America rations the plug: who may be cut, who must hold on, who pays
On 3 August 2026 the Governor of Texas said ERCOT, which runs most of the state’s grid, was weighing over 474 gigawatts of requests to connect. That is more than five times the most power Texas has ever used at once, and about nine tenths of it is data centres.
Twelve days earlier, a fault on a power line in Dominion Energy’s part of Virginia cleared exactly as designed. Even so, about 3,800 megawatts of data centres switched themselves off the grid and onto their backup generators.
One grid has too many data centres asking to come on. The other has data centres that leave when nobody asked them to. America’s answer to both is the same: it is not speeding up the connection, it is rationing it.
The reason is the wires. A firm connection is the grid’s promise to carry a site’s power at all times, short of an emergency. Before giving one, the grid operator lists the new lines, substations and transformers it needs, and those take years. A data centre’s shell goes up far faster.
So grids are starting to sell the connection in tiers. The full tier still waits for the wires. A lower one switches on sooner, but the grid may cut it first when lines are strained, or lets it on only if the site brings its own power.
The second rule points the other way. A site must ride through a disturbance, staying connected through a brief dip or swell in voltage rather than fleeing to its generators. The grid wants data centres to stay on when it wobbles and to come off only when it asks. That makes what backs a site up, a generator, a battery or software that slows its chips, part of the connection itself.
Explain it like I don't work in finance
Picture a block of flats whose fuse box is full, with the electrician booked for years. The landlord still lets new tenants in, on terms: your heater goes off first when the fuse runs hot, or you bring a camping stove.
And when a bulb flickers, you may not yank every plug at once. Nobody gets a bigger fuse sooner. The landlord only decides who goes dark first.
Someone pays for the wait. PJM runs the grid from Chicago to Virginia, and its auction to keep power stations ready for 2028 fell 6,831 megawatts short of its own reliability target. New plants and upgrades added only 525 megawatts. The gap reaches household bills.
A megawatt the grid may cut is not the same product as a firm one. For anyone building a data centre in America, the question is no longer only how much power, but which kind, and on what terms it can be taken away.
For investors
Texas has far more asking than it can connect
The clearest picture of the rationing is the Texas queue, measured at each gate a request must pass. The Governor’s 474 GW is the top. ERCOT told the state Senate on 29 July that only about 205 GW had the studies to enter its first batch. By 20 August, 6,608 MW across 17 sites had cleared every check except approval to switch on.
Texas has been asked for 474 gigawatts, and has cleared 6.6
Requests to connect large loads to ERCOT at each gate, in gigawatts, summer 2026
Then ERCOT paused even that last approval for data centres and crypto mines of 75 MW or more, while it audits every request. Under a seventieth of the queue reached the final gate, and the gate is shut while the audit runs; ERCOT aims to file its reports by 10 December 2026.
What Texas offers instead is a lower plug. A Provisional Controllable Load Resource connects early if ERCOT may reduce it “during periods of localized stress”. A Withdrawal-Limited Private Use Network brings its own plant and takes only a capped amount from the grid. ERCOT is not clearing the queue. It is writing the terms on which most of it may connect.
The regulator is pressing the same split on six grids
The federal regulator is putting the same question to every organised grid it oversees. On 18 June 2026 FERC told six regional grid operators, PJM among them, to defend their rules for large loads or change them. PJM’s tariff “appears to be unjust and unreasonable”, it said, and it asked whether PJM could manage without an interim, non-firm service while network upgrades are built.
The grids are converging on a menu. Southwest Power Pool’s lower tier, CHILLS, was accepted on 5 June: power as available, for up to seven years. On 9 October PJM’s transmission owners told members that capacity “will not be procured” for interim customers until their upgrades are built. ISO New England and NYISO are designing non-firm services, and PJM files its answer in November 2026.
The firm plug is not getting faster. A cheaper one the grid may cut is being added beneath it, grid by grid.
Households pay for the power stations that are not built
Where the plug is rationed, everyone else pays for the shortfall in power stations. PJM serves 67 million people across 13 states and Washington DC. Its capacity auction pays power stations to stand ready in a future year, and the last megawatt needed sets the price for all.
PJM's auction has sat at its price cap for three auctions running
Price paid to power stations to stand ready, in dollars per megawatt of capacity per day, by delivery year
The cap has not called up new plant. The 2028/2029 auction cleared only 525 MW of new plant and upgrades, against a shortfall of 6,831 MW. PJM’s market monitor attributes $6.3 billion, 38.2% of that auction, to data centre load.
The tiers meet the bill here. PJM’s interim tier buys no capacity for a site, so a site the grid may cut adds nothing to the auction until its upgrades land.
Virginia guards against the opposite risk, the site that never fills. Data centres took 28% of Dominion Energy’s Virginia sales in 2025, and from January 2027 its largest customers sign 14 year terms with minimum bills, backed by collateral. Dominion’s stated aim is “to minimize both cross-rate class subsidies and stranded costs”. The floor protects households from the empty site, not the full one.
The sites the grid cannot cut are leaving on their own
The strongest evidence that rationing must run both ways is PJM’s own account of one summer morning. At 07:56 on 22 July 2026 a fault on a 230 kV line in Dominion’s zone cleared properly, with no relay misoperations. Still, about 3,800 MW of data centre load switched to backup generators.
The first wave of transfers raised the voltage, and the higher voltage set off a second. Frequency rose to 60.092 Hz, a sign the grid suddenly had more power than demand.
July's drop was the largest load transfer in PJM's history, and not its first
Data centre load that switched to backup generation after a normally cleared fault, in megawatts
The backup that protects each building is what threw the load at the grid. A data centre is built never to lose power, so at the first flicker it leaves. One site doing that protects its servers. A zone full of them doing it together moved more than seven times the new plant PJM’s last auction added, in one morning, with no warning.
So the grid will now decide when a site may leave
The fix is to make the building hold on, and the rule is being written. PJM proposed on 6 October that sites of 50 MW or more, with at least 25 MW behind power electronics, ride through set dips and swells. A site must stay on for a full second at up to 120% of normal voltage, the kind of swell that set off July’s second wave.
Switching to backup power does not “by itself” count as compliance. New projects must comply from 16 January 2027, and existing sites by 16 January 2028, with limited exceptions where a site shows it cannot. ERCOT’s own rule exempts sites approved before 14 November 2025.
The national rule is following. NERC’s alert of 4 May 2026 “does not create a mandatory obligation”, but FERC has ordered NERC to file binding standards by 31 December 2026. Rationing now has two halves: the grid decides whom it may cut, and forbids the rest to leave on their own.
What backs a site up is now part of its connection
A site that may be cut, or must hold on, needs something to lean on, and the newest answer is the chip itself. By 19 March 2026 Google had deals with five utilities, covering 1 GW, to “limit or shift a portion” of its machine learning work at peak. At an Oracle site in Phoenix, software cut the power of 256 Nvidia chips by 25% for three hours, with no battery.
On 16 September Google, Nvidia and Emerald AI formed an alliance seeking faster connections for “credible, verifiable flexibility commitments”. No public record yet shows a connection approved on a slowed chip.
Batteries were the obvious candidate, and the record says they are mostly doing something else. California’s fleet earns about four times as much from shifting cheap midday power into the evening as from steadying the grid, and on its twelve hardest summer evenings of 2026, gas led every time. That is a time-shifter. Nobody publishes how many batteries sit at data centres, so they stay a line in this story, not a beat.
What could leave the rationing on paper
The case against is that tiers can be written and never used. Developers want firm service, and CAISO’s own proposal records that they “would only agree to be flexible temporarily”. SPP’s CHILLS has been accepted since June, yet no site is shown running on it. And nobody has published what riding through costs, so PJM’s limited exceptions could swallow its rule.
Each objection is a reason to test use rather than paperwork, and the claim below does. It counts only a site of 50 MW or more actually switched on under a tier the grid may cut first, in at least two grids, by the end of 2027. Texas’s power to shed every large load in an emergency does not count, because that is load shedding, not a tier.
For holding on, the test is the outcome. Another transfer of 1,500 MW or more in 2028, after existing sites were due to comply, would show the rule changed nothing on the ground.
Nobody gets a bigger fuse sooner
Texas has more data centres asking than it can carry, and Virginia had data centres leaving that nobody asked to go. Both are answered by terms, not wires. The fuse box is still full and the electrician still booked, so the landlord now writes the tenancy: whose heater goes off first, who must not yank every plug at once, and who pays the bill meanwhile. For a data centre in America, the terms of the connection are now part of the product.
The cut-first plug can outlast the gear queue it waits on
The word “interim” hides a long wait. A site on a cut-first tier waits for the grid’s upgrades, and those upgrades need transformers and switchgear from makers with years of orders on their books. SPP’s CHILLS lets a site take power as available for up to seven years. Siemens Energy’s grid business holds about four years of work, and Hitachi Energy about three.
The cut-first tier's term runs past the makers' order books
SPP's CHILLS term from its acceptance, against each maker's years of work on its books, from its latest filing
So the waiting room’s lease is as long as the queue it waits on. If the seven years run from a site’s own switch-on, which is my reading of SPP’s order rather than its words, a data centre that switches on under CHILLS now could still be cut-first in 2033. In the equipment queue, the gear, not the chip, sets a site’s start date. Here the same gear sets how long a site lives on terms the grid may take away.
That decides who gains. The winners are sites that can live with being cut: chips that can slow, or plant of their own, as in building your own power. The losers are sites whose customers need firm power from the first day.
The load that fell off by accident is the reserve PJM could not buy
July’s morning reads differently once the two rules are put together. About 3,800 MW of data centre load left Virginia’s grid in one morning, by accident. PJM’s latest auction came 6,831 MW short of its reliability target and bought only 525 MW of new plant and upgrades.
The load that left in one morning was seven times the new plant PJM's auction bought
Megawatts: PJM's shortfall and new plant in the 2028/2029 auction, against load that can leave or slow
Load that can leave in seconds is a reserve, if it leaves on the grid’s call rather than at the first flicker. The two rules together make that so: a site must hold on through a fault, and come off when the grid asks. As PJM’s transmission owners describe its interim tier, no capacity is bought for such a site, so it adds nothing to the auction households pay for. PJM’s market monitor put $6.3 billion of the last one on data centre load.
So households are spared that bill, and the site pays: it must own the backup, or chips that slow, as with Google’s 1 GW of deals. The first readings come soon: PJM’s filings to FERC in November 2026 and its 2029/2030 auction in December.
What would prove me wrong
The first test is use, not paperwork. If fewer than two grids have a real site of 50 MW or more on an early tier the grid may cut first by the end of 2027, the rationing is on paper only.
The second is whether holding on works. PJM’s earlier transfers were about 1,500 MW each. If one that size recurs in 2028, after existing sites were due to comply, the rule did not change how the fleet behaves.
Research notes
No initiation note rests on these claims. The power a site brings itself is the subject of the notes on GE Vernova and Bloom Energy. The protection and power gear inside the site are covered by Vertiv, Schneider Electric and Eaton, each resting on its own claim. Whether any of this is already in share prices is a question for these notes, not for the journal.
The claims on record, word for word
The test this piece stands behind, quoted word for word from the claims page:
America rations the plug (30 October 2026)
Claim: “America is rationing the grid connection rather than speeding it up: by the end of 2027, data centres of 50 megawatts or more will be running on an early-connection tier the grid may cut first in at least two of its seven organised grids, and once PJM’s ride-through deadline for existing sites passes, its data centres will stop dropping to backup power in blocks of 1,500 megawatts or more after normally cleared faults.”
Breaks if: “By 31 December 2027, fewer than two of ERCOT, PJM, MISO, SPP, NYISO, ISO New England and CAISO have a load of 50 MW or more, mainly a data centre, actually energised under an early-connection tier the grid may cut first, such as ERCOT’s Provisional Controllable Load Resource, SPP’s CHILLS, PJM’s Interim NITS or CAISO’s flexible interim service, shown in a filed service agreement or the operator’s own report; curtailment rights that apply to all large loads during emergency load shed, such as Texas Senate Bill 6, do not count. Or, between 16 January 2028 and 31 December 2028, PJM reports a transfer of 1,500 MW or more of large load to backup generation or off the grid after a normally cleared fault.”
Two earlier claims on the plug and the battery:
The grid’s new problem is not the power AI uses, it is the power it drops (22 September 2026)
Claim: “The grid’s new problem with AI data centres is not the power they draw. It is the gigawatts they can drop in under a second.”
Breaks if: “Ride-through turns out to be a cheap settings change rather than a constraint, or Virginia’s 3.8 gigawatt trip proves to be a local fault rather than how these sites are built.”
Where later evidence bears on it: the fault “cleared properly” and this was PJM’s third such transfer in about two years, which supports the second limb. The first is open: nobody has published what compliance costs. Where I was wrong: the piece said NERC’s alert “binds” planners and operators; NERC says it “does not create a mandatory obligation”. Binding standards will come from NERC, which FERC ordered on 16 July to file them by 31 December 2026. PJM times the event to the minute, so “under a second” was my reading, not PJM’s figure.
The battery has quietly stopped being an energy product (22 September 2026)
Claim: “Large batteries are increasingly bought as connection equipment that lets a site switch on early, not as a way to trade electricity.”
Breaks if: “New storage stops clustering where computing load is and spreads out with renewable generation instead, or connection waits shorten enough that opening early stops being worth paying for.”
Where later evidence bears on it: California’s 17.3 GW fleet is a time-shifter, earning about four times as much from shifting energy as from balancing. My count of the 2026 summer narrows the claim to, at most, the newest buyers beside data centres. The Texas figure I relied on counts grid batteries, which also trade, and no one publishes the share at data centre sites. The nearest primary case, storage behind Google’s Michigan site, is utility supply, not a connection condition.
The figures behind the charts
Every large American site now meets three questions: firm service or a tier the grid may cut, whether it must ride through faults, and who carries the wait.
| Grid | The cut-first tier | Must it ride through faults? | Who carries the wait |
|---|---|---|---|
| ERCOT (outside FERC) | Curtailable under local stress, or own generation with a capped draw (approved 18 June 2026) | Yes for new large computational loads; sites approved to energise, or with signed agreements, before 14 November 2025 exempt | New large loads can be cut in rotating outages (Senate Bill 6, 2025) |
| PJM | Interim service, non-firm until upgrades are built, no capacity bought (9 October 2026) | Proposed from 2027, existing sites included (6 October 2026) | Households, through the capacity auction |
| Southwest Power Pool | SPP’s CHILLS: as available, up to seven years (accepted 5 June 2026) | Not shown | Not shown |
| MISO, NYISO, ISO New England, CAISO | Non-firm or interim flexible services in design (August 2026) | MISO drafting | Not yet filed |
| Dominion Energy, Virginia | None; firm, with a 14 year minimum bill from January 2027 | At least six automatic reclosing attempts | The customer if it never fills; shared if it does |
“Not shown” means I found no primary statement. ERCOT’s and Dominion’s ride-through terms are as PJM summarised them on 8 September 2026. MISO’s and Southwest Power Pool’s plans are known only from a counsel’s summary (3 August 2026).
Texas. The 474 GW, five times the record peak and about 90% data centres are the Governor’s (3 August 2026); ERCOT’s own briefing gave 474 GW as of June. The 205 GW (preliminary), the 17 loads and 6,608 MW, the pause at 75 MW and the audit date are from ERCOT’s documents of 29 July and 20 August 2026. Senate Bill 6 makes loads of 75 MW or more connected after 2025 curtailable in rotating outages, and lets ERCOT order large loads onto their backup generators in an emergency. Under a seventieth is 6.6 GW against 474 GW, my arithmetic.
FERC. The PJM show cause order (EL26-67, 18 June 2026) suggests a large load is a new site of 50 MW or more above 69 kV. CAISO’s straw proposal is dated 11 August 2026.
PJM’s capacity auction.
| Delivery year | Held | Result |
|---|---|---|
| 2025/2026 | July 2024 | $269.92 per MW-day, up from $28.92 |
| 2026/2027 | July 2025 | $329.17, the cap |
| 2027/2028 | December 2025 | $333.44, the cap; about 6,500 MW short |
| 2028/2029 | July 2026 | $325, the cap; $16.4 billion; 6,831 MW short of the reliability requirement; 525 MW of new plant and upgrades |
PJM’s market monitor counts 21,470 MW of data centre load in the 2028/2029 forecast and attributes $6.3 billion, 38.2% of the latest auction, to it (27 July 2026). More than seven times is 3,800 MW against 525 MW, my arithmetic.
Virginia. Dominion’s 10-K (23 February 2026) gives the 28% and defines its high load customers as 25 MW or more at a load factor of 75% or more; its 8-K (14 September 2026) names the class GS-5 and sets minimum payments for generation, transmission and distribution. Virginia’s legislative auditor found shared generation and transmission are recovered over 20 to 40 years, far past a 14 year term, and estimated data centre growth could add $14 to $37 a month to a typical household’s generation and transmission costs by 2040, in 2024 dollars (JLARC Report 598, 9 December 2024).
The 22 July event. PJM’s deck gives about 3,800 MW, two stages, no relay misoperations and 60.092 Hz. Inside Lines calls it the largest load transfer in PJM’s history, after transfers of about 1,500 MW on 10 July 2024 and 17 February 2025.
PJM’s ride-through proposal (6 October 2026). It covers sites of 50 MW or more with at least 25 MW behind power electronics.
| Voltage, share of normal | Must stay connected for |
|---|---|
| Above 1.10, up to 1.20 | 1 second |
| 0.90 to 1.10 | Continuously |
| 0.80 to 0.90 | 2 seconds |
| 0.50 to 0.80 | 0.5 seconds |
| Below 0.50 | 0.25 seconds |
Frequency from 58.8 to 61.2 Hz must be ridden through continuously, and 57.0 to 58.8 Hz or 61.2 to 61.8 Hz for 299 seconds. Projects entering PJM’s studies from 16 January 2027 must comply; existing sites by 16 January 2028, or a year after switching on. PJM targets a FERC filing in November. FERC’s order to NERC (16 July 2026) asks for standards, with registry criteria, by 31 December 2026 and a workplan for more by 1 March 2027.
Slower chips. Google’s first deals were announced on 4 August 2025. Indiana’s regulator approved the first contract in April 2026 (Cause 46276), with credits and penalties; its size and delivery are confidential. The Phoenix trial is from Nature Energy (5 December 2025). DTE Energy says Google’s 1 GW Van Buren Township load ramp is supported by up to 1,600 MW of renewables, 480 MW of storage and 350 MW of demand response (8-K, 28 September 2026); it does not say how much of the 350 MW the computing itself delivers.
Batteries. EIA reported on 20 February 2026 that developers planned 24 GW of utility-scale batteries in 2026, 12.9 GW or 53% of it in Texas; EIA counts grid batteries, and nobody publishes how many sit at data centres. From CAISO’s five-minute fuel mix I counted the summer evenings, June to September, on which batteries were the largest source at the net peak, demand less wind and solar.
| Summer | Evenings batteries led, of 122 |
|---|---|
| 2023 | 0 |
| 2024 | 10 |
| 2025 | 52 |
| 2026 | 66 |
California’s fleet was 17,307 MW at 30 September 2026. CAISO’s market monitor finds batteries earned about $35 per kW-year from energy against $9 from regulation in 2025. In March 2025 PJM estimated a four-hour battery counted 94% against summer risk and 43% against winter risk, and named data centres as a main reason risk is moving to winter. Its binding ratings have run 50%, 58%, 59% and 57% for 2026/2027 to 2029/2030; its non-binding outlook falls to 42% for 2030/2031.
Who is exposed if America rations the plug
This maps who operates in each layer. It is not a recommendation, and naming a company is not a view on its shares. Figures in the piece are disclosures used as evidence for the mechanism.
- Rules, grids and wires
- FERC, NERC, ERCOT and PJM set the terms. Dominion Energy, Oncor and American Electric Power build the lines.
- Backup on site
- GE Vernova makes turbines, Caterpillar generators, Bloom Energy fuel cells and Tesla Energy batteries.
- Protection and slower chips
- Schneider Electric, Eaton and Vertiv make switchgear and backup power. Emerald AI makes software that slows AI work for Google and others.
- On the other side
- Developers counting on firm power from day one, and PJM households.
Sources: Office of the Texas Governor (3 August 2026); ERCOT Senate briefing (29 July 2026), regulator presentation (20 August 2026); PUCT (18 June 2026); Texas Senate Bill 6 (2025); FERC show cause orders (18 June 2026), abeyance orders (14 August 2026) and order to NERC (16 July 2026); PJM transmission owners (9 October 2026); NEPOOL counsel (3 August 2026); CAISO straw proposal (11 August 2026); PJM auction reports (30 July 2024 to 14 July 2026); Monitoring Analytics (27 July 2026); Dominion Energy 10-K (23 February 2026) and 8-K (14 September 2026); JLARC Report 598 (9 December 2024); PJM event review (31 July, 11 August 2026) and ride-through proposal (8 September, 6 October 2026); NERC Level 3 alert (4 May 2026); EIA (20 February 2026); CAISO fuel mix (my count), Key Statistics and market monitor (2026); PJM storage ratings (13 March 2025 to 19 August 2026); DTE Energy 8-K (28 September 2026); Google (4 August 2025, 19 March 2026); Indiana regulator, Cause 46276 (April 2026); Nature Energy (5 December 2025); Nvidia (16 September 2026). Leads: a YouTube review, "State of AI data centres 2026", and two videos by Leo Cui. Personal research, not investment advice.