Electricians have a name for the power a device draws while it is switched off. They call it phantom load: the television on standby, the charger with nothing attached, pulling current for work no one can see. The American grid is now being rebuilt around a second kind. It lives in utility queues, tax statutes and investor decks. It has a megawatt figure and a target year. Some of it will become concrete and copper. Most of it, on the only evidence anyone has bothered to collect, will not. And the people paying for it cannot tell which is which, because nobody who benefits from the confusion is required to.
This note follows a single watt through five places it might live and asks one question at each stop: is this real yet? The answers are shorter than the forecasts, and cheaper for everyone except the people who end up holding them.
Thirty gigawatts of maybe
what got connected
In March 2023 the utility that serves Columbus, Ohio stopped signing up new data centers. It was not out of power. It was out of certainty. Developers had been arriving with requests the size of small cities, and the interest on AEP Ohio's books passed 30 gigawatts: more than 50 customers, more than 90 sites, every one asking for firm service and none yet obliged to take it.12 The utility could not build for all of it. It also could not tell which requests were serious, because a request cost nothing to make. Every one of them looked the same on paper.
So Ohio's regulators changed the paper. From July 2025, any data center wanting more than 25 megawatts signs a contract of up to twelve years, pays for at least 85% of what it reserves whether or not it uses it, and posts financial assurance up front.2 Then the utility watched what happened to the line.
Fig 1The line outside the door
What data centers asked for, what they studied, and what they signed. Switch to Texas to see the same shape at five times the scale.
In Ohio, about 19 of every 100 gigawatts asked for became a signed contract once signing cost money.
Data table for Fig 1
| Item | GW | Encoding |
|---|---|---|
| AEP Ohio: interest in the queue when it froze new agreements (at least) | 30 | ghost |
| AEP Ohio: formal load studies, 36 sites | 13.0 | ghost |
| AEP Ohio: signed since the tariff, Feb 2026 | 5.6 | copper |
| AEP Ohio: all data center agreements incl. pre-tariff (17,861 MW) | 17.9 | copper |
| ERCOT: pending large-load requests, ~90% data centers, Aug 2026 | 474 | ghost |
| ERCOT: all-time peak demand, July 22, 2026 | 91.1 | copper |
| ERCOT: data center additions BNEF expects over the next four quarters | 1.2 | ghost |
| Hypothetical: 474 × 0.90 × 0.188 | 80.2 | breaker |
Five and a half gigawatts signed. Thirteen went as far as a formal study. The rest, four-fifths of what had frozen a utility's planning, were never customers. They were options, written free of charge against the public's grid.1 An analyst at Enverus described the queue before the tariff without ceremony: anyone with an acre of land they thought a data center might fit on was filing an interconnection request.6
Texas has the same line, five times longer. When Governor Abbott ordered an audit of every data center in the state's connection queue this August, the requests waiting there added up to about 474 gigawatts. The whole Texas grid peaked at 91 twelve days earlier, an all-time record.45 BloombergNEF expects about one gigawatt of data centers to actually connect there over the next year.3 Switch the chart above to Texas and the signed bar nearly disappears.
Hold that ratio. Everything downstream, the forecasts, the power plants, the capacity bills, is sized against the line. And the line is free to stand in.
What was promised
what was forecast
Forecasts are supposed to narrow as the future gets closer. These have climbed, and they climbed because their inputs are the queues from section 1.
The Electric Power Research Institute said in 2024 that data centers might use up to 9% of America's electricity by 2030. In 2026 it said 9% was the floor.78 Berkeley Lab, whose December 2024 report anchored the debate, pushed its ceiling to 15% in June and, in the same document, revised the past downward: the 2024 figure came in at the bottom of the range it had projected, mainly because fewer chips shipped than the data had said.910 History came in soft. The forecast rose anyway. S&P's 451 Research raised its 2030 number by more than half in under a year.1112 And two respected shops, Goldman Sachs and 451, still disagree by a factor of two about how much power data centers drew in 2025, a year that is over.1312 They count different things in different units. Neither is lying. That is the point. A number nobody can falsify is a number nobody has to defend.
Fig 2The ratchet
Each bar is a published forecast for data centers' share of U.S. electricity, placed on the date it came out. They arrive one at a time; press play to run them again, or show them all at once. The copper line is what was actually measured.
Four forecasts in two years. The newest floor is where the oldest ceiling was.
Data table for Fig 2
| Source | Published | Target | Share of U.S. electricity |
|---|---|---|---|
| EPRI | May 2024 | 2030 | 4.6% to 9.1% |
| LBNL | Dec 2024 | 2028 | 6.7% to 12.0% |
| EPRI | Feb 2026 | 2030 | 9% to 17% |
| LBNL | Jun 2026 | 2030 | 9.5% to 15.3% (reference 11.8%) |
| LBNL, measured | Jun 2026 | 2024 | 4.7% |
A forecast that lives in a report can be revised. A forecast that becomes a law is harder to reach, and the clearest case is Virginia. The state wrote a sales tax exemption for data center equipment in 2008 to lure an industry; it took effect in 2010 with a sunset of 2020. In 2016 the General Assembly moved the sunset to 2035, and the statute now lets operators who meet investment and hiring thresholds push their own exemption to 2040 or 2050.1415 The industry it was written for builds server halls costing billions, and the exemption has grown with them. In fiscal 2023 the legislature's auditors put its cost at $685 million. In fiscal 2024, $1.02 billion, before the local and regional shares abated alongside it. In fiscal 2025, $1.9 billion state and local.161718 The equipment inside those halls, as the next section shows, is written off in five or six years and outrun in one or two. The tax break that bought it has a longer life than anything it buys.
Fig 3The forecast that became a law
What Virginia's data center sales tax exemption cost, by fiscal year, above how long it has been on the books.
Nearly tripled in two fiscal years. About $267 million of the fiscal 2024 loss would otherwise have funded K-12 schools in the counties that host the data centers.
Data table for Fig 3
| Year | Cost | Basis |
|---|---|---|
| FY 2023 | $685 million | JLARC via Governing, state share |
| FY 2024 | $1.02 billion state; ~$1.39 billion with local and regional | Good Jobs First; Governing reports "$1 billion" |
| FY 2025 | $1.9 billion state and local | Good Jobs First, from state figures |
| Of which K-12 schools, FY 2024 | ~$267 million | Good Jobs First estimate |
| In force | July 2010 to June 2035; extendable to 2040 or 2050 by MOU | Code of Virginia § 58.1-609.3; SB 64 (2016); VEDP |
Four states, Georgia, Texas, Ohio and Virginia, now lose more than $1 billion a year each to data center sales tax exemptions, and each has raised its estimate since last fall.18 The legislators who approved these were handed forecasts in the same units and with the same confidence as the ones in Fig 2. Virginia did not build a mechanism to report what its exemption cost until fiscal 2021, thirteen years after writing it.16 The incentives run to 2035 and beyond. The forecast got a statute. The statute never got a meter.
What it is worth, and for how long
how long the asset lasts against the debt on it
The watt we are following was bought to feed a chip, and the chip is the strangest asset on the grid: the most expensive thing in the building and the first thing to go out of date.
Nvidia ships a new architecture roughly every one to two years, each several times more capable than the last: Hopper in 2022, Blackwell in 2024, Rubin in 2026.19 The companies buying them by the hundred thousand have, over the same period, lengthened the accounting life of their servers. Microsoft and Google to six years. Meta to five and a half, a change it told investors would trim about $2.9 billion from its 2025 depreciation expense.1920 Amazon went the other way, shortening a class of servers to five years and citing, in its own filing, "the increased pace of technology development, particularly in the area of artificial intelligence."21 Read those two moves together. Same market, same chips, same year: one buyer says the machines last longer, another says they last less, and each is right about its own income statement. The companies say retired training chips keep earning for years on lighter work, and that is where they go. The chips that stop being frontier become the inference fleet in section 4, kept warm at a fifth of full power, waiting for a prompt.19
between Nvidia architectures, as Nvidia ships them
over which the owners now write the servers off
when the debt on Meta's Hyperion campus matures
What is not in dispute is where the money comes from. Meta's Hyperion campus in Louisiana was financed with $27 billion of debt issued by a joint venture Meta owns one-fifth of, maturing in 2049, kept off Meta's balance sheet.22 CoreWeave borrows against the chips themselves and the customer contracts on them.22 The loan outlives the collateral by a decade or two, and the collateral outlives its own edge by a similar margin. Someone is betting the building will find tenants for chips that do not yet exist, and the bet is denominated in megawatts the grid has to honor whether or not the tenant arrives.
Meanwhile the equipment that feeds the building has gone parabolic. GE Vernova's book of gas turbine orders and slot reservations went from 83 to 116 gigawatts in six months; it is taking reservations for 2031 delivery, and its first-half 2026 orders are priced ten to twenty points above last winter's.232425 Wood Mackenzie expects turbine prices near $600 a kilowatt by 2027, close to triple 2019.26 A turbine is not a two-year purchase. It is being bought against a queue that is four-fifths bids, to feed chips whose makers outrun them every year or two, on the strength of a forecast that has already been wrong about the past.
Fig 4Written off slower than it is outrun
One AI server bought in 2026, on the schedules its owners state: what the books say it is worth, how many generations ship past it, and what it does while both clocks run.
By 2029, when the books say the server is half used up, between one and three newer generations have shipped past it, and it has spent four hours in five waiting, drawing a fifth of its power to do it.
Data table for Fig 4
| Item | Value | Source |
|---|---|---|
| Server useful life, accounting | 6 years (Microsoft, Google); 5.5 years (Meta); 5 years (Amazon, a subset) | CNBC; Meta Q4 2024 results; Amazon 10-K |
| Nvidia architecture cadence | roughly 1 to 2 years (Hopper 2022, Blackwell 2024, Rubin 2026) | CNBC |
| Inference server utilization assumed for 2030 | 20% | LBNL 2025 Update |
| Idle draw as a share of full power | 20%, measured | LBNL 2025 Update, citing Brookhaven |
Put every clock on one line and the mismatch is the picture. The party that sets each term is the party that benefits from its length.
Fig 5The clocks
How long each piece of the bargain lasts, drawn on the same timeline, as each party states it. Hover or tap a bar for the source.
Tap a bar.
Data table for Fig 5
| Item | Duration | Source |
|---|---|---|
| Nvidia architecture cadence | a new generation every 1 to 2 years | CNBC |
| AI server, accounting life | 5 to 6 years | Microsoft, Google, Meta, Amazon disclosures |
| AEP Ohio data center contract | up to 12 years | PUCO order, July 2025 |
| PJM capacity commitment | 1 delivery year, bought 3 years ahead | PJM auction rules |
| Meta Hyperion joint-venture debt | to 2049 | credit analysis of the SPV |
| Virginia data center sales tax exemption | 2010 to 2035; extendable to 2040 or 2050 | Code of Virginia; SB 64 (2016); VEDP |
The unevenness runs through the buildings too. The industry's average efficiency, the ratio of power into the building to power that reaches a chip, has been stuck at 1.54 for six years: 54 cents of overhead on every dollar of computing. Google runs its fleet at 1.09.2829 More than a quarter of operators work in buildings sixteen years old or older, where the upgrades that would close that gap no longer pencil out.28 The campuses with their own generation, their own cooling loops and their own engineers sit at one end of that spread. The older halls, and the communities around them, sit at the other.
The chips are outrun every year or two and written off over six. The loan runs to 2049. The tax break runs to 2035, or 2050 for those who ask. The bill outlives all of them.
What it does all day
what it does while nobody is asking
Here is the part that gave this note its title, and the reason the watt we are following is worth less than its price tag. The most careful forecast in the country, Berkeley Lab's, moves more on one assumption than on any other: not how many chips get built, but how much power they draw while they wait.
The servers that answer prompts are sized for the busiest minute of the day and spend most of the rest of it idle. The lab now assumes they will be working about one hour in five by 2030. It also assumes an idle AI server still pulls a fifth of its full power, a figure measured at Brookhaven National Laboratory. People who run these fleets told the lab it could be half, because the machines are kept warm so answers come back fast. Nobody has published field data either way.10 Nudge those two numbers and the national forecast for 2030 rises by a fifth, more than any other scenario the lab tested.10 The single largest uncertainty in American load growth is how much electricity will be spent on nothing.
hours an inference server is assumed to be working, 2030
of full power it draws while waiting, measured at Brookhaven
swing in the 2030 national forecast from those two numbers alone
That is phantom load in the literal sense: a reservation of power held at a fifth of full draw so a machine can answer inside a second. It is also depreciating while it waits. Every idle hour is one of the fifty thousand or so in a six-year book life, spent on a chip already a generation behind. The forecast counts the reservation as demand. The balance sheet counts the chip as an asset. The meter counts the standby current. Only the meter is measuring anything.
Fig 6One chip's day
Each block is fifteen minutes of one AI server's day. Copper is work. Ghost is waiting, still drawing power. Slide to change how busy it is, then flip to see what the waiting could be worth to the grid.
Working 4.8 hours of 24. Average draw 36% of full power, and waiting supplies 44% of that.
Data table for Fig 6
| Item | Value | Source |
|---|---|---|
| Inference server utilization assumed for 2030 | 20% (30% in the high case) | LBNL 2025 Update |
| Idle draw as a share of rated power | 20% measured; 50% industry hypothesis | LBNL 2025 Update, citing Brookhaven |
| Effect of 30%/30% on the 2030 national forecast | +20.6% (649 to 782 TWh) | LBNL 2025 Update, Table 2 |
| New load the existing grid could absorb at a 0.25% curtailment rate | 76 GW; about 85 hours a year with some curtailment; events average 1.7 hours; half the load keeps running about 90% of that time | Duke Nicholas Institute, as reported |
Average draw = full power × (working share + idle share × (1 − working share)). At 20% and 20%: 0.2 + 0.2 × 0.8 = 36% of full power.
That waiting is not only waste. It is also the most valuable thing a data center could offer a grid, if it were willing. Researchers at Duke estimated that the existing American system could absorb 76 gigawatts of new load, roughly everything the forecasts are arguing about, if that load agreed to be curtailed for a quarter of one percent of its maximum uptime: about 85 hours a year with some curtailment, in events averaging 1.7 hours, with half the load still running through most of them.3031 Flip Fig 6 to the year view and look for the copper. That is the whole ask.
The number has since been quoted in testimony and trade decks without its authors' warnings: it leaves out transmission constraints, which could erase much of the headroom, and it assumes data centers will actually deliver the flexibility they are credited with.3031 A machine that is idle four hours in five is not the same as a machine that will power down on request. The first is a fact. The second is a contract nobody has signed, and until someone signs it the flexibility is one more phantom: counted in the plan, absent from the meter.
Who pays
where the forecast becomes a line on a bill
PJM runs the grid for 67 million people across thirteen states and Washington, D.C.33 Once a year it holds an auction that pays power plants to promise they will show up on the hottest afternoon three summers from now. The price depends on how much demand PJM forecasts against how much supply appears, and every customer in the footprint pays a slice. That auction is where a queue turns into a bill.
PJM's independent market monitor has been doing the arithmetic. The forecast the July 2026 auction cleared against carried 21 gigawatts of data center load, three-quarters of it not yet on the system. The auction hit its price cap and still came up short of the reliability target.2734 The monitor's finding, across the last four auctions: data center load, actual and forecast, accounts for $29.4 billion of what the capacity market charged, 46% of the total, and the results are "almost entirely the result of large load additions from data centers."3534 Its proposal is to make data centers buy their own capacity in a separate auction. PJM has agreed about the cause, if not yet the cure.35
Fig 7The bill
Top: four years of PJM capacity charges, with the share its market monitor attributes to data center load. Bottom: what the same monthly increase means to households at different incomes.
An extra $10 a month is 0.2% of a $60,000 income. For a household already spending 20% of its income on energy, that is the same dollars on top of a burden nearly seven times heavier.
Data table for Fig 7
| Item | Value | Source |
|---|---|---|
| PJM capacity charges, four auctions 2025/26 to 2028/29 | $63.6 billion | IMM (29.4 ÷ 0.462) |
| Attributed by the IMM to data center load | $29.4 billion (46.2%) | IMM presentation, July 2026 |
| Data center load in the 2028/29 forecast; of which above embedded (not yet on the system) | 21,470 MW; 16,436 MW | IMM presentation, July 2026 |
| Crude per-person share, 67 million people, four delivery years | about $440, or $110 a year | copper.vision arithmetic |
| Energy burden, higher-income households | about 3% of income | EESI |
| Energy burden, low-income, renter, Black and Hispanic households | up to 20% of income | EESI |
| U.S. households behind on utility bills entering 2026 | about 1 in 6 | NEADA via Consumer Reports |
The dollars in Fig 7 are careful counterfactuals, not invoices, and hedged customers escape some of them. The direction is not in dispute. And a flat charge is not flat. Low-income, renter, Black and Hispanic households spend up to a fifth of their income on energy, against about 3% for higher-income households; one American household in six entered this year behind on its utility bill.3637 Nearly half of the country's data centers sit in census tracts that already carry above-median environmental burdens.38 Slide the income in Fig 7 down and watch the same ten dollars change meaning.
Files for several times what it will build, because a request is free and a place in line is worth money. Until Ohio put a price on the reservation, the cost of asking was zero.
pays: nothing, until it signs
Utilities and PJM feed the queue into the load forecast. If the forecast runs high, the plants get built and the cost passes through in rates. The error is never charged back to whoever made it.
pays: nothing; recovers it
Writes a 25-year exemption for an asset its owners write off in six and replace in two, then installs the meter for what it costs thirteen years later.
pays: $1.9 billion a year, foregone
Signed nothing. Carries the capacity charge for load that is three-quarters forecast, on a bill that already takes up to a fifth of income at the bottom of the ladder.
pays: the reservation, plus the error
There is a version of this that bills the reservation to whoever asked for it. Ohio's tariff does that at the utility: sign for up to twelve years, pay for at least 85% of what you reserve. PJM's monitor wants the same logic in the capacity market. Until one of them reaches the auction, the forecast is charged to everyone, and the party who guessed high never sees the invoice.
Three rules
what to do with a number you cannot verify
Follow the watt back up the page. It was asked for five times over, by parties who paid nothing to ask. It was forecast in units nobody shares, by parties who pay nothing when the forecast is wrong. It was subsidized by a tax break written before the industry it now serves existed, set to run a decade past the equipment it buys and, for operators who file the paperwork, a quarter century. It spends four hours in five waiting inside a chip that will be outrun before the loan on the building is a tenth of the way to maturity. And it shows up on bills that fall hardest on the people least able to pay, none of whom were in the room.
None of the people producing these numbers is lying. The forecasts, the queues, the exemption ledgers and the depreciation schedules each answer a different question, and they all wear the same units. The damage begins when a bid is read as a plan, a projection as a price, or a standby machine as demand. Every mismatch on this page runs the same direction: the cost of being wrong lands on someone other than the person who guessed. Three rules would stop most of it.
- Count what is signed, not what is requested. Ohio did this and four-fifths of its queue vanished. Texas is doing it now. Every forecast built on the old queues is still in circulation, and still being priced into auctions.
- Charge the asker for the reservation, not the neighbors. Ohio does this at the utility. PJM's own monitor is asking for it in the capacity market. A reservation with a price attached is the only kind that tells you anything.
- Publish the meter, not the model. The fleets hold the idle-power data that would settle the forecast, and the states hold the exemption ledgers that would have caught Virginia a decade earlier. Virginia only built a mechanism to report its exemption's cost in fiscal 2021, thirteen years after writing it.16 An asset that is depreciating in the dark is not an asset anyone should be lending against, or building power plants for.
Twenty-four of Ohio's thirty gigawatts were never coming. Everything built to serve them was real. So is the bill.