Chapter 1
American electricity demand barely moved for twenty years
American electricity demand barely moved for twenty years. From 2005 to 2019 it grew about a tenth of a percent a year, and the machinery built around that fact — the auctions that procure generating capacity, the rate cases that decide who pays for wires, the reliability standards that govern what may be connected — assumed a system that would grow slowly and predictably if it grew at all. Since 2020 demand has grown about 1.7 percent a year, and data centres are the primary reason. That single reversal is the subject here: not what is being built, but what the electricity system does about it, and whose bill absorbs the difference.
Chapter 2
The strain showed first in price
The strain showed first in price, then in cost allocation. PJM's capacity auction — the market that pays generators to be available years ahead — has now cleared at its administrative price ceiling three times running, and even at the ceiling the most recent one fell 6.8 gigawatts short of the supply its own reliability standard requires; only 525 megawatts of new generation cleared. PJM's results report simulates that without a cap it would have cleared at $554.72 a megawatt-day region-wide and $776.69 in the Chicago area. Monitoring Analytics, the independent monitor of PJM's markets, attributes about $6.3 billion of that auction's $16.4 billion in charges to data-centre demand, and its president puts the cumulative figure at $29.4 billion over four auctions. That is the wholesale half. The retail half travels a quieter route: a Harvard Electricity Law Initiative review of nearly fifty rate proceedings found utilities using rate structures and non-public contracts to move the cost of infrastructure built for data centres onto ordinary customers, largely out of public view. By the end of 2025 the average residential price had reached about 19 cents a kilowatt-hour, some 27 percent above 2019; utilities requested more than $29 billion in increases in the first half of 2025 alone, double the year before, and households owed a record $25 billion in unpaid utility bills. Individual operators kept promising that none of it was theirs — a claim that keeps being made and has not yet been settled by any tariff or commission ruling. How much of that rise data centres actually caused is, on the backward-looking record, smaller than the wholesale numbers imply: Congress's research service, surveying the studies in September 2026, reports that the main driver of price increases from 2019 to 2025 was utility investment in ageing and hardening infrastructure, and that the states with the fastest data-centre growth generally saw prices fall — the extra demand spread fixed costs across more sales. The forward-looking models disagree with the rear-view mirror, projecting national increases of 6 to 29 percent as the load arrives.
Chapter 3
Two things then changed the terms
Two things then changed the terms. The first was physical. A transmission fault in Ashburn, Virginia caused hyperscale facilities to transfer themselves to backup power, removing roughly 3.8 gigawatts from the PJM system in seconds, on the reliability regulator's preliminary count, — and revealed that no reliability standard governs how a load that large is required to behave. Nobody shed it; the data centres left on their own protection schemes, which is how a disappearance of that size and "no load was shed" can both be true. The second was institutional, and followed directly. Federal regulators ordered the North American Electric Reliability Corporation to write mandatory standards for "computational loads" and criteria for registering the entities behind them by the end of 2026, with a second phase due the following March; NERC's first proposal would have registered any entity with at least 20 megawatts of connected load at a single high-voltage interconnection and at least a megawatt of computation behind it — a threshold reaching owners, tenants, colocation providers and hosts alike, without saying which of them carried the obligation. In August it revised that: two registrations, a Computational Load Owner and a Computational Load Operator, so the duty has a named holder, and — as compliance advisers read the revision — a threshold lifted to 50 megawatts. Weeks later PJM's board directed its own filing: a registry, a backstop procurement, and an interim service that would cut new large load within ten minutes of instruction, ahead of the emergency measures that reach everyone else — and at the end of July PJM lodged the backstop procurement as an actual case at FERC (docket ER26-3380-000), fixing the calendar the board had left open: a one-time auction bid this autumn, results by early December, accepted supply capped at $555 a megawatt-day against the $325 ceiling that bound the base auction. What began as directions to write rules has become filings and rulings. At the end of September federal regulators accepted the PJM procurement but suspended it until the end of February 2027, sent the question of who pays for it to a hearing and invited PJM to try again; the auction did not open. NERC's board approved the two registrations in October, subject to federal approval, and its reliability standards passed their final ballot days later without a numeric ride-through requirement; neither is in force. The same month FERC accepted PJM's registry and a 50-megawatt definition of a large load but rejected the interim service that would have curtailed new data centres first, holding that who is cut in a shortage is a retail question for the states.
Chapter 4
Three questions are open
Three questions are open, and they are the ones the record will settle. The first is who pays. Nothing in the federal architecture assigns retail costs to a named customer — PJM's board said so in terms, and said state action would be essential — so the allocation is being decided commission by commission, with the operator's own filing spelling out the default if a state does nothing: the cost falls on all load in the zone, including load that is not large, for fifteen years. The first federal ruling on the question came on that very filing: regulators suspended it on the ground that PJM had not shown its allocation charges the customers who cause the cost. Congress has tried and not answered — the House voted 417–3 in September 2026 for a bill asking states to consider making large loads pay, and two weeks later the Senate blocked it, 57–43, the opposition coming from senators who wanted it to require rather than ask. The first commission to answer in the affirmative is Virginia's, which ordered Dominion Energy to write a tariff charging data centres and other large loads directly for the transmission their arrival requires — a network or substation upgrade that would not have been built but for a large customer is assigned to that customer — building on the 25-megawatt GS-5 rate class it created for January 2027. It is only a direction to develop the tariff, and the commission conceded it "may not address all instances", but it is the first concrete state instrument pointing the buildout's wire costs back at the buildout. The second is whether any of this produces measurement. The first registry is now in force: PJM's has utilities report each large load's location, ramp schedule and supply to the operator and regulators — figures no American operator publishes — but only totals will be made public, so how much the new regime knows and how much the public learns are now separate questions. The third is whether the demand is real. The five-year national peak forecast has risen more than six-fold in three years, to 166 gigawatts, and the same analysis that reported it cautioned that utility filings likely overstate 2030 data-centre load by roughly 40 percent; PJM raised its ten-year outlook and cut its near-term peaks in the same revision, on stricter vetting of speculative requests. PJM's own market monitor counts 42 per cent of the data-centre growth behind its latest shortfall as delayed or cancelled. The grid is being rebuilt, and repriced, around a number nobody has yet had to prove.
And now
Who pays for the grid’s emergency fix is being rewritten. In September FERC suspended PJM’s backstop purchase of new supply because the cost would land on all customers. On October 22 PJM showed members a redesign that charges the zones where large loads are growing, in proportion to their share of that growth. It also drops the opt-out for municipal and co-op utilities. PJM said it would file it by October 29. The purchase would run next spring, with results in May 2027.
The same fight is spreading. On October 23 the Organization of MISO States, backed by 11 of MISO’s 15 state commissions, asked FERC to reject its 120-day fast lane for loads over 200 megawatts. Their reason: nothing stops the transmission costs moving onto existing customers. Google filed the same objection.
The rules themselves are hardening. California’s grid operator approved interruptible interim service for data centres awaiting upgrades on October 28. It will publish what each upgrade costs. NERC’s board adopted the first data-centre reliability standards on October 26, still with no ride-through number. Grid operators’ answers to FERC are due November 16. Virginia’s Dominion–NextEra merger goes to hearing on November 17.
Where it stands today →