From June 2027, the operator of America’s largest electricity grid will be able to cut power to data centres when supply runs short — putting them behind households and ordinary businesses in the queue.
PJM Interconnection serves 67 million customers across thirteen states and Washington, D.C. Its board proposed the curtailment framework on 27 July 2026. It is a reversal: for a decade, grid operators competed for data-centre load. PJM has now decided it has more than it can carry.
What is being proposed
Two connected measures:
A backstop capacity auction, proposed 27 July, to procure generation the regular market has failed to deliver.
The Interim Resource Adequacy Service (IRAS), filed with the Federal Energy Regulatory Commission on 13 August. Under it, new “large loads” — overwhelmingly data centres — must either bring their own generation or accept lower priority than existing customers during emergencies.
Curtailment applies to facilities of 50 megawatts or larger. Affected operators would receive between thirty minutes and several days of notice, and compensation.
The board’s justification was blunt: “Existing consumers should not bear higher capacity costs caused by new large loads” that do not bring or contract for new supply.
The number that forced the issue
PJM’s June 2026 capacity auction cleared at roughly $16.4 billion — about eight times the price of a few years earlier.
Capacity auctions are how the grid pays generators to guarantee availability at future peak demand. That cost is passed to consumers. An eightfold increase is a bill arriving in millions of homes.
Joseph Bowring, PJM’s Independent Market Monitor, attributed a specific share to a specific cause: data-centre load has “increased capacity costs by $29.4 billion over the last four auctions.”
Even at those prices, supply did not arrive. The most recent auction fell about 6.8 gigawatts short of PJM’s 20 per cent reserve-margin target — the safety buffer above expected peak demand. High prices were meant to call forth new generation. They called forth less than required.
The scale of what is coming
PJM forecasts 30 to 34 gigawatts of new large-load demand by the early 2030s, potentially reaching 70 gigawatts by 2038. Data-centre electricity demand overall could rise roughly fourfold by 2035.
Seventy gigawatts is national-scale consumption arriving inside one regional grid, in about a decade.
The mismatch is one of clock speeds. A large data centre can be built in one to two years. A gas plant takes three to five, a transmission line five to fifteen, nuclear longer still. Demand can be added faster than supply can be built, and no market design changes those construction times.
That is the whole problem, and it is why the response is rationing rather than procurement.
Who is expected to pay
The dispute underneath is about cost allocation.
Grid costs have traditionally been socialised: everyone pays through rates, on the theory that everyone benefits from a reliable system. That works when demand growth is incremental and broadly distributed. It works less well when a handful of very large, very concentrated new loads drive most of the increase.
PJM’s answer is that new large loads should carry the costs they cause. The administration’s parallel “Ratepayer Protection Pledge” pushes in the same direction, requiring data-centre operators to build, bring or buy new generation.
Kent Chandler of the R Street Institute defends curtailment as the “only defensible and sustainable way” to protect consumers from cost-shifting.
Where the disagreement is
Clara Summers of the Citizens Utility Board argues the plan does not resolve who is actually responsible, and that state regulators are not ready either: “no state that I know of yet is completely prepared.”
Tom Rutigliano of the Natural Resources Defense Council is sceptical that the final proposal changes much, saying PJM ended up “not too far off from where they started.”
Julia Hoos of Aurora Energy Research questions the sufficiency of the supply targets: the procurement is ambitious, but “it’s nowhere near close enough if all this large load shows up.”
Notice these are not the same objection. One says the plan misallocates responsibility; one says it is cosmetic; one says it is directionally right but too small. They are compatible, and together they suggest a measure that buys time rather than solving anything.
The diesel question
Data centres keep diesel generators for backup. Federal rules allow those to run up to fifty hours a year for demand-response events.
So curtailment does not necessarily mean a data centre goes dark. It may mean it switches to diesel — moving the emissions and the air quality impact from a distant power plant to the local neighbourhood, which is frequently not the neighbourhood benefiting from the facility. Environmental groups have raised this directly, and it is a real cost that does not appear in the capacity-price arithmetic.
Why this is a national story
PJM is first because it has the most data centres, particularly in northern Virginia. It will not be the last. ERCOT in Texas, MISO in the Midwest and the western grids face versions of the same arithmetic.
The wider point is that the AI build-out has run into a physical constraint that capital cannot dissolve. Money can buy chips quickly. It cannot buy transmission lines quickly, because the limit is permitting, supply chains and construction time.
PJM’s proposal is what happens when that constraint becomes binding: the grid stops treating new industrial load as an unqualified good and starts asking it to pay its own way, or wait.
The timeline
- 13 August 2026 — IRAS filed with FERC, with a requested 60-day review.
- 30 September to 21 October 2026 — backstop capacity auction window, targeting new supply for the June 2028 delivery year.
- June 2027 — data-centre curtailment authority takes effect.
- Early 2030s — the 30–34 GW of forecast large-load growth arrives, or does not.
What a capacity auction is, and why the price moved so far
The mechanism is unfamiliar to most people who pay for its results, and the eightfold price move is unintelligible without it.
An electricity market pays for two distinct things. Energy is the electricity actually delivered. Capacity is a commitment to be available at some future peak, whether or not the plant ends up running.
Capacity exists because a grid must meet its highest demand hour of the year, not its average. A plant that runs only during a handful of summer afternoons cannot recover its costs from energy sales alone, and would close — leaving the grid short precisely when it is most stressed. The capacity auction pays for that standby availability.
Prices in such an auction are set by the balance between the required commitment and the supply offered. When supply comfortably exceeds requirement, prices sit near the cost of the marginal plant. When it does not, prices rise sharply toward the administrative cap, because there is no competitive discipline left.
PJM’s June 2026 auction cleared at roughly $16.4 billion — about eight times a few years earlier — and still fell about 6.8 gigawatts short of its reserve-margin target. That combination is the diagnostic. High prices with unmet requirement means the shortage is not a pricing problem. Supply did not respond because supply cannot be built on auction timescales.
Why generation cannot simply be built
The obvious response — build more plants — runs into constraints that money does not resolve quickly.
Interconnection queues. Connecting a new generator requires studies of its effect on the existing network. Queues across US grid operators have run to years, and PJM has been among the most congested.
Equipment lead times. Large gas turbines and high-voltage transformers are made by a small number of manufacturers with order books measured in years. A transformer ordered today may arrive after the delivery year it was meant to serve.
Transmission. Generation in the wrong place is not useful. New lines require siting approval across multiple jurisdictions and routinely take a decade or more, with local opposition a reliable feature.
Permitting. Environmental review, air permits and local approvals add years independent of construction.
Against that, a hyperscale data centre can be sited, built and energised in roughly one to two years. The asymmetry is the entire problem, and it is why an operator facing 30 gigawatts of forecast new load reaches for demand-side measures rather than supply-side ones.
Where the electricity actually goes
A useful corrective to the framing: AI training runs are not the whole story, and treating data-centre load as synonymous with AI overstates the case.
A large facility’s draw is dominated by two things — the servers themselves and the cooling required to remove the heat they produce. Cooling can account for a substantial share of total consumption, which is why operators site facilities in cool climates and invest heavily in cooling efficiency.
The load also includes everything else the internet runs on: cloud services, storage, streaming, corporate systems. AI has accelerated the growth curve sharply, but it is added on top of a base that was already rising.
The distinction matters for policy. A curtailment framework aimed at “AI data centres” would be unworkable, because a facility does not partition its power draw by workload type. PJM’s rule applies by size — 50 megawatts and above — because size is the only thing a grid operator can actually observe and act on.
What this means for the people paying the bills
The consumer question is simpler than the market mechanics: does this stop bills rising?
Partly, and prospectively. The framework applies to new large loads. The $29.4 billion in additional capacity costs that PJM’s market monitor attributes to data-centre demand over four auctions has already flowed into rates. The measure is designed to prevent the next increment, not to reverse the last one.
It also shifts a cost rather than eliminating it. If new facilities must bring their own generation, that generation still has to be built, and its cost still lands somewhere — in the price of the compute, and eventually in the price of the services running on it. The policy question is who carries it, not whether it exists.
Sources
- Utility Dive, “PJM board proposes backstop capacity auction, data center curtailment plans,” 28 July 2026 — utilitydive.com
- PJM Inside Lines, “PJM Proposes Framework To Connect Data Centers Without Compromising Reliability, Affordability,” 13 August 2026 — insidelines.pjm.com
- TechCrunch, “Data centers may face temporary power cuts to prevent blackouts on largest US grid,” 28 July 2026 — techcrunch.com
- Canary Media, “PJM’s big new data center plan: Make the states figure it out,” August 2026 — canarymedia.com
