auctions; require new data center loads to hold 15-year bilateral contracts with new generation that permit no cost or risk shifting to other customers; and impose curtailment obligations on wholesale service to data centers during a transition period. That approach echoes the White House’s Ratepayer Protection Pledge, under which hyperscalers committed to build, bring, or buy the generation their data centers need. FERC has not adopted the monitor’s proposal.
What it means for the retail bill
Return to the question this desk started with: how does wholesale reach the retail bill? The monitor’s arithmetic supplies the wholesale half of the answer, and it is unusually clean. Data center load, through the capacity market alone, added about a tenth to the wholesale cost of power in the first half of 2026. Capacity is procured on forecast, paid on commitment, and allocated to customers on peak-load contribution. The retail half is slower and murkier: hedging, rate cases, and state allocation rules delay and redistribute the cost, which is why two households in different states — or on different rate plans in the same state — will experience the same wholesale shock very differently.
The policy question is whether that delay should be spent fixing the forecast. If 42 percent of the priced growth has already been canceled or delayed, and the procurement machinery has no mechanism to revise its target, every additional auction run on the stale forecast compounds the overcharge. The December auction for the 2029/2030 delivery year — in which the monitor says 20,539.7 MW of data center load remains included — the FERC paper hearing, and the new Large Load Registry are the three places where the forecast could get corrected.
The market, as currently designed, will keep pricing the forecast until someone changes the design. The monitor has now shown, in its own arithmetic, both what the forecast has cost and how much of it has already evaporated. The $29.4 billion is the price of the old forecast. The 4,017 MW is the measure of its error. What happens next depends on whether the institutions that run the market update the forecast before they charge for it again.