Data center power and grid interconnection to 2028
What the queues actually say, and why the planners have started discounting them.
One grid operator's public queue holds 1,091 active interconnection requests for 254,591 MW, read on 5 November 2026. Of every request that operator has ever resolved, 80.8% was withdrawn. The same operator has 20,285 MW under construction. Both numbers are true, and only one of them is a forecast.
This is a forecast rather than a market summary: seven structural shifts with the evidence under each, the loop that drives them, five scenarios with odds at December 2027 and December 2028, what commoditises and what keeps its scarcity, and eighteen dated calls a named public source will settle. Every reading carries a source grade. Every judgment carries a confidence. The odds are judgments too, and the page says so.
A queue four fifths of which was withdrawn, and a price at its cap
Active in one operator's queue
1,091 active requests, summer net capacity, counted from the operator's own public feed on 5 Nov 2026.
Of resolved requests, withdrawn
FD estimate: 2,325 withdrawn ÷ 2,878 resolved, over the whole history in that feed.
Capacity price per MW-day, at the cap
The 2028/29 delivery year, cleared 14 Jul 2026 against $28.92 for 2024/25.
The planner's own cut to its 2028 peak
PJM's 2026 forecast against its 2025 forecast for the 2028 auction year, a 2.6% reduction.
By the end of 2028 the queue is no longer the forecast, and the thing being rationed is a firm commitment rather than a request
The grid does not get built faster. What changes is who is allowed to stand in the line, and what they have to post to stay there.
We put the rationed-access market at 41% by December 2028 against 29% a year earlier, and today's open-queue market down from 34% to 13%. The second panel is what actually decides a project's date. Each column sums to 100, and both are Frontier Desk judgments rather than model output.
The financing market
Which shape compute financing is in at each date. Frontier Desk judgments, not model output
| Scenario | Dec 2027 | Dec 2028 |
|---|---|---|
| Open queueToday's shape: anyone can file, most withdraw, and the queue number is quoted as if it were demand | 34% | 13% |
| Rationed accessDeposits, curtailment terms and firm-commitment tests decide who holds a position | 29% | 41% |
| Build-your-ownLarge loads stop waiting and bring generation with them, behind or beside the meter | 18% | 21% |
| Demand disappointsRequests keep rising while firm commitments fall, and the forecasts are cut again | 11% | 17% |
| Reliability eventA shortfall or a curtailment of large load forces an emergency rule change | 8% | 8% |
How a new large GPU purchase is funded
The dominant structure for financings signed in each year. Today the secured facility still leads
| Structure | Dec 2027 | Dec 2028 |
|---|---|---|
| Transmission interconnection dateThe project's date is set by the queue it is standing in | 45% | 27% |
| Turbine or equipment delivery slotThe date is set by a manufacturing slot the owner has already paid to reserve | 24% | 29% |
| On-site or co-located generationThe date is set by what the owner can build on its own site | 18% | 26% |
| A curtailable or flexible-load agreementThe date is set by what the owner will agree to give up when the grid is short | 13% | 18% |
Odds are Frontier Desk judgments at the two horizons, set against the triggers and early markers that decide each one in Scenarios, and the method is in Sources and method.
Bottom line
Eight forward-looking judgments, each with the evidence under it and the confidence we hold it at.
- 01The interconnection queue is roughly five times larger than the capacity it will produce, and the operator's own record says so.Reading MISO's public project feed on 5 November 2026, 4,061 requests have been filed. Of the 2,878 that have resolved, 553 reached Done and 2,325 were withdrawn — FD estimate: 2,325 ÷ 2,878 = 80.8% withdrawn, 19.2% completed. The active queue holds 1,091 requests for 254,591 MW of summer net capacity. FD estimate: 254,591 MW × 19.2% = about 48,881 MW, which is what that queue is worth at the operator's own historical completion rate. Against the active 254,591 MW, exactly 20,285 MW across 128 projects is under construction after signing an interconnection agreement — FD estimate: 20,285 ÷ 254,591 = 8.0%. Anyone quoting a queue total as a demand forecast is quoting a number with an 81% historical failure rate attached.Confidence: high. Every figure is a count Frontier Desk performed over the operator's own public feed, and the method is published in band 15 so anyone can repeat it. The one inference is that past withdrawal rates predict future ones, and band 11 names what would break it.
- 02The planner that matters most has already cut its own near-term forecast while its queue grew.PJM's 2026 Long-Term Load Forecast, dated January 2026, states that it "is lower than the 2025 Long-Term Load Forecast in the near term through 2032 due to updates to the electric vehicle forecast, economics and large load adjustments". The cut to the summer peak for the 2028 capacity auction year is 4,414 MW, or 2.6%. PJM says why: near-term years now need "firm" commitments such as an Electric Service Obligation or Construction Commitment, while longer-term projects are treated as non-firm and derated. The same report still projects 3.6% a year of summer peak growth over ten years. Both halves are the story: the long-run number is enormous and the near-run number is being marked down as the evidence standard rises.Confidence: high. The cut, its size and its stated reason are quoted from the planner's own report. The reading that an evidence standard, rather than a change in demand, caused the near-term cut is the planner's own explanation rather than our inference.
- 03The reliability body says data centers drive most of the demand growth, and in the same paragraph says the near-term projects are slipping.NERC's 2025 Long-Term Reliability Assessment finds that "New data centers for artificial intelligence and the digital economy account for most of the projected increase in North American electricity demand over the next 10 years", with aggregated summer peak forecast to rise by over 224 GW — 69% higher than the previous year's 132 GW ten-year projection — and winter peak by 245 GW. It also records that ERCOT and PJM have both revised load forecasts since its data collection, and that "Both forecasts indicate that some large-load projects have slowed or failed to materialize within the shorter-term horizon, while interconnection requests for later years continue to increase." Thirteen of 23 assessment areas face resource adequacy challenges.Confidence: high. One authoritative assessment, quoted on both sides of its own argument. What it does not settle is which of the two readings dominates by 2028, which is what the scenarios are for.
- 04The load is real where it has already connected, and the filings now show it in revenue.Dominion Energy reports a regulated electric sales class it defines as "customers in Virginia, including certain data centers, with actual or anticipated forecast demand of 25 MW or higher and annual load factor of 75% or higher". That class billed $2,156 million in the nine months to 30 September 2026 against $1,308 million a year earlier — FD estimate: $2,156M ÷ $1,308M = 1.65 times, a 64.8% rise. It is now $2,156M of Virginia Power's $11,153M of revenue from contracts with customers — FD estimate: $2,156M ÷ $11,153M = 19.3%. Whatever the queue turns out to be worth, the part that has already connected is growing faster than any other customer class on that utility's income statement.Confidence: high. The class definition, both figures and the totals are quoted from one registrant's Form 10-Q for the quarter ended 30 September 2026, and the two ratios are shown. The class includes some non-data-center customers, which the filing says and we repeat.
- 05The capacity market has told the same story four times, and the fourth time it told it at the cap.PJM's 2028/2029 Base Residual Auction cleared at the $325.00 per MW-day price cap on 14 July 2026, with 138,317.8 MW UCAP cleared and total procured capacity 6,831.3 MW below the reliability requirement, at a 14.4% installed reserve margin against a 20% target. The clearing price was $28.92 for delivery year 2024/25. The product of cleared supply and clearing price went from $2.2 billion for 2023/24 to $16.4 billion for 2028/29 — FD estimate: $16.4bn ÷ $2.2bn = 7.5 times in five delivery years. Total new generation and uprates in the 2028/29 auction were 524.7 MW UCAP. A price that rises 11-fold and calls forth 525 MW is not a signal the market is failing to hear. It is a signal that price is not the constraint.Confidence: high. Every figure is from the market operator's own results report. The inference that price is not the binding constraint is ours, and the new-build figure is the evidence we rest it on.
- 06The binding constraint is moving from the queue to the factory, and one manufacturer's cash flow shows exactly when.GE Vernova's cash from operating activities for the nine months to 30 September 2026 included a $19.5 billion inflow from contract liabilities and current deferred income, "driven by down payments on orders and slot reservation agreements at Power", against $2.4 billion in the same period of 2025. Its remaining performance obligation was $196,721 million, of which equipment-related RPO was $104,054 million with 33% expected within one year and 60% within two — FD estimate: $104,054M × 40% = $41,622M scheduled more than two years out. Customers are paying cash years ahead to reserve a manufacturing slot. That is what a queue looks like when the thing being queued for is a machine rather than a permission.Confidence: high on the readings, moderate on the reading of them. The figures are quoted from the manufacturer's own quarterly filing. That slot reservations are now a tighter constraint than interconnection in some regions is our inference, and it is not something a single manufacturer's filing can establish.
- 07The regulator has moved from watching to ordering, and the deadline has already passed.At its 18 June 2026 open meeting the Federal Energy Regulatory Commission opened six Section 206 show-cause proceedings — Docket Nos. EL26-67-000 through EL26-72-000 — one against each regional transmission organisation, requiring each to justify its existing tariff for large loads or file changes within 60 days, and to file an informational report within 30 days describing how it will ensure adequate generation for existing and new large loads. In the same meeting it acted on rehearing, compliance and paper hearing in the PJM co-located load docket, EL25-49-002. Two independent professional analyses published 23 June 2026 describe the same actions in the same terms. Whatever each region files, the rules for connecting a large load in 2028 will not be the rules of 2026.Confidence: moderate. The dockets, the dates and the deadlines come from two independent named professional analyses that agree, rather than from the Commission's own order text, which we could not retrieve. They are graded C and attributed in the sentence that carries them.
- 08A site with firm power is now worth more than a site with capital, and that inversion is the whole investment case.Hold the readings together. Capacity cleared at its cap and called forth 525 MW of new build. The queue's historical completion rate is 19.2%. Equipment backlog runs two years and more. The planner derates anything without a firm commitment. Every one of those is a statement that money cannot buy a date. What can buy a date is an existing interconnection agreement, a reserved turbine slot, a brownfield site with capacity already behind it, or a willingness to be curtailed. 57% (judgment) that a US large-load interconnection position or energised brownfield site changes hands in a disclosed transaction valued above $1 million per MW by 31 December 2028.Confidence: moderate. The four readings under it are each high-confidence and independent. The conclusion that they price a queue position is an inference, and no public source yet states a dollar-per-MW figure for one, which is why the call is written to create that evidence rather than to rely on it.
The baseline
Where data center power and interconnection stand on 5 November 2026. This band is collection, not analysis: every row is a reading from a named public source or a count Frontier Desk performed over a public dataset, graded A, B or C, with the arithmetic shown and no inference attached.
The readings below are the position the rest of this report forecasts from: a queue far larger than anything that will be built, an operator record showing four in five requests withdrawn, a planner cutting its own near-term forecast while that queue grows, a capacity price at its cap calling forth almost no new generation, and a connected data center load already visible in a utility's revenue.
Active in one operator's queue
1,091 active requests, summer net capacity, counted from the operator's own public feed on 5 Nov 2026.
Of resolved requests, withdrawn
FD estimate: 2,325 withdrawn ÷ 2,878 resolved, over the whole history in that feed.
Capacity price per MW-day, at the cap
The 2028/29 delivery year, cleared 14 Jul 2026 against $28.92 for 2024/25.
The planner's own cut to its 2028 peak
PJM's 2026 forecast against its 2025 forecast for the 2028 auction year, a 2.6% reduction.
Every reading, with its source
Publicly observableGrades describe the source, not the conclusion. A — record: a grid operator's own published data or results report, a filing with a securities regulator, a government publication, or a reliability body's own assessment. B — party: an organisation's own page or announcement about itself. C — crowd: reputable named press and professional analyses of a regulator's actions, carried with their attribution.
Confidence describes Frontier Desk's inference from those readings: high where two or more independent sources agree and none contradicts, moderate where the direction is corroborated but the magnitude is not, low where a single source or an analogy carries it. The odds on this page are judgments of the same kind, not the output of a model, and they are stated as judgments wherever they appear.
| Reading | Source | Figure | Grade |
|---|---|---|---|
| Active interconnection requests in one operator's queue | The operator's own public project feed, counted 5 Nov 2026 | 1,091 requests, 254,591 MW summer net | A · record |
| Requests in that feed that have ever resolved | The same feed, counted 5 Nov 2026 | 2,878 of 4,061: 553 Done, 2,325 Withdrawn | A · record |
| Historical completion rate in that feed | FD estimate from the same feed | 19.2% (553 ÷ 2,878); 80.8% withdrawn | A · record |
| Capacity withdrawn over that history | The same feed, counted 5 Nov 2026 | 395,238 MW | A · record |
| Projects under construction after signing an interconnection agreement | The same feed, counted 5 Nov 2026 | 128 projects, 20,285 MW — 8.0% of active MW | A · record |
| Median time from request to in-service, for projects that finished | FD estimate over 512 of the 548 Done records in the same feed | 3.2 years | A · record |
| Median age of the requests sitting in the active queue today | FD estimate over the same feed | 1.6 years; the oldest is 5.8 years | A · record |
| Largest active fuel types by capacity in that queue | The same feed, counted 5 Nov 2026 | Not stated 103,425 MW; solar 47,881; battery 39,641; gas 25,400; wind 24,547 | A · record |
| PJM capacity clearing price, delivery year 2028/29 | The market operator's own auction results report | $325.00 per MW-day, at the cap; $28.92 for 2024/25 | A · record |
| Capacity cleared and the shortfall against the requirement | The same report | 138,317.8 MW UCAP cleared; total procured 6,831.3 MW below the reliability requirement | A · record |
| Installed reserve margin cleared against target | The same report | 14.4% against a 20% target, the second year running more than a point below | A · record |
| New generation and uprates in that auction | The same report | 524.7 MW UCAP | A · record |
| Cleared supply times clearing price | The same report | $16.4 billion for 2028/29, against $2.2 billion for 2023/24 | A · record |
| PJM's own revision to its near-term load forecast | PJM 2026 Long-Term Load Forecast, January 2026 | Lower through 2032; −4,414 MW (−2.6%) for the 2028 auction year | A · record |
| Why the near-term forecast fell | The same report | Near-term years now need firm commitments; longer-term projects are non-firm and derated | A · record |
| PJM ten-year growth, as still forecast | The same report | Summer peak +3.6% a year to 222,106 MW in 2036; net energy +5.3% a year | A · record |
| North American ten-year peak demand growth | NERC 2025 Long-Term Reliability Assessment | Summer +224 GW, 69% above the prior year's 132 GW projection; winter +245 GW | A · record |
| What drives it, in the assessment's own words | The same assessment | New data centers for AI and the digital economy account for most of the projected increase | A · record |
| What the same assessment says about near-term projects | The same assessment | Some large-load projects have slowed or failed to materialize in the shorter-term horizon | A · record |
| Assessment areas facing resource adequacy challenges | The same assessment | 13 of 23 over the next 10 years | A · record |
| Announced and confirmed generator retirements | The same assessment | Over 105 GW of peak seasonal capacity over 10 years, about 10 GW lower than the prior year | A · record |
| Change in the resource mix in one year | The same assessment | Fossil capacity fell 21 GW from 2024 to 2025; battery, wind and solar peak-hour capacity rose 23 GW | A · record |
| US data center electricity consumption | US Department of Energy, 20 Dec 2024 | 176 TWh in 2023, about 4.4% of US electricity, from 58 TWh in 2014 | A · record |
| The same report's 2028 projection | The same publication | 325 to 580 TWh, approximately 6.7% to 12% of total US electricity | A · record |
| ERCOT large-load interconnection requests | Utility Dive, 5 Aug 2026, reporting ERCOT and the Governor of Texas | 474 GW of requests, approximately 90% data centers | C · crowd |
| Data center revenue at one utility's large-load class | Dominion Energy Form 10-Q, quarter ended 30 Sep 2026 | $2,156M for nine months against $1,308M, a 64.8% rise; 19.3% of Virginia Power customer revenue | A · record |
| Turbine manufacturer down payments and slot reservations | GE Vernova Form 10-Q, quarter ended 30 Sep 2026 | $19.5bn inflow from contract liabilities and deferred income, against $2.4bn a year earlier | A · record |
| That manufacturer's equipment backlog and its timing | The same Form 10-Q | $104,054M equipment RPO; 33% within one year, 60% within two, 97% within five | A · record |
| Wholesale power prices | EIA Short-Term Energy Outlook, 6 Oct 2026 | US average $52 per MWh in 2026, 11% above 2025; PJM forecast to rise 41% | A · record |
| US average retail electricity price, all sectors | EIA Electric Power Monthly, August 2026 preliminary | 14.87 cents per kWh against 14.15 a year earlier; industrial 9.64 against 8.98 | A · record |
| Federal regulatory action on large loads | Two independent professional analyses, both 23 Jun 2026 | Six Section 206 show-cause proceedings, EL26-67-000 to EL26-72-000, opened 18 Jun 2026 | C · crowd |
One dataset carries more of this forecast than any other, and it is worth saying exactly what was done to it. MISO publishes its generator interconnection queue as a public JSON feed with one record per request: a queue date, a study cycle, a service type, a summer net megawatt figure, an application status and a post-agreement construction status. Frontier Desk downloaded that feed on 5 November 2026 and counted it.
The counts are these. 4,061 records. By application status: 1,091 Active, 2,325 Withdrawn, 548 Done, 5 Legacy Done, 75 Pending Revision Approval, 17 Pending Transfer. Summing summer net megawatts over the active records gives 254,591 MW; over the withdrawn records, 395,238 MW. By post-agreement status, 128 records totalling 20,285 MW are Under Construction and 92 totalling 14,243 MW are In Service. For the 512 of the 548 Done records carrying both a queue date and an in-service date with a positive gap, the median elapsed time is 3.2 years.
Two things follow that are not in dispute. The first is that the headline queue number and the build number differ by more than an order of magnitude. The second is that the difference is not a mystery or a scandal: it is the ordinary behaviour of an option that is cheap to take and expensive to exercise. The forecast in this report is mostly about what happens when the option stops being cheap.
Hold those readings together and the next two years are already implied. A queue with an 81% withdrawal rate will either be filtered by rules or discounted by planners, and both are now happening. A capacity price at its cap that calls forth 525 MW will not be fixed by a higher cap. A manufacturer taking cash two years ahead is a constraint capital cannot relieve. And a utility billing 65% more to its large-load class than a year ago is proof that the part of this which is real is very real indeed.
Structural shifts underway
Seven shifts are already running. Each separates what the sources say from what we infer from them: the evidence first with its grade, then the reading with its confidence, then how fast, then the specific thing that would reverse it. These are the inputs to the scenarios.
From an open queue to a rationed oneAccelerating
Speed: Already running; settled by 2028
What the sources say. Counting MISO's public feed on 5 November 2026 gives 2,325 withdrawn requests against 553 completed, an 80.8% withdrawal rate over 2,878 resolved requests, and 395,238 MW of withdrawn capacity. PJM's 2026 Long-Term Load Forecast states that near-term forecast years "need 'firm' commitments, such as Electric Service Obligation (ESO)/Construction Commitments (CC), while longer-term projects will be considered 'non-firm' and will be derated because of their greater uncertainty". NERC's 2025 assessment records that Texas lawmakers have granted ERCOT authority to curtail new large loads to prevent grid emergencies.
Evidence grade: A on all three — an operator's own feed counted by us, a planner's own report, and a reliability body's own assessment.
What we read into it. A queue with an 81% failure rate is not a demand signal; it is a free option. Everything happening in 2026 is a different party taking the option's price up: deposits and readiness tests at the operator, derating at the planner, curtailment authority at the legislature, and show-cause proceedings at the regulator. Rationed access is where this ends, and the practical consequence for an owner is that a queue position stops being something you file for and becomes something you qualify for.
Confidence: high. Three independent sources, each describing a different mechanism pointing the same way, and the withdrawal rate is a count anyone can reproduce from the same feed.
What would reverse it. A federal or state rule that speeds connection without adding a commitment test — connect-and-manage applied broadly, for instance — which would let the queue stay open and push the filtering onto operations instead.
Planners start discounting the queue in publicRunning
Speed: Already visible; standard by 2027
What the sources say. PJM's 2026 Long-Term Load Forecast, dated January 2026, is "lower than the 2025 Long-Term Load Forecast in the near term through 2032", with the summer peak for the 2028 capacity auction year cut by 4,414 MW or 2.6%, the 2026 third-incremental-auction year by 2,564 MW or 1.6%, and the 2031 planning year by 1,630 MW or 0.8%. NERC's 2025 assessment records that ERCOT and PJM have both revised forecasts since its data collection and that "some large-load projects have slowed or failed to materialize within the shorter-term horizon, while interconnection requests for later years continue to increase".
Evidence grade: A on both — a planner's own report and a reliability body's own assessment.
What we read into it. The shape of the revision is the point: cuts are deepest two years out and shallowest five years out. That is exactly what you would expect if the problem were evidence rather than demand — near-term years now demand a firm commitment and most projects cannot produce one, while far years still accept a request. Expect every planner to publish a derating method within two years, and expect the published methods to differ enough that the same project implies different load in different regions.
Confidence: high. The cuts, their sizes and the planner's stated reason are all quoted. The forecast that other planners copy the method is our inference, supported by the regulator's show-cause proceedings requiring each operator to defend its own practice.
What would reverse it. PJM's next forecast restoring the near-term cut, which would mean the firm-commitment test is being satisfied at scale and the derating was a timing artefact rather than a structural filter.
Price stops clearing the market, and the constraint moves to physical supplyRunning
Speed: Decisive through 2028
What the sources say. PJM's 2028/2029 Base Residual Auction cleared at the $325.00 per MW-day cap, procuring 6,831.3 MW UCAP less than the reliability requirement at a 14.4% installed reserve margin against a 20% target, with total new generation and uprates of 524.7 MW UCAP. The clearing price was $28.92 for delivery year 2024/25 and $16.4 billion of cleared supply times price for 2028/29 against $2.2 billion for 2023/24. GE Vernova's nine-month cash flow included a $19.5 billion inflow from contract liabilities "driven by down payments on orders and slot reservation agreements at Power", against $2.4 billion a year earlier.
Evidence grade: A on both — the market operator's own results report and the manufacturer's own quarterly filing.
What we read into it. A capacity price that rises eleven-fold and produces 525 MW is a market telling you the supply curve is vertical in the relevant range. The reason is on the manufacturer's balance sheet: buyers are paying cash two years ahead for slots, and 40% of the equipment backlog is scheduled beyond two years. When the binding constraint is a factory rather than a price, raising the price moves the queue order and not the quantity.
Confidence: high on the readings. Moderate on the generalisation, because one manufacturer's backlog is not the whole turbine market and the auction result also reflects a cap that was set administratively.
What would reverse it. A capacity auction clearing below cap with a reserve margin at or above target, or equipment RPO falling and the share scheduled beyond two years falling with it.
Large loads stop waiting and bring generation with themAccelerating
Speed: 12 to 36 months
What the sources say. The Federal Energy Regulatory Commission opened six Section 206 show-cause proceedings on 18 June 2026, Docket Nos. EL26-67-000 to EL26-72-000, requiring each regional transmission organisation to justify or revise its large-load tariff within 60 days and to report within 30 days on how it will ensure adequate generation, and acted on rehearing, compliance and paper hearing in the PJM co-located load docket EL25-49-002, according to two independent professional analyses both published 23 June 2026. A trade body note of 23 July 2026 records development moving onto federal land, naming an Air Force proposal for AI data centers on three bases in Alaska and Army negotiations over facilities at Fort Bliss and Dugway Proving Ground.
Evidence grade: C · crowd on both — two professional analyses that agree, and a trade body summarising a bank's research. Neither is the regulator's own order text, and the page says so.
What we read into it. When the queue is the constraint, the rational move is to leave it. Co-location, on-site generation and federal land are three versions of the same decision: accept a worse cost of energy in exchange for a date you control. The regulator's response is not to stop it but to write rules for it, which is the clearest signal available that this is a permanent feature rather than a loophole. Expect the share of new large-load capacity that is self-supplied to rise every quarter to 2028.
Confidence: moderate. The direction is well supported and the specific dockets are corroborated by two sources, but we could not read the Commission's own orders and no public dataset yet counts self-supplied large-load capacity.
What would reverse it. A FERC or state decision making co-located load pay full transmission charges with no reduction for reduced grid usage, which would remove most of the economic advantage of leaving the queue.
Flexibility becomes a product, and curtailment becomes a termRunning
Speed: Standard by 2028
What the sources say. NERC's 2025 assessment records that "Lawmakers in Texas have provided ERCOT with curtailment management authority over new large loads to prevent grid emergencies and established funding to speed new generating capacity to the grid", and that load forecasts now reflect "agreements between utilities and owners of connecting loads, such as facility peak demand, load flexibility, and some operating characteristics". The FERC show-cause proceedings are described by both professional analyses as focused in part on the treatment of flexible or curtailable loads.
Evidence grade: A on the assessment; C · crowd on the description of the FERC proceedings.
What we read into it. A grid short of capacity will always pay more for load it can turn off than for load it cannot, and the moment that is written into a tariff it becomes a price. For an owner the practical question changes from whether you can get power to how many hours a year you will give back and at what notice. Expect curtailable service to be cheaper and faster to obtain than firm service by 2028, and expect the discount to be large enough that the decision is a real one rather than a formality.
Confidence: moderate. The legislative authority and the forecasting practice are quoted from an authoritative source. That the discount becomes large enough to change behaviour is our judgment, and no published tariff yet prices it at scale.
What would reverse it. Large loads refusing curtailable terms at scale, visible as firm-service requests continuing to dominate the queues after the tariff revisions land, which would mean the discount is not worth the operational risk.
The connected load is real, and it shows up in revenue before it shows up in forecastsRunning
Speed: Already here
What the sources say. Dominion Energy's regulated electric sales include a "High load" class defined as "customers in Virginia, including certain data centers, with actual or anticipated forecast demand of 25 MW or higher and annual load factor of 75% or higher". That class billed $706 million in the quarter to 30 September 2026 against $503 million a year earlier, and $2,156 million for nine months against $1,308 million. Virginia Power's total revenue from contracts with customers for nine months was $11,153 million. The US Department of Energy reports data center consumption at 176 TWh in 2023, about 4.4% of US electricity, from 58 TWh in 2014, projected at 325 to 580 TWh by 2028.
Evidence grade: A on both — a registrant's own filing and a government publication.
What we read into it. Every argument in this report about queues being soft has to be set against a utility whose large-load class grew 64.8% in a year. The honest reading is that the demand is genuine and the queue is a bad measure of it: load that has signed, connected and is paying shows up here, and load that has filed a request shows up in the queue, and the two differ by the whole distance this forecast is about. An owner should trust the revenue line and discount the queue line, in both directions.
Confidence: high. The figures are quoted and the ratios are shown. The class definition includes customers that are not data centers, which the filing states and we repeat rather than netting out.
What would reverse it. That class's growth rate falling below the utility's commercial class for two consecutive quarters, which would be the first filed evidence that connected data center load is slowing rather than its queue.
The grid's cost lands on everyone, and that becomes a political constraintRunning
Speed: 2027 to 2028
What the sources say. EIA's Short-Term Energy Outlook of 6 October 2026 forecasts US wholesale electricity prices averaging $52 per megawatt hour in 2026, "11% more than in 2025", with PJM prices expected to rise 41%. EIA's Electric Power Monthly for August 2026 shows average retail prices of 14.87 cents per kWh across all sectors against 14.15 a year earlier, with industrial at 9.64 against 8.98. The Governor of Texas ordered an audit of data centers in the ERCOT queue and said any project failing the verification and audit process "must be denied", as reported by Utility Dive on 5 August 2026 against a queue of 474 GW of requests, approximately 90% of them data centers.
Evidence grade: A on both EIA series; C · crowd on the Texas reporting, which is reputable named press quoting a public order.
What we read into it. Capacity costs that went from $2.2 billion to $16.4 billion in five delivery years are paid by load, and most load is not a data center. The politics of that arrive before the economics are settled, which is why an audit arrived before a rate case. For an owner the risk is no longer only whether power is available; it is whether the terms on which it was promised survive an election. Expect at least two more states to impose a large-load-specific approval, tariff or audit by the end of 2028.
Confidence: moderate. The price series are authoritative and the Texas action is reported by a named outlet quoting a named official. The forecast that it spreads to other states is a judgment, and the mechanism — cost allocation politics — is an inference rather than a stated cause.
What would reverse it. Retail prices flattening while large-load connections continue, or a state adopting a tariff that visibly insulates other customers from large-load costs and defusing the issue.
The causal model
The loop that produced everything above, in six steps, with the figure that shows each one is running. Three arms reinforce it and three brake it; the scenarios are what happens when one arm or one brake wins.
- 1
Request
A developer files for interconnection, cheaply and early
1,091 active requests - 2
Forecast
The planner adds some of it to the load forecast
+3.6% a year, 10 years - 3
Price
Capacity and energy prices rise to call forth supply
$325.00 per MW-day - 4
Order
Buyers pay ahead to reserve generation and equipment
$19.5bn in down payments - 5
Filter
Firm-commitment tests and deposits remove the weakest requests
80.8% withdrawn - 6
Revise
The planner cuts the near-term forecast and the cycle restarts
−4,414 MW for 2028
↻ Each round of filtering makes the near-term forecast smaller and the far-term queue larger, which makes the next filter stricter. The loop has run once in public already, in the step from PJM's 2025 forecast to its 2026 one.
The three arms, and the three brakes
Read the loop and the scenarios stop being moods. Rationed access is B1 winning. Build-your-own is R3 winning. Demand disappoints is B1 and B3 together. A reliability event is B2 arriving before any of the brakes have finished working. The odds below are our judgment about which gets there first.
How it evolves, 2027 and 2028
Five phases to the end of 2028, each named for what changes in it, with the Frontier Desk judgments that attach to that window. Every figure here is one of the calls in band 13, carried unchanged.
The tariffs come back, and the first derating methods are published
Six regional transmission organisations have responded to show-cause proceedings opened on 18 June 2026, and the Commission has to act on what they filed. Expect at least one tariff revision accepted with a large-load-specific study process, expect at least one operator to publish a deposit or readiness requirement, and expect the next round of planner load forecasts to carry an explicit derating method rather than an adjustment buried in a table. The tell is a published rule rather than a published number.
A large-load-specific tariff is accepted at one or more RTOs: 76% · a second planner publishes an explicit derating method: 64%
The queue stops being the headline and the firm number starts
Once two planners publish derating methods, the comparison everyone makes changes from requests to firm commitments, and the two numbers diverge enough to be newsworthy. Expect the first region to publish both side by side, expect the ratio to be startling, and expect at least one developer to complain in a filing that the test is arbitrary. This is the window in which the investment case for a brownfield site with an existing agreement becomes obvious to everyone rather than to specialists.
A grid operator publishes firm versus requested large load side by side: 58% · PJM's next forecast cuts the near term again: 51%
Equipment, not interconnection, becomes the quoted lead time
On the central path the queue filter does its work and the binding constraint moves visibly to the factory. Expect turbine and transformer lead times to be quoted in project announcements the way interconnection dates are quoted now, expect slot reservation payments to keep rising on manufacturers' balance sheets, and expect the first disclosed resale of a reserved slot. The equivalent in the financing market is covered in our companion forecast on compute financing.
Equipment RPO beyond two years keeps rising: 61% · a reserved equipment slot is disclosed as transferred or resold: 38%
Capacity for delivery year 2028/29 starts, at the cap
The delivery year that cleared at $325.00 per MW-day begins on 1 June 2028, with the market 6,831.3 MW short of its own reliability requirement at a 14.4% reserve margin. This is where a forecast stops being a forecast. Expect the cost to reach retail bills in the regions concerned, expect the political response that follows, and expect the first serious public argument about whether large loads should pay a different capacity rate from everyone else.
A state adopts a large-load-specific tariff, approval or audit: 72% · a reliability event involving large-load curtailment: 25% by end-2028
Rationed access, or the forecasts come down again
On the central path a queue position is something a developer qualifies for, self-supply has taken a visible share of new large load, curtailable service is cheaper than firm, and the published load forecasts have stabilised at a level well below today's queue totals and well above today's connected load. On the alternative path the firm commitments keep failing to appear, the forecasts are cut a third time, and the capacity price falls off its cap with the shortfall unresolved for the opposite reason.
Rationed access: 41% · demand disappoints: 17% · an interconnection position changes hands above $1m per MW: 57%
Scenarios
Five ways data center power can go. Each carries Frontier Desk odds at 31 December 2027 and 31 December 2028, the trigger that puts you on it, the early markers that tell you first, and what it means on each side of a project. Each column sums to 100, and every signpost in band 12 names the scenario it moves.
Open queue
Today's shape continues. Filing stays cheap, queues keep growing, withdrawal rates stay near four in five, and the published totals keep being quoted as demand by people who have not read the withdrawal column. The reason this falls from 34% to 13% is not that the queues shrink. It is that every party with authority over them — operator, planner, legislature and regulator — is currently acting to put a price on a position, and most of those actions land inside the forecast window.
Trigger: None. This is the no-trigger path: the show-cause proceedings produce modest changes, deposits stay small, and derating stays confined to one planner.
Early markers: Withdrawal rates holding near 80%; queue totals still rising; no second planner publishing a derating method; deposits unchanged.
For an owner: A queue position is cheap and nearly worthless, so file early, file in more than one place, and treat the date as an option rather than a plan.
For utilities: Study costs keep rising against requests that mostly evaporate. The operator carries the cost of other people's options, which is precisely why this scenario ends.
Rationed access
The modal destination by the end of 2028. A queue position requires a deposit, a site control test, a firm commitment or all three; planners count only what clears those tests; and the published forecast separates requested load from committed load. Queues shrink without demand falling, because the filtering removes options rather than projects. An existing interconnection agreement becomes the most valuable thing a site can have, and the gap between a qualified developer and an unqualified one becomes the main source of returns.
Trigger: Two or more regional operators adopting firm-commitment or deposit requirements that materially cut filings, and a second planner publishing a derating method.
Early markers: Deposit or readiness requirements adopted; queue filings falling while connected load rises; firm and requested load published side by side.
For an owner: Qualify before you file. Site control, a signed offtake, an equipment slot and a credible construction schedule are now the application, and the project without them has no date at any price.
For utilities: Study backlogs fall, cost recovery improves, and the planning forecast becomes defensible. This is the outcome every operator is currently working toward.
Build-your-own
Large loads stop waiting and bring generation with them: on-site turbines, co-located nuclear or gas, federal land, behind-the-meter supply. The grid becomes one option among several rather than the default. Note the shape of the odds: this grows but does not take over, because the equipment that self-supply needs is the same equipment the grid needs, and the backlog is shared. It is a way around the queue, not around the factory.
Trigger: A FERC decision confirming a reduced transmission charge for co-located load, or an equipment manufacturer prioritising direct-to-load orders over utility orders.
Early markers: Co-located capacity announced per quarter; federal-land projects proceeding; behind-the-meter generation appearing in operators' load forecasts as a deduction.
For an owner: The best date available if you can afford the energy cost and the operating complexity. Model the fuel and the maintenance honestly, because a private grid is a business you are now in.
For utilities: Load leaves the system while the fixed costs stay, which is the cost-allocation argument the political brake is built on.
Demand disappoints
Requests keep rising while firm commitments fall, planners cut forecasts a second and third time, and the gap between the queue and the connected load turns out to have been mostly speculation. Capacity prices come off their cap for the wrong reason. This is the second-likeliest 2028 outcome and the one least represented in current project economics, and the evidence for it is already in NERC's own wording about projects that have slowed or failed to materialise.
Trigger: A second consecutive PJM near-term cut, or a large-load class at a major utility growing more slowly than its commercial class.
Early markers: Queue filings rising while firm commitments fall; a utility's large-load revenue growth decelerating; announced projects cancelled rather than delayed.
For an owner: Time is the risk. A project committed into this scenario is competing for a tenant that does not arrive, which argues for contracted capacity before construction rather than after.
For utilities: Stranded investment in transmission and generation built for load that never connects. The derating methods now being published exist precisely to prevent this.
Reliability event
A shortfall, an extreme weather event or a forced curtailment of large load produces an emergency rule change. NERC already finds 13 of 23 assessment areas facing resource adequacy challenges, PJM cleared 6,831.3 MW below its reliability requirement, and Texas has already given its operator curtailment authority over new large loads. The odds do not rise with the horizon because each year of build-out adds both the exposure and the resources; this is a tail, not a trend.
Trigger: Any firm load shed attributed to resource adequacy in a major market, or a large load curtailed under emergency authority.
Early markers: Reserve margins falling further below target; emergency alerts rising; an operator invoking large-load curtailment authority for the first time.
For an owner: Everything depends on what you agreed before the event. A curtailable contract is a liability you planned for; a firm contract in an emergency is a negotiation you did not.
For utilities: Rules change quickly and retroactively, and the party holding the least documented position loses the argument.
Winners and losers
Who gains and who loses on the central path, and why. This band is analysis throughout: direction is stated against today, not against each other, and nothing here is a reading.
Owners of sites, load and generation
| Segment | Direction | Why |
|---|---|---|
| Owners of brownfield sites with existing interconnection capacity | Clear winner | An energised position is the one thing a rationed queue cannot create. With 8.0% of one operator's active queue under construction and a 3.2-year median build time for projects that finish, a site that is already connected skips the entire filter. |
| Industrial owners with large retired or under-used electrical service | Clear winner | A closed plant with 40 MW of live service is a data center site in a market where 40 MW of new service takes years. This is the single most under-priced asset class the forecast identifies, and the owners mostly do not know it. |
| Developers holding speculative queue positions with no site control | Clear loser | Every filter now being built is aimed at exactly this position. Deposits, readiness tests and firm-commitment standards convert a free option into a funded obligation, and the 80.8% withdrawal rate says what happens to most of them. |
| Operators of existing dispatchable generation near load | Clear winner | A capacity market clearing at its cap with 524.7 MW of new build is a market paying a scarcity rent to whatever already exists. Co-location makes a plant beside a data center worth more than the same plant elsewhere. |
| Loads willing to be curtailed | Winner, selectively | A grid short of capacity pays for flexibility, and Texas has already written the authority into law. An owner whose process can stop for a hundred hours a year has a date everyone else is waiting for. |
| Owners relying on a 2028 connection date promised in 2025 | Loser | The near-term forecasts have already been cut once, equipment backlogs stretch past two years, and the firm-commitment tests arrived after many of these dates were given. Re-confirm the date in writing before you commit the capital. |
Utilities, operators and the supply chain
| Segment | Direction | Why |
|---|---|---|
| Regulated utilities with large-load tariffs and connected demand | Clear winner | One utility's large-load class grew 64.8% year on year to $2,156 million over nine months. This is rate base growth with a creditworthy counterparty, and it is the cleanest exposure to the theme available anywhere. |
| Turbine, transformer and switchgear manufacturers | Clear winner | $19.5 billion of down payments and slot reservations in nine months, against $2.4 billion a year earlier, is customers funding the factory. Pricing power lasts exactly as long as the backlog does. |
| Grid operators and their planning departments | Winner in authority, loser in workload | They gain the power to ration and the obligation to defend how. Six show-cause proceedings mean every one of them now has to justify its rules to a federal regulator on a deadline. |
| Merchant developers of solar and storage in crowded queues | Loser | Solar at 47,881 MW and battery at 39,641 MW are the largest named fuel types in one active queue. The filters coming are not aimed at them, but they are standing in the line those filters shorten. |
| Residential and small commercial customers | Clear loser | Capacity costs went from $2.2 billion to $16.4 billion in five delivery years and retail prices rose 5.1% year on year across all sectors. The political brake in band 04 is this group noticing. |
Frontier Desk is the advisor to the company that owns the site, the load or the generator, never a utility and never a broker. The first conversation covers what you actually hold, what a lender or a buyer will make of it, and what to establish before you commit capital.
Commoditised or scarce
What becomes a commodity in this market and what keeps its scarcity value, with the date we think each crosses. This band is a forecast, and the band after it argues against itself.
Commoditising, and by when
| What | By when | Why |
|---|---|---|
| A speculative interconnection request | 2027 | An option that costs nothing is worth nothing once everyone holds one. The filters arriving in 2027 will make the request expensive and the position scarce, which are opposite things. |
| Queue totals as a market statistic | 2027 | Once two planners publish derating methods, quoting a raw queue number marks the speaker rather than the market. An 80.8% withdrawal rate is public and recomputable. |
| Solar and battery capacity in a crowded queue | 2028 | The two largest named fuel types in one active queue, at 47,881 MW and 39,641 MW, compete with each other for the same study slots and the same interconnection upgrades. |
| Generic greenfield data center land | 2028 | Land has never been the constraint. A site without electrical service is an agricultural asset with a story attached, and the market is learning to price it that way. |
| Capital for grid-connected projects | Already | Financing is abundant and dates are not. Our companion forecast on compute financing covers what has happened to the cost of capital over the same period. |
Keeping its scarcity value, and why
| What | How long | Why |
|---|---|---|
| An executed interconnection agreement with energised capacity | Through 2028 and beyond | 128 projects totalling 20,285 MW are under construction after signing one, against 254,591 MW of active requests. The agreement, not the request, is the asset. |
| A reserved turbine or transformer delivery slot | Through 2028 | 40% of one manufacturer's $104,054 million equipment backlog is scheduled beyond two years, and customers are paying cash to hold places in it. |
| Dispatchable generation sited near load | Through 2028 and beyond | A capacity auction clearing at its cap with 524.7 MW of new build is paying scarcity rent to what already exists, and co-location adds a locational premium on top. |
| Willingness and ability to be curtailed | Through 2028 | A grid short of its reliability requirement will pay for load it can turn off. Very few loads can actually do it, which is what keeps the premium. |
| Regulatory and community standing in a host jurisdiction | Through 2028 and beyond | An audit arrived in Texas before a rate case did. The right to build is becoming as scarce as the ability to, and it cannot be bought in a queue. |
The case against this table.
The strongest argument against this table is that we have generalised one operator's queue to the country. MISO is not PJM and neither is ERCOT, whose large-load process is new, differently designed and reported at 474 GW of requests with roughly 90% data centers. An 80.8% withdrawal rate measured over a history dominated by merchant solar and wind requests may say very little about data center load requests, which are backed by different sponsors with different costs of abandonment. If data center requests withdraw at half that rate, the implied capacity doubles and the rationed-access scenario arrives later and matters less.
The second argument is that we have read a forecast cut as a structural signal when it may be an accounting one. PJM's 4,414 MW reduction for 2028 came from reclassifying projects as non-firm, not from any project cancelling. If those projects simply produce their Electric Service Obligations over the next year, the cut reverses, the queue was right all along, and the whole first half of this report is a story about paperwork. We hold demand disappoints at 17% rather than higher for exactly this reason, and signpost 02 is designed to catch the reversal within one forecast cycle.
What to do about it
What the forecast implies for the two readers this report is written for: the owner of a site, a load or a generator, and the adviser or investor underwriting one. Frontier Desk is an advisor, not a utility, not a developer and not a broker.
If you own a site, a load or a generator
| Decision | What the forecast says |
|---|---|
| Do you have a queue position or an agreement? | These are different assets and the gap between them is the subject of this report. A position is an option with an 80.8% historical failure rate behind it. An executed agreement with energised capacity is one of 128 in one operator's whole system. Know which you hold before you price it. |
| What is your real lead time? | Add three numbers, not one: the study and agreement time, the equipment delivery slot, and the construction time. One operator's median from request to in-service for projects that finished is 3.2 years, and 40% of one manufacturer's equipment backlog sits beyond two years. |
| Should you be in the queue at all? | If your load can be served behind the meter or beside an existing plant, the queue is a choice rather than a requirement. You pay more for energy and less for time, and in a rationed market time is what is scarce. |
| What will you give up to go first? | Curtailable service is the cheapest date available and almost nobody can use it. If your process tolerates interruption, say so early and in writing, because it is the strongest negotiating position in a short market. |
| Do you own idle electrical service you are not using? | A closed or shrunken industrial site with live service is a scarce asset in this market, and most owners of one have it on the books as land and a building. Establish what the service is worth before you sell the site that carries it. |
| How exposed are you to a rule change? | Six regional operators have been ordered to rewrite their large-load rules and at least one state has ordered an audit. Terms agreed in 2025 may not survive to 2028. Re-confirm dates, charges and curtailment obligations in writing, annually. |
If you underwrite or invest in these projects
Underwrite the agreement, not the queue. A diligence pack that quotes a queue total without the withdrawal rate alongside it is quoting a number with an 81% failure rate. The operator's own feed publishes both and the count takes minutes.
Price the equipment slot separately from the project. A reserved manufacturing slot is a transferable, scarce, cash-settled asset, and it is increasingly the thing that sets the date. Ask whether the slot deposit is paid, who holds the contract, and what happens to it on a change of control.
Treat the political term as a real one. Total capacity cost rose seven and a half times in five delivery years — FD estimate: $16.4bn for DY 2028/29 ÷ $2.2bn for DY 2023/24 = 7.5 times, which is the bill, not the eleven-fold rise in the per-MW-day clearing price over four delivery years. That lands on voters, and the response arrives before the rate case. A project whose economics need a tariff to stay unchanged through 2028 is carrying a risk nobody has priced.
Frontier Desk is the advisor to the company that owns the site, the load or the generator, never a utility and never a broker. The first conversation covers what you actually hold, what a buyer or a lender will make of it, and what to establish before you commit capital: Get your company ready.
See who buys and funds sites, load and generation like yours.
A queue position, an executed agreement and an energised site are three different assets with three different buyers. A short conversation covers which one you hold and who is looking for it right now.
The evidence base
The present-state record the forecast rests on, as a reference band. Everything here is a reading from a named public source or a count over a public dataset, with its date and its arithmetic; the inferences are in the bands above.
The queue, counted rather than quoted
Publicly observable| Measure | What it counts | Figure | What it tells the forecast |
|---|---|---|---|
| Records in the feed | Every interconnection request the operator has ever published | 4,061 | The denominator for everything below |
| Active | Requests still in the process | 1,091 requests, 254,591 MW summer net | The number usually quoted as demand |
| Withdrawn | Requests abandoned by their sponsor | 2,325 requests, 395,238 MW | More capacity has been withdrawn than is currently active |
| Done | Requests that completed the process | 553, including 5 legacy records | The numerator of the completion rate |
| Completion rate | Done ÷ (Done + Withdrawn) | 19.2% — FD estimate: 553 ÷ 2,878 | What a queue megawatt has historically been worth |
| Implied capacity | Active MW × completion rate | About 48,881 MW — FD estimate: 254,591 × 19.2% | The queue, discounted by the operator's own record |
| Under construction after a signed agreement | Projects past the agreement and building | 128 projects, 20,285 MW — 8.0% of active MW | What is actually being built right now |
| In service | Projects energised, as flagged in the feed | 92 projects, 14,243 MW | What finished |
| Median request-to-service time | For 512 of the 548 Done records carrying both dates | 3.2 years | The lead time a finished project actually took |
| Median age of the active queue | Time since the queue date, for active requests | 1.6 years; oldest 5.8 years | Most of the queue is recent, which is why its outcome is unknown |
| Requests entering in the last 90 days | Active requests with a queue date in that window | 222 requests, 68,681 MW | The inflow is still running at scale |
| Largest fuel types by active capacity | Summer net MW by stated fuel type | Not stated 103,425; solar 47,881; battery 39,641; gas 25,400; wind 24,547 | Two fifths of the active queue does not state a fuel |
Every row is Frontier Desk's own count over the operator's public project feed, downloaded on 5 November 2026 and summed as described in band 15. The feed's own field names are used unchanged: applicationStatus for the status column, summerNetMW for capacity, postGIAStatus for construction state. Anyone can repeat the count from the same URL; the method is in Sources and method and the URL is in the sources list.
The funnel, in megawatts
Publicly observableOne operator's public queue, counted by Frontier Desk on 5 November 2026. Log scale, because the top bar and the bottom bar differ by more than an order of magnitude
Source: Frontier Desk's own count over MISO's public project feed, downloaded 5 Nov 2026
What the capacity market has said, four auctions running
Publicly observableCapacity clearing price by delivery year
Publicly observablePJM Base Residual Auction resource clearing price in the RTO, dollars per MW-day, from the operator's own results report
Source: PJM's own Base Residual Auction results report, read 5 Nov 2026
What that capacity costs, by delivery year
Publicly observableCleared supply times clearing price, in billions of dollars, from the same reports. The 2028/29 auction produced 524.7 MW of new generation and uprates
Source: PJM's own Base Residual Auction results report, read 5 Nov 2026
What the planners say, and what they revised
Publicly observable| Reading | Source | Figure | What it tells the forecast |
|---|---|---|---|
| Ten-year summer peak growth, PJM RTO | PJM 2026 Long-Term Load Forecast, Jan 2026 | 3.6% a year; 222,106 MW in 2036, a 65,733 MW increase | The long-run number, unrevised |
| Ten-year net energy growth, PJM RTO | The same report | 5.3% a year; 1,437,629 GWh in 2036, a 581,554 GWh increase | Energy grows faster than peak, which is what a high-load-factor customer does |
| Near-term revision for the 2028 auction year | The same report | −4,414 MW, −2.6%, against the 2025 forecast | The planner marking down its own near-term number |
| Near-term revision for 2026 and 2031 | The same report | −2,564 MW (−1.6%) for 2026; −1,630 MW (−0.8%) for 2031 | The cut is deepest two years out and shallowest five years out |
| Why the near-term fell | The same report | Firm commitments required near term; longer-term projects non-firm and derated | An evidence standard, stated by the planner as the cause |
| Zones adjusted for data center growth | The same report | AEP, ATSI, APS, BGE, COMED, DAYTON, DLCO, JCPL, METED, PECO, PEPCO, PL, DOM and PS | Fourteen of the RTO's zones, named |
| Ten-year North American summer peak growth | NERC 2025 Long-Term Reliability Assessment | Over 224 GW, 69% above the prior year's 132 GW ten-year projection | The scale of the revision upward, one year earlier |
| Ten-year North American winter peak growth | The same assessment | 245 GW, outpacing summer | Winter is where the reliability risk concentrates |
| Areas facing resource adequacy challenges | The same assessment | 13 of 23, including MISO, PJM, ERCOT, SPP and New England | Where the risk is, by name |
| Announced and confirmed retirements | The same assessment | Over 105 GW over 10 years, about 10 GW less than the prior year's projection | Retirements are slowing, which is itself a response to this demand |
| US data center electricity consumption in 2023 | US Department of Energy, 20 Dec 2024 | 176 TWh, about 4.4% of US electricity, from 58 TWh in 2014 | The last authoritative measurement before the current build |
| Its 2028 projection | The same publication | 325 to 580 TWh, about 6.7% to 12% of US electricity | A government range almost two-fold wide, for the horizon of this report |
Three gaps in this evidence base are worth stating plainly, because the forecast above is weaker where they are.
First, the queue counts are from one operator. We chose MISO because it publishes a complete per-request feed that anyone can download and recompute; PJM and ERCOT do not publish their queues in a form that supports the same count, which is why the ERCOT figure in this report comes from named press rather than from our own arithmetic. Second, no public source states what an interconnection position or an energised brownfield site has actually sold for, which is why the central call about $1 million per MW is written to create that evidence rather than to rest on it. Third, the FERC dockets are carried from two independent professional analyses rather than from the Commission's own order text, which we could not retrieve, and they are graded accordingly.
Key assumptions
The assumptions the odds depend on, and the evidence that would break each one. Each is an analytical commitment, not a reading.
| Assumption | Breaks if |
|---|---|
| Historical withdrawal rates are a usable guide to what today's queue will produce. | Data center requests withdrawing at materially lower rates than the generator requests that dominate the historical record, which would roughly double the implied capacity. |
| PJM's near-term cut reflects a durable evidence standard rather than a one-off reclassification. | PJM's next long-term forecast restoring the near-term reduction as projects produce their firm commitments. |
| Equipment lead times stay long enough to bind through 2028. | Equipment remaining performance obligations falling, or the share scheduled beyond two years falling, at two or more major manufacturers. |
| Price does not call forth meaningful new generation inside the forecast window. | A capacity auction producing several gigawatts of new build rather than hundreds of megawatts, which would mean the supply curve is not vertical after all. |
| Co-located and behind-the-meter load keeps a transmission cost advantage. | A FERC or state decision requiring co-located load to pay full transmission charges with no reduction for reduced grid usage. |
| Connected data center load keeps growing faster than other customer classes. | A large-load class at a major utility growing more slowly than its commercial class for two consecutive quarters. |
| The political response stays at the level of tariffs, approvals and audits rather than moratoria. | Any state enacting a moratorium on new large-load connections, which would move build-your-own and demand disappoints up together. |
| Reserve margins below target do not produce firm load shed inside the window. | Any firm load shed attributed to resource adequacy in a major organised market. |
| Curtailable service is offered at a discount large enough to change behaviour. | Tariff filings that price flexible service at or near firm service, which would leave curtailment a reliability tool rather than a market. |
| Interconnection rights become transferable enough to have a market price. | Operators prohibiting or heavily restricting transfer of queue positions and agreements, which would leave the value trapped with the original applicant. |
Signposts
A signpost is a reading that moves the odds if it crosses the threshold shown. Each one names the scenario or scenarios it moves, so no reading on this page floats free of the forecast it is supposed to update.
| Indicator | Now | Threshold | Scenario it moves | What it would mean | Checked |
|---|---|---|---|---|---|
| Withdrawal rate in one operator's public feed | 80.8% of 2,878 resolved requests | Below 70%, or above 85% | Open queue · Rationed access | Below 70% means the queue is firmer than its history and the filters matter less; above 85% means the opposite | Monthly, from the feed |
| Active capacity in that queue | 254,591 MW across 1,091 requests | A fall of 10% or more | Rationed access | The filtering has started to bite, and a position is becoming scarce rather than free | Monthly, from the feed |
| Capacity under construction after a signed agreement | 20,285 MW, 8.0% of active | Above 12%, or below 6% | Rationed access · Demand disappoints | Above 12% means the pipeline is converting; below 6% means it is stalling at the agreement stage | Monthly, from the feed |
| PJM's next near-term load revision | −4,414 MW for the 2028 auction year | A second consecutive cut, or a restoration | Demand disappoints · Open queue | A second cut moves demand disappoints sharply up; a restoration says the first cut was paperwork | PJM's annual load report |
| Planners publishing an explicit derating method | 1 | 3 or more | Rationed access | Derating has become standard practice, and raw queue totals stop being quoted by anyone serious | RTO load forecast reports |
| PJM capacity clearing price | $325.00 per MW-day, at the cap | Clearing below the cap, or new build above 2,000 MW | Open queue · Demand disappoints | Below cap with the shortfall unresolved reads as demand disappointing; new build at scale reads as the physical brake releasing | Annual auction results |
| New generation and uprates cleared per auction | 524.7 MW UCAP | Above 2,000 MW | Rationed access · Reliability event | The supply curve is not vertical after all, and the physical constraint is easing faster than forecast | Annual auction results |
| Equipment backlog scheduled beyond two years | 40% of $104,054m — FD estimate | Below 30%, or above 50% | Rationed access · Build-your-own | Below 30% means the factory constraint is easing; above 50% means it is the binding one through 2028 | Manufacturers' quarterly filings |
| One utility's large-load class revenue growth | +64.8% year on year over nine months | Below its own commercial class growth | Demand disappoints | The first filed evidence that connected data center load is decelerating rather than its queue | Quarterly filings |
| State-level large-load approvals, tariffs or audits | At least one state, ordering an audit | 3 or more states | Build-your-own · Demand disappoints | The political brake has become a national constraint rather than a Texas one | State legislatures and commissions |
| Co-located load rules | Six show-cause proceedings open, one co-location docket on rehearing | A final rule setting a reduced transmission charge | Build-your-own | Leaving the queue is cheaper by an amount a regulator has now fixed, and self-supply accelerates | FERC dockets and named analyses |
| Firm load shed attributed to resource adequacy | None in the current window | Any instance in a major market | Reliability event | The tail has arrived, and rules change faster than any forecast on this page | NERC and operator event reports |
| Disclosed sale of an interconnection position or energised site | None on the public record | Any disclosed price per MW | Rationed access | The first number the market has ever had for the asset this whole report says is scarce | SEC filings and named press |
| US average retail electricity price, all sectors | 14.87 cents per kWh, from 14.15 | Above 16.0 cents | Demand disappoints · Build-your-own | The political brake tightens, and large-load-specific cost allocation becomes the main regulatory question | EIA Electric Power Monthly |
Calls
Each call is dated, carries a Frontier Desk judgment as a percentage, and names the public source that settles it. No call on this page is settled by a social post, an unnamed source or a private record. We score each one when it resolves, with the Brier score: the squared gap between our odds and what happened, where 0 is perfect.
| Call | Resolves | Odds | Settled by |
|---|---|---|---|
| A large-load-specific tariff or study process is accepted at one or more RTOs | 31 Dec 2026 | 76% | FERC dockets and the RTOs' own tariff filings |
| A second planner publishes an explicit large-load derating method | 30 Jun 2027 | 64% | RTO long-term load forecast reports |
| A grid operator publishes requested against firm large load side by side | 31 Dec 2027 | 58% | RTO load forecast reports and planning documents |
| PJM's next long-term load forecast cuts the near term again | 31 Dec 2027 | 51% | PJM's annual Long-Term Load Forecast |
| Active capacity in MISO's published queue falls 10% or more from 254,591 MW | 31 Dec 2027 | 44% | MISO's own public project feed |
| The withdrawal rate in that feed stays above 70% of resolved requests | 31 Dec 2028 | 79% | MISO's own public project feed |
| Capacity under construction after a signed agreement exceeds 12% of active queue capacity | 31 Dec 2028 | 36% | MISO's own public project feed |
| Equipment RPO scheduled beyond two years keeps rising at a major turbine manufacturer | 31 Dec 2027 | 61% | SEC EDGAR, 10-Q and 10-K |
| A reserved generation-equipment delivery slot is disclosed as transferred or resold | 31 Dec 2028 | 38% | SEC EDGAR and named press |
| A PJM capacity auction clears below its price cap | 31 Dec 2028 | 39% | PJM's own auction results reports |
| A PJM capacity auction procures more than 2,000 MW UCAP of new generation and uprates | 31 Dec 2028 | 33% | PJM's own auction results reports |
| Three or more states adopt a large-load-specific approval, tariff or audit requirement | 31 Dec 2028 | 72% | State legislatures and public utility commissions |
| A US interconnection position or energised brownfield site changes hands above $1 million per MW, disclosed | 31 Dec 2028 | 57% | SEC filings and named press |
| Firm load is shed in a major organised market and attributed to resource adequacy | 31 Dec 2028 | 25% | NERC and grid operator event reports |
| A major utility's large-load class grows more slowly than its commercial class for two consecutive quarters | 31 Dec 2028 | 29% | SEC EDGAR, 10-Q |
| US average retail electricity price, all sectors, exceeds 16.0 cents per kWh | 31 Dec 2028 | 47% | EIA Electric Power Monthly |
| Power market: open queue / rationed access / build-your-own / demand disappoints / reliability event | Jan 2028 / Jan 2029 | 34/29/18/11/8 · 13/41/21/17/8 | Triggers and markers in Scenarios |
| What sets a project's date: interconnection / equipment slot / on-site generation / flexible-load agreement | Jan 2028 / Jan 2029 | 45/24/18/13 · 27/29/26/18 | RTO filings, manufacturers' filings and project announcements |
Questions to test this
Questions a sharp reader would use to break the argument.
- 01An 80.8% withdrawal rate is measured over a history dominated by merchant renewables. What would it take to measure it separately for data center load, and does any operator publish the fields that would let an outsider do it?
- 02PJM cut its 2028 peak by 4,414 MW by reclassifying projects as non-firm. If every one of those projects produces an Electric Service Obligation next year, was the cut information or accounting — and how would a reader tell at the time?
- 03A capacity price that rises eleven-fold produces 525 MW of new build. If price cannot call forth supply inside five years, what is the capacity market actually procuring, and who is paying $16.4 billion for it?
- 04One utility's large-load class grew 64.8% while its queue is said to be speculative. Which of the two is the better forward indicator, and has anyone published a back-test?
- 05If a reserved turbine slot now sets the date more often than an interconnection agreement does, should it be a tradable instrument — and what stops a secondary market in manufacturing slots from existing already?
- 06Co-location is a way around the queue and not around the factory. At what share of new large load does self-supply start to lengthen everyone else's equipment lead time rather than shorten their queue?
- 07The political brake arrived in Texas as an audit before it arrived as a tariff. Which state writes the first large-load cost-allocation rule that other states copy, and what does the copied rule most likely say?
- 08Every scenario here except one assumes nothing breaks. With 13 of 23 assessment areas flagged and PJM clearing 6,831 MW below its own requirement, is an 8% chance of a reliability event through 2028 defensible, or complacent?
Sources and method
Where the direction comes from
Direction comes from Frontier Desk's standing coverage of compute and the build-out behind it — who sells capacity, who prices it, who buys it, and who owns the physical build. The public surface of that work is at /compute and /markets.
The standard every claim had to clear
A figure, a quotation or a market fact is carried only if an independent primary or reputable named source states it — a grid operator's own published data or results report, a filing with a securities regulator, a government publication, a reliability body's own assessment, or named press and professional analysis quoting a named official or docket — and is dropped otherwise. Every source on this page was opened and read on 5 November 2026. No figure on this page comes from a private record, from a company Frontier Desk advises, from any piece of work it is doing, or from an unnamed source.
How the queue was counted
The queue figures are Frontier Desk's own counts, not quotations. We downloaded MISO's public generator interconnection project feed on 5 November 2026, which returns one JSON record per request, and grouped it on three of its own fields: applicationStatus for Active, Withdrawn and Done; summerNetMW summed for capacity; and postGIAStatus for Under Construction and In Service. Elapsed time uses each record's queueDate and inService fields over a narrower base: the 548 records whose applicationStatus is Done, excluding the five carrying the legacy Done marker, where both fields are present and the gap is strictly positive and under twenty-five years. That leaves 512 of those 548, excluding 36 with implausible, missing or same-day dates, and their median is 3.2 years. The feed URL is in the sources list and the counts can be reproduced from it in a few lines of code.
How a reading was graded
Processing graded every reading A for a record, B for a party's own published statement about itself, and C for a crowd source. Two readings are graded C and attributed in the sentence that carries them: the ERCOT queue total, which comes from named press reporting the operator and the Governor of Texas rather than from a dataset we could count; and the FERC show-cause dockets, which come from two independent professional analyses published the same day rather than from the Commission's own order text.
How the readings became a forecast
Analysis turned the readings into structural shifts, a causal model, phases, scenarios and odds, and every judgment on this page carries its own confidence and the reason for it, kept visibly separate from the grade on the evidence under it. Where a figure is derived rather than quoted it is marked FD estimate and its arithmetic is shown beside it.
What the odds are, and what they are not
The odds on this page are Frontier Desk judgments at two fixed horizons, not the output of a model, and the page says so wherever they appear. They are internally disciplined rather than simulated: each column sums to 100, a call that is a scenario arriving is held at or below that scenario's odds, and a call that is a necessary marker of a scenario is held at or above it. They are intended to be scored and to be wrong in public.
What comes back, and when
Dissemination is this page. Feedback is the signposts, each of which names the scenario it moves, and the Brier-scored calls. The queue counts are recomputed from the same feed before every update. The next update is 5 December 2026.
What we dropped, and why
Several claims that would have strengthened this report did not clear the standard above, and are not on the page in any form.
Aggregated national queue totals
Widely quoted figures for total US interconnection queue capacity come from compilations we could not reproduce from the underlying operator data in this run. The queue arithmetic on this page is confined to the one operator whose complete per-request feed is public, and the report says so each time it generalises.
Data center capacity figures from market trackers
Commercial trackers publish megawatt totals for data centers under construction by market. Their methods are not published and we could not check them, so none of them is on this page. Where capacity figures appear they come from a registrant's own filing, such as the 210 MW campus and 1,410 MW of aggregate contracted critical IT load in our companion compute-financing forecast.
Lead times quoted by developers
Interconnection and turbine lead times quoted in interviews and conference panels are not checkable by a reader. The lead times on this page are either computed from the operator's own dated records or read off a manufacturer's own disclosed backlog timing.
Prices for queue positions
Figures circulate for what an interconnection position or an energised site has sold for. We found no disclosed transaction stating one and have carried none, which is why the central call on this subject is written to be settled by a future disclosure.
What we still cannot verify
We could not retrieve the Commission's own text for the six show-cause orders, could not obtain ERCOT's large-load queue as a countable dataset, and could not find any public source stating a realised price per megawatt for an interconnection position. Each of those gaps is named where it bears on a judgment.
- 01The interconnection queue as a countable dataset: one JSON record per request, with applicationStatus, summerNetMW, postGIAStatus, queueDate and inService fields. Downloaded and counted by Frontier Desk on 5 Nov 2026, giving 4,061 records, 1,091 active for 254,591 MW, 2,325 withdrawn for 395,238 MW, 553 done, 128 under construction for 20,285 MW, and a 3.2-year median request-to-service time over 512 of the 548 Done records: MISO's public project feed; the interactive queue it serves
- 02The capacity auction, in the market operator's own results report: 138,317.8 MW UCAP cleared at the $325.00 price cap for delivery year 2028/2029, total procured capacity “6,831.3 MW UCAP below the RTO Reliability Requirement”, a 14.4% installed reserve margin against a 20% target, “the second BRA in a row where PJM has cleared more than one percentage point below the IRM”, total new generation and uprates of 524.7 MW UCAP, and cleared supply times clearing price of $16.4 billion against $2.2 billion for delivery year 2023/24 and $28.92 per MW-day for 2024/25; read 5 Nov 2026: PJM 2028/2029 BRA results report
- 03The planner marking down its own near-term forecast: the 2026 Long-Term Load Forecast “is lower than the 2025 Long-Term Load Forecast in the near term through 2032”, with summer peak changes of −2,564 MW (−1.6%) for 2026, −4,414 MW (−2.6%) for the 2028 auction year and −1,630 MW (−0.8%) for 2031, because near-term years “need ‘firm' commitments, such as Electric Service Obligation (ESO)/Construction Commitments (CC), while longer-term projects will be considered ‘non-firm' and will be derated”; alongside ten-year growth of 3.6% a year in summer peak and 5.3% a year in net energy; dated January 2026, read 5 Nov 2026: PJM 2026 Long-Term Load Forecast
- 04The reliability assessment: “New data centers for artificial intelligence and the digital economy account for most of the projected increase in North American electricity demand over the next 10 years”; aggregated summer peak “forecast to rise by over 224 GW … 69% higher than last year's 10-year growth projection of 132 GW”, winter peak by 245 GW; 13 of 23 assessment areas facing resource adequacy challenges; over 105 GW of announced retirements; and the countervailing finding that “some large-load projects have slowed or failed to materialize within the shorter-term horizon, while interconnection requests for later years continue to increase”; read 5 Nov 2026: NERC 2025 Long-Term Reliability Assessment
- 05US data center electricity consumption of 176 TWh in 2023, “about 4.4% of total U.S. electricity”, from “58 TWh” in 2014, projected at “325 to 580 TWh” or “approximately 6.7 to 12% of total U.S. electricity” by 2028; published 20 Dec 2024, read 5 Nov 2026: US Department of Energy
- 06The large-load class that has already connected: regulated electric sales to “customers in Virginia, including certain data centers, with actual or anticipated forecast demand of 25 MW or higher and annual load factor of 75% or higher” of $706m in the quarter and $2,156m for nine months to 30 Sep 2026, against $503m and $1,308m a year earlier, within Virginia Power's $11,153m of nine-month revenue from contracts with customers; read 5 Nov 2026: Dominion Energy Form 10-Q
- 07The factory constraint, on a manufacturer's own cash flow statement: a $19.5bn nine-month inflow from contract liabilities and current deferred income “driven by down payments on orders and slot reservation agreements at Power”, against $2.4bn in the same period of 2025; remaining performance obligation of $196,721m including equipment-related RPO of $104,054m “of which 33%, 60%, and 97% is expected to be recognized within 1, 2, and 5 years”; quarter ended 30 Sep 2026, read 5 Nov 2026: GE Vernova Form 10-Q
- 08Wholesale power prices: “Wholesale electricity prices in the United States average $52 per megawatt hour in 2026 in our forecast, 11% more than in 2025”, with PJM prices expected to rise 41%; released 6 Oct 2026 with inputs finalised 1 Oct 2026, read 5 Nov 2026. The link is to the archived October release rather than the STEO landing page, because the next release overwrites the landing page on 10 Nov 2026: EIA Short-Term Energy Outlook, October 2026
- 09Retail electricity prices for August 2026, preliminary, against August 2025 final: all sectors 14.87 against 14.15 cents per kWh, residential 18.16 against 17.57, commercial 14.47 against 13.87, industrial 9.64 against 8.98; read 5 Nov 2026: EIA Electric Power Monthly, Table 5.3
- 10The Texas large-load queue and the audit ordered against it: “474 GW of interconnection requests”, of which “[a]pproximately 90 percent of the new power requests are data centers”, and the Governor's statement that “[a]ny data center project that fails to comply with the verification and audit process to protect the reliability and resilience of the Texas electric grid must be denied”, alongside ERCOT's 3 Aug 2026 delay of its Batch Zero study; reported 5 Aug 2026 by Robert Walton, read 5 Nov 2026: Utility Dive
- 11The federal regulator's actions of 18 Jun 2026, as described by two independent professional analyses both published 23 Jun 2026: six Section 206 show-cause proceedings at Docket Nos. EL26-67-000 through EL26-72-000, one per RTO/ISO, with 60 days to justify or revise large-load tariff provisions and 30 days to file an informational report on ensuring adequate generation; plus an order on rehearing, compliance and paper hearing in the PJM co-located load docket, EL25-49-002; read 5 Nov 2026: Duane Morris; Holland & Knight
- 12Where power sits in the financing of the build-out, including the observation that “[p]ower availability is increasingly shaping both where data centers are built and how they are financed” and the federal-land examples; a trade body note of 23 Jul 2026 attributing Morgan Stanley and Barclays research, read 5 Nov 2026: Structured Finance Association
- 13A 15-year lease of 210 MW of critical IT load to a high investment-grade hyperscaler for about $5.2bn of contracted revenue with delivery expected in the first half of calendar 2028, and 1,410 MW of aggregate contracted critical IT load; the one capacity figure on this page taken from a registrant's own filing rather than a tracker; quarter ended 31 Aug 2026, read 5 Nov 2026: Applied Digital Form 10-Q
- 14Our own public work on how a lender reads a compute project, which this forecast sits beside; each page checked 5 Nov 2026: Finance a GPU cluster; How lenders underwrite a GPU cluster loan; The markets we map
- 15Our companion forecast on the financing side of the same build-out, published 5 Nov 2026 and read the same day: GPU and neocloud compute financing to 2028
Find out what your site, your load or your position is actually worth.
Frontier Desk is the advisor to the company that owns the asset. We read a site the way a buyer, a lender or a hyperscaler will — the service already behind it, the agreement or the position, the equipment, and the terms that could change — and tell you what you hold and what to establish before you commit capital. Answer what you have and the first conversation covers the rest.
If you own an industrial site with electrical service you no longer fully use, that service may be the most valuable thing on the deed. The same review establishes what it is and who would want it.
Four in five requests were withdrawn, and the fifth took three years. Everything else in this market is a consequence of those two numbers.
Frontier Desk outlook, 5 November 2026. Frontier Desk acts as the advisor to the company that owns the site, the load or the generator, through its relationships with buyers, lenders and investors, and a cybernetic system of its own — which is why the forecast stays current. Frontier Desk is not a broker-dealer, lender, utility or investment adviser. Every figure is quoted from a named public source read on the date stated, or marked as a Frontier Desk estimate or count with its arithmetic shown. The odds are Frontier Desk judgments, not model output, and may be wrong. Nothing here is an offer, a recommendation to buy or sell any security, or advice on any particular project. Companies and grid operators are named because their public filings and published data are the record this market has, not because they are connected to Frontier Desk in any way.