Today, five of the six regional grid operators must answer the Federal Energy Regulatory Commission’s June show-cause orders on large loads, with Southwest Power Pool to follow on November 20. The filings will set the terms on which data centers connect to the transmission system and who pays for the upgrades they trigger. They land with a number hanging over them: utility forecasts filed with FERC put data center load about 40 percent above credible industry projections through 2030.
That figure comes from Grid Strategies, whose annual load-growth report is the industry’s closest thing to a shared scoreboard. Its November 2025 report, Power Demand Forecasts Revised Up for Third Year Running, Led by Data Centers, put five-year peak demand growth at 166 gigawatts by 2030 — more than four times its 2023 projection of 38 gigawatts. About 90 gigawatts of that growth, roughly 55 percent, is tied to data centers. But the report’s own authors did not believe their headline number. “We think the data center number is too high,” John Wilson, the firm’s specialist in planning and utility regulation, told Latitude Media when the report was published. Data center market analysts, the report said, indicate growth is unlikely to require much more than 65 gigawatts through 2030 — suggesting the FERC-filed forecasts overstate data center load growth by 40 percent.
The arithmetic is worth stating plainly, because the gap is not a rounding dispute. Ninety gigawatts of claimed data center load against roughly 65 gigawatts that independent trackers expect to materialize leaves about 25 gigawatts unaccounted for — Wilson’s own estimate of the overstatement. The date matters: this is a November 2025 finding, not a new November 2026 analysis. What is new is the context — a year of dockets, auctions, and filings that turned a forecasting caveat into the central cost-allocation question in grid planning.
How a forecast becomes 25 gigawatts too big
The mechanism is not mysterious, and Grid Strategies named it when the report came out: speculative and duplicate interconnection requests. A developer with one project shops the same megawatts to several utilities to keep its siting options open. Each utility folds the request into its load forecast, so the same campus is counted two or three times in the national total — and none of the forecasts knows about the others. The FERC filings Grid Strategies aggregated are an accumulation of reservations, not a count of commitments.
That is the same mechanism this publication documented in PJM last month, at the scale of a single region. PJM’s independent market monitor, Monitoring Analytics, calculated in a September 25 filing 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. The monitor’s arithmetic cut PJM’s claimed 6,831-megawatt capacity shortfall for 2028/29 to about 3,055 megawatts. Different measurements, same direction. The queue holds megawatts that will never be built, and the forecast carries them as if they will.
Two cautions belong on the national figure. First, the 40 percent compares two different kinds of estimates — utility filings against market analysts’ project trackers — not a measurement of projects that have actually failed. Some of the gap is real overstatement; some of it is the difference between counting reservations and counting commitments. Second, even the lower number is staggering. If the truth is, as Wilson put it, somewhere in the middle — about 25 gigawatts less than utilities declare — the system still faces roughly 140 gigawatts of peak growth in five years. The load wave is real; it is smaller than the number being planned to, and planning to the wrong number has a price.
What the gap does to transmission planning
That price shows up first in transmission planning. A load forecast is the input to the regional transmission plans — PJM’s Regional Transmission Expansion Plan, MISO’s Transmission Expansion Plan, and their equivalents — that decide which lines get built, where, and at what cost. Transmission is lumpy — a line is built or it is not — but a plan sized on inflated peaks systematically overbuilds where speculative load clustered, and overbuilding is not free. The steel and wire go into the rate base, and the rate base is paid for by everyone on the system.
Berkeley Lab’s Queued Up 2026 edition puts the median time from interconnection request to operation at 61 months for power plants built in 2025, so a planner who waits for certainty will always be late — and one who builds on the forecast will sometimes build for load that never arrives. The industry’s answer has traditionally been to plan on the forecast and socialize the risk. The forecast-credibility gap is an argument that the forecast is no longer a safe thing to socialize.
Texas is the sharpest example of what happens when the queue stops being credible as a work plan. Governor Greg Abbott ordered an audit of every data center in ERCOT’s interconnection queue in August — a queue reported at 474 gigawatts, more than five times the state’s record peak — with no project moving forward until the audit is complete. A queue that size cannot be a forecast; it is a list of options, and planning as if every option will be exercised is how a system ends up building for ghosts.
The filings due today, read as a forecasting test
That is the test to apply to the filings due today. The November 2 edition of this publication laid out the five questions each filing must answer — application and study, cost shifting and transparency, co-location, flexible service, and proximate generation. The forecast-credibility gap cuts across all five, but it bites hardest on the second: what prevents upgrade costs for one large load from landing on everyone else.
The strongest protections on offer are the ones that make a reservation expensive to hold and verifiable when claimed. PJM’s Large Load Registry, which FERC accepted in October and which took effect October 12, requires load-serving entities to report location, in-service dates, peak load, ramp, telemetry, and backup generation for every large load. It is, as this desk argued last month, the beginning of a cure — but only if the forecast uses the data it collects. ISO New England’s proposal, filed November 13, pairs three new service options with a $50,000 application fee, a $150,000 study deposit, and a readiness deposit of $10,000 per megawatt. MISO’s 120-day Large Load Addition Resource Study, filed October 2, studies a load above 200 megawatts together with new generation in the same zone, tying the two together for 15 years. Each is a filter for separating a project with site control, financing, and a construction schedule from a reservation filed to hold a place in line.
The cost-recovery agreement is the backstop for when the filter fails. FERC has pressed every region for a standard, pro forma agreement — operator, transmission owner, and the transmission customer taking service for the large load — backed by financial security, obliging the load to cover the costs incurred to serve it even if the project is canceled or runs below its anticipated demand. Without one, an upgrade built for a load that never arrives sits in the rate base and is paid for by everyone else. MISO has said its cost-shifting protections are still to come in its November 16 response — the gap to watch in today’s filings. Indiana’s Strategic Energy Growth Task Force recommended in July that large-load customers pay 100 percent of project costs; TVA’s tariff, effective October 1, charges new and expanded data center load above five megawatts roughly $1.5 million per megawatt to reserve capacity. The direction is unmistakable: a reservation must carry a cost or a commitment, and the cost must land on the customer that triggers it.
The other side of the gap: underbuilding
The other side deserves a fair hearing: discounting forecasts too aggressively carries its own risk. The independent trackers that Grid Strategies cites are not neutral arbiters. They count projects that are financed and under construction, and will miss load that is real but early. The disagreement among forecasters is itself a measure of the uncertainty: Berkeley Lab’s June 2026 update puts United States data center use at 192 terawatt-hours in 2024, with a reference case of 649 terawatt-hours by 2030, inside a range of 521 to 843. The International Energy Agency’s base case puts United States data center use at about 426 terawatt-hours. The two most cited public estimates for the same country and the same year differ by more than 200 terawatt-hours. A planner who discounts to the lowest estimate and is wrong will have underbuilt — and underbuilding shows up as scarcity pricing and load that cannot connect.
The North American Electric Reliability Corporation’s latest long-term assessment found 13 of 23 assessment areas at elevated or high risk of supply shortfalls within five years, with data centers a main driver. Grid Strategies has argued that assessment overstates the risk by overestimating how much data center load will connect — the same dispute, in reliability language. PJM’s January 2026 long-term forecast showed the correction beginning: it trimmed the summer 2028 peak by 4.4 gigawatts, or 2.6 percent, citing stricter vetting of planned data centers. The question is whether the correction is keeping pace with the error.
Congress prices the gap
The forecast-credibility gap has reached Congress, with a dollar figure attached. Grid Strategies estimated in a white paper published this month that the Bipartisan American Affordability and Jobs Act, Senate bill 5653, would generate about $5.5 billion a year in transmission charges from data centers, roughly $55 billion over ten years — about $35 billion for existing transmission costs and $20 billion incremental — against roughly $2 billion a year that data centers currently pay. The estimate rests on 40.3 gigawatts of new data center capacity outside Texas, from Enverus’s third-quarter high-confidence tracker, all assumed to take firm service. About 82 percent would land in PJM and MISO, at $2.31 billion and $2.25 billion a year respectively. A separate provision, the Ratepayer Protection Floor, could recover another $136 billion in generation costs over ten years if states fully implement it — potential exposure, not a forecast. The bill has not passed, and its prospects are uncertain. But the $55 billion rests on 40.3 gigawatts of high-confidence capacity — a number that already discounts the queue. It uses the tracker, not the forecast.
What the gap costs if nothing changes
The clearest illustration of the cost is already on the docket in PJM. On October 29, PJM filed its revised Reliability Backstop Procurement — the one-time, out-of-market purchase of 6,831 megawatts of new capacity that FERC accepted in September and suspended to February 28, 2027. The revised filing keeps the 6,831-megawatt target, subject to recalculation as executed bilateral contracts are netted out, at offers capped at $555 per megawatt-day for up to 15 delivery years, with costs assigned to zones by forecast large-load growth. At the offer cap, 6,831 megawatts costs roughly $1.38 billion per delivery year; the monitor’s corrected shortfall of about 3,055 megawatts would cost about $619 million a year on the same arithmetic. The load behind those contracts is measured in summer 2028 — after the decision, and after the bills start. That is the gap converted directly into money: a procurement sized on a forecast, billed on a forecast, verified only when the load arrives or does not.
A line approved on a forecast that overstates load by 40 percent is a line whose cost is recovered from customers who did not cause it. The gap is a two-way ratchet: overbuild now, underbuild later.
What to watch
Four things will tell us whether today’s filings close the gap or paper over it. The first is whether any filing commits the operator to discounting its load forecast by project milestones — site control, financing, construction progress — rather than treating every request as firm. A registry that collects data but does not feed the forecast is a filing cabinet, not a fix. The second is the cost-recovery language, above all in MISO’s response, where the protections are still outstanding. The third is the 30-day answer period, which will test the weakest joints — the $250-per-megawatt-hour re-dispatch trigger in New England, interim-service duration in California and SPP, and the netting rules in New York. The fourth is the Texas audit: the first state-level count of how much of a 474-gigawatt queue is real.
The deeper question is the one the monitor’s filing posed in PJM and Grid Strategies’ report posed nationally: who has to prove the load is real? For two years the burden ran the other way: the forecast was taken as given, and skeptics had to prove it wrong. The 40 percent gap, PJM’s 4,016.9 megawatts, and the 474-gigawatt queue in Texas have shifted that burden. From here on, anyone who wants to procure capacity, build transmission, or raise rates on a data center forecast will have to answer it with evidence rather than a reservation: how much of that load is real?