Companies that plan to build data centers ask the local electric utility to reserve far more of the grid than they will ever use, because asking costs little next to what a place in line is worth. Exelon, which owns Commonwealth Edison in Chicago and northern Illinois, says it has screened the speculative requests out: its count of likely data center load fell from about 18 gigawatts in February to about 11 in July, and, by its account, 4 of those 11 have signed payment guarantees backed by $1 billion of collateral. We read one agreement as filed and the four cancellation notices at the federal regulator: every agreement carries a fee for leaving early, promised to the other customers, and in the two exits that are complete no fee appears. Whether that is a screen that works or a free option with a signature on it is a question Britain's House of Commons argued in 1845, when it doubled the cash deposit it demanded from a railway's promoters. The instrument we build to answer it, reading a commitment by how it ends, is built for a larger job: the order books, slot reservations and remaining performance obligations that hold up the AI trade, more than $1.3 trillion of them at Oracle and Microsoft alone, are backlogs of the same kind, disclosed far less.
Where this began. My AI assistant and I read Jennifer Hiller's report in The Wall Street Journal of October 13, 2026, "'Phantom' Data Centers Are Demanding More From the Power Grid Than They Need," and that morning sent Vista's research desk the question it raised for us: whether Exelon's screen of speculative requests is real, or a free option with a signature on it. The desk sharpened the question, and this inquiry is the answer; the full story is near the end, under The idea.
A utility that makes a would-be customer pay to hold its place in line ends up with customers who mean it; a utility that takes every request at face value builds for some who were never coming. Two things have to be true for the screen to work: the customers must be made to commit, and the money they pay must stay with the customers who would otherwise pay for the wires and substations the utility builds for them, the equipment that utilities call plant. The second condition matters because a regulated utility collects, each year, the cost of its wires plus a return for its owners, divided among every customer in its zone, the part of the regional grid that Commonwealth Edison, ComEd, serves. If the payments that hold a place in line are credited against everyone else's bills, the screen protects them; if they become the utility's profit, the screen is a toll the neighbors pay.
Exelon says it has done the first. ComEd now requires any customer planning 50 megawatts or more, a megawatt being a million watts and a gigawatt a thousand of them, to sign a transmission security agreement, a TSA, a promise to pay for ten years of service on the high-voltage grid at the size the customer itself wrote down, whether or not it ever draws the power; 4 gigawatts of the 11 have signed, Exelon says, backed by $1 billion of collateral, security a customer posts that the utility can draw for what it is owed. The second condition now has an answer, read in one agreement as filed and in the cancellation files at the Federal Energy Regulatory Commission, FERC, which oversees the high-voltage grid. Every ComEd agreement carries a fee for leaving early, "tens of millions of dollars for the first 180 days" in the one case ComEd described, rising after that "to hundreds of millions of dollars," and ComEd told FERC that such fees would be credited to its transmission cost of service "for the benefit of other customers," a promise FERC noted. Yet in the four cancellation notices on file, two of which ended agreements, no fee appears: one developer gave notice, in ComEd's words, just before its agreement took effect, so by our reading no fee fell due; one left by mutual consent, where the fee is owed unless ComEd waives it in writing, and nothing says which happened; FERC rejected the other two notices, so those agreements stand, and one is before a court. The protection is real on paper and untested in practice.
So what do the four notices tell us? They settle less than the press coverage suggests. Read the right way, though, they say something the coverage missed. Each named a withdrawn or cancelled application for service, which fits an option, a place in line it need never use, and a slipped project alike. We borrowed a method from climate science, which faces the same problem, and asked a different question, the exit fingerprint: not what reason the notice gives, but when the exit came, through which door, and at what cost. Both completed exits came while leaving cost the least. A cheap option would leave then, but so would most slipped projects, real data centers whose plans were delayed or fell through: their trouble can come at any time, and since the cheap window has covered most of these agreements' lives, chance alone puts most of their exits inside it. So for these agreements timing can never settle the question. What the next year can show is whether exits keep coming once the fee has stepped up, and through which door. A door is a way out the agreement allows: some cost nothing, such as leaving before the agreement takes effect or by mutual consent with the fee waived; some owe the fee, leaving for convenience among them; and some are kept for the utility's own delay, a regulator's refusal or an outside event. The test that reads them was frozen on October 20, before any exit it will score.
In Britain in the 1840s no railway could be built until Parliament passed a law for it, and a promoter who wanted one had to lodge his maps and his list of the land the line would cross with the clerk of the peace of each county "on or before the thirtieth day of November," as a notice in the London Gazette of November 1845 reminded landowners. That month The Times counted 1,263 new railway projects announced in 1845 alone. Andrew Odlyzko, a mathematician, describes the climax in his 2010 history of the mania: several hundred cabs converging on a government office at midnight, express trains carrying documents colliding, "bundles of papers conveyed secretly in a hearse." And before Parliament would look at any of those plans, it required the promoter to lodge cash, a set share of the money his investors had pledged.
A hundred and eighty years later the promoters are back, and the thing they are promoting is a building full of computers.
The artificial intelligence boom has made electricity the thing a data center cannot get, and a developer who wants to build one must ask the local utility to connect it. Asking is cheap. A developer who has not yet found a tenant, or financing, can file for power at five sites and build at one. Each request is a free option on a scarce thing: the right, but not the duty, to take hundreds of megawatts if the project happens, and to walk away if it does not. Hiller reports a research firm's estimate: Wood Mackenzie counts only 23 percent of the data center requests disclosed to utilities as "high confidence." The Journal's word for the requests that will never be built is phantoms; we use it for a type of request, never for a named developer, because a real project fails for many reasons. A utility that builds for phantoms builds plant that nobody uses, and the cost goes into its revenue requirement, the yearly sum its regulator lets it collect from every customer in the zone. Jeanne Jones, Exelon's chief financial officer until October 5, put it to the Journal plainly: if the grid is built and the data center walks away, "someone's got to pay for it."
I know the type from the other side of the table. In 2008, in the recession, my company hired a consultant to call our newer competitors and ask whether they would sell. Under non-disclosure agreements they showed us their books, and the refrain, as I set it down in The Scholars' Treasure, was that "next year we'll surely be at $5 million in sales," and the price should be set by that (ch. 12, p. 150). A request built on hoped-for revenue is not a lie. It is an option, and the other side of the table has to decide what it is worth.
ComEd announced its first transmission security agreements on January 6, 2026, with eight large customers representing "a forecasted new load greater than 6.5 GW." As of February 12, Exelon counted, in the words of its May deck, a "committed data center pipeline" of about 18 gigawatts, projects "in an official phase of engineering with deposits paid," about 45 percent of them "secured with TSAs." On July 30 the picture was redrawn: about 11 gigawatts of "high probability load," about 40 percent of it covered by TSAs "backed by ~$1B in collateral," and a separate future pipeline of about 25 gigawatts.
From 18 to 11 is a fall of about 39 percent, and it would be easy to read it as demand walking away; but the two figures carry different definitions, and the decks do not say what moved between them. Ms. Jones told the analysts, in a third-party transcript, that the company had "weeded out speculative projects," that 4 gigawatts of the 11 have signed TSAs, and that the other 7 "predated the TSA process but are further along." Exelon's capital plan, $41.7 billion over 2026 to 2029, was unchanged, she said, "because we have not put in speculative projects"; at least $1 billion of the transmission Exelon ties to committed data center load sits outside the plan, "largely expected in 2030 and beyond," though the deck says new business, the row that holds it, may need spending before 2030. One smaller thing: the data center slide and Ms. Jones call the $1 billion collateral; the next slide, and Exelon's release of August 4, say "Customer Protections," which may be collateral or may be a promise.
Illinois made the agreements compulsory this spring. The tariff, a utility's published rules and prices, that the Illinois Commerce Commission approved for ComEd in March 2026 requires a TSA of any customer whose projected load reaches 50 megawatts within ten years, and requires it "to guarantee that its annual payments for transmission service" will not, for ten years, fall below what the customer's own load ramp, its schedule of how fast its demand will grow, would pay. If a customer pays less than it promised in any year, the shortfall is collected from it and "credited to the transmission revenue requirement borne by transmission customers in the ComEd zone." Read that clause twice: it says where a shortfall payment goes.
The first agreements were signed before that tariff took effect, and they are the ones now coming apart, so we read one as filed, ComEd's agreement with Karis Critical, which ComEd told FERC is a standardized form, and whose mutual-consent clause FERC's order on a second agreement, Monarch's, quotes word for word. Three things in it matter. The first is the minimum, owed once the data center is running. The agreement commits the customer to "80% of the sum" that a customer "with 80% of the monthly loads" on its schedule would pay, which read literally, 80 percent of 80 percent, is 64 percent of the scheduled charges and, if one of the two is redundant, 80; the tariff that followed asks for the whole ramp. The second is the fee. Section 14 and a schedule, Exhibit E, whose amounts are blacked out in every public copy, set an early termination fee by the year in which the customer leaves. The third is the doors. We counted eleven ways out. Five owe the fee, leaving for convenience, no reason given, among them. Five owe nothing: leaving before the agreement takes effect, by our reading, since the fee runs from that day; a utility delay beyond a length the public copy blacks out; a refusal by the Illinois commission, the grid operator or FERC; force majeure, an outside event, when ComEd is the party that ends the agreement; and a change in law. The eleventh, mutual consent, owes the fee "unless" ComEd waives it "in writing." Every exit under the agreement's delay and termination sections requires the customer to withdraw its application irrevocably; a project that returns must ask again. And what the customer forfeits is what it owes on the day it leaves, not its collateral: ComEd may draw the letter of credit, a bank's promise to pay a set sum if its customer does not, "for any and all unpaid costs or amounts owed to it," and must release the rest. Before service that is the fee for the date, and the fee is smallest at the start, which is where the screen must bite.
FERC Commissioner Judy Chang's concurrence in a sister case, a statement that agrees with a ruling and keeps its reservations, reads today like a list of things to check: the commission "has not independently assessed the terms of the Agreement," the cost of whatever upgrades are needed would be "rolled into transmission rates that all customers pay," and the upgrades are paid for over forty years while the agreements, in her example, run eight to ten. Her statement also carries ComEd's own figures, 6,700 megawatts of new load at $950 million of upgrades, which ComEd argued would lower other customers' rates, estimates she called "preliminary and subject to substantial changes."
Then the agreements began to come apart. Ending a filed agreement takes a notice that FERC accepts or rejects, and a rejected notice leaves the agreement in force. Karis Critical, whose agreement FERC had accepted on February 17, withdrew its application for service "immediately prior to the effectiveness" of the agreement, in ComEd's words, and ComEd filed on March 17 to cancel; FERC's staff accepted on May 8. The agreement took effect on February 21; the fee runs from that date and the notice came before it, so by our reading none fell due. In the summer ComEd filed to cancel two more, with Red Energy Partners and with Monarch Rock Air, on July 21, and a fourth, with PowerHouse Hillwood, on July 24. FERC accepted Monarch's on September 18 under the agreement's mutual-consent clause, which makes the fee payable "unless" ComEd waives it "in writing"; neither ComEd nor FERC says whether it was paid or waived. FERC rejected Red Energy's the same day because ComEd named no provision of the agreement that allowed it; no refiling had appeared by October 20, so that agreement stands.
The Hillwood case is where the protection's terms are being tested. Hillwood's project, Utility Dive reports, is a 1.8 gigawatt data center in Joliet, Illinois. ComEd's notice gave two grounds: a retail application that ComEd itself had cancelled, for noncompliance with its retail tariff, and a breach of the credit support terms. FERC rejected the notice on September 22, so that agreement stands too. Hillwood posted a letter of credit on April 14, 2026, in the amount of $1. Hillwood's position, as the order records it, is that the agreement names no amount for the period before its first year of service, so a $1 letter met a duty to post something within 30 days; ComEd's is that "a $1.00 letter of credit is indistinguishable from no letter at all." The notice came four days after Hillwood sued ComEd, in a case now in federal court in Chicago, and FERC left the letter of credit terms to that court. ComEd's answer of September 15 put the fee's size on the public record. FERC Commissioner David Rosner: "A $1 letter of credit protects no one."
What ComEd says a developer pays when its project dies, in the one case where it described it, is not the fee but its deposit under ComEd's retail rules, which ComEd's declaration in the Hillwood case says "would be drawn upon" for its actual expenses, a circuit breaker vendor's cancellation fee and the engineering studies, with the rest refunded. That money repays costs; it is not a forfeit that lowers anyone's bill. None of it proves that any of the four developers was a phantom: Monarch's chief operating officer, in a local newspaper, the Rock River Current, describes a revised timeline rather than an abandoned project. To get past that sentence we needed a different instrument.
In 2021 the Nobel Prize in physics went in part to Klaus Hasselmann, a German oceanographer who had solved a problem that looks nothing like ours and is exactly ours. The climate swings on its own, and Hasselmann showed in 1976 why: a slow system such as the ocean, fed by the fast random shocks of the weather, wanders over years and decades, so a trend can build up from nothing but accumulated noise. A warming trend is therefore not evidence of a human cause until the swings the system makes by itself have been measured. His answer, set out in 1979 and 1993 and called optimal fingerprinting, now underlies the attribution statements of the Intergovernmental Panel on Climate Change: write down in advance the pattern each cause would leave; let the data count most where the patterns differ and the noise is quiet; and say how much more data it would take for a signal to stand out. Vista's Library of decision models holds it, and we bring it to ComEd's cancellations, where nobody, so far as we can find, has thought to use it.
The method is not really about climate. It fits any question of this shape: two causes would leave different traces; which one left these? A sales slump reads one way if a large customer has quietly gone and another if the season is to blame; the trace is in which accounts shrank and when. An order book whose cancellations bunch just before a deposit turns non-refundable looks like a book of options; one whose cancellations scatter through the year looks like projects that ran into trouble. A stock that falls on news about the company leaves a different mark across its peers than one that falls with its whole industry. The trillion-dollar backlogs behind the AI trade are commitments of the same kind as ComEd's, and the same reading could tell their options from their demand once the companies say how their commitments end. But first, let's get it dialed in on ComEd's four notices; then we'll point it at AI.
As climate science writes it, the observations y, one number for each place measured, are a signal pattern g scaled by an amplitude a, how much of the cause is present, plus the system's own variability u, whose covariance C records how large the swings are at each place and how they move together:
y = a x g + u
The best linear test divides the pattern by the noise, so that a place where the pattern is strong and the swings are small counts for much and a noisy place for little. That weighted pattern is the optimal fingerprint,
f = C^-1 x g
and projecting the observations onto it gives the amplitude and its uncertainty, the dot being the sum of products across places and g' the pattern as a row:
a_hat = (f . y) / (f . g)
var(a_hat) = 1 / (g' x C^-1 x g)
The signal-to-noise ratio is the largest any linear filter reaches, which is what optimal means, and with m independent periods it grows as sqrt(m):
SNR = a x sqrt(g' x C^-1 x g)
Detection is an amplitude significantly greater than zero. Attribution asks that it be consistent with one when g carries the size the cause itself predicts, with rival causes estimated beside it, Hasselmann's multi-pattern method of 1997,
y = a_1 x g_1 + a_2 x g_2 + ... + a_p x g_p + u
and what the patterns leave unexplained should be no larger than the noise allows, the check Myles Allen and Simon Tett added in 1999.
Carried to our problem, the method gets a plain name: reading a commitment by how it ends. An option and a real project look exactly alike on the day they are signed; the difference shows only in the leaving, and the mark the leaving makes is the exit fingerprint. So y is no longer temperatures: for each exit it is when the exit came relative to the date the fee steps up, through which door, and whether a fee appears afterward. The two causes are types of request, never named developers, and each pattern is the probability of those features under its cause, drawn from the agreement's own terms. The first, g_O, is a request held as an option, built only if a tenant turns up: it leaves when leaving costs the least, before its agreement takes effect or inside the first 180 days, through a door that costs nothing, and it rarely pays. The second, g_R, is a real project that slipped: it leaves whenever its trouble comes, so it lands in the cheap window as often as chance puts it there, or more if it sees its trouble coming, uses the door kept for the utility's delay when the delay is the utility's, and pays when it leaves for its own reasons. The noise C is how those features vary among exits of a known kind; climate science estimates it from model runs with the cause switched off, and ComEd offers no comparison group yet, so what follows takes the slipped project's own chance as the noise and states each result as a bound. One confession: we knew the four dates when we wrote these patterns down, and a pattern counts as evidence only when it was fixed before its data; one more reason the test that counts is the one frozen for exits still to come.
With two patterns the option's share of the exits is a ratio of amplitudes,
s = a_O / (a_O + a_R)
and it is the haircut the method proposes: B signatures read as B x (1 - s) of demand. Four notices are far too few for the regression; what they allow is the weights, one feature at a time.
Now the record, and first the part that weighs nothing. With noise uncorrelated across features, the fingerprint's weight on feature k is the contrast between the patterns over its noise variance:
f_k = (g_O,k - g_R,k) / C_kk
Every notice says the application for retail service was withdrawn or cancelled, and every one says the parties will have no remaining obligations. But the agreement demands a withdrawn application of anyone who leaves, and "no remaining obligations" is as true of a fee waived as of a fee paid: for both, g_O,k equals g_R,k and the weight is zero whatever the noise. Carl Hempel's paradox of 1945, told in the later literature with a white shoe, asks what a white shoe says about the claim that all ravens are black; the Library's entry on which evidence counts answers with the likelihood ratio, the form the fingerprint takes for a yes-or-no feature:
LR_k = P(feature k | option) / P(feature k | slipped project)
Evidence both explanations predict equally has a ratio of exactly 1 and carries no weight however many times you see it. Four of four notices carry both features, each at 1; the stated record cannot tell the causes apart, and neither can Exelon's drop from 18 to 11 gigawatts, which a screen and a delay would both produce.
What the record does show is timing and door. Counted from the day each agreement took effect, the four notices came at 24, 153, 153 and 131 days, all inside the first 180. In that window the fee is at most about $55,556 a megawatt on the one project ComEd described, a little over a fifth of the $250,000 a megawatt that Exelon's $1 billion of collateral averages over 4 gigawatts. None used the door kept for the utility's own delay, but that door opens only when ComEd forecasts a long delay, which nothing shows, so this weighs nothing either. Two exits are complete, and neither shows a fee; the other two notices were rejected, and Hillwood's date was ComEd's choice.
The equations say how little that proves. Take the option at its strongest, always leaving inside the window, and a slipped project's exit as timed by events unrelated to the fee, as likely on one day as another. An agreement read A days after FERC accepted it has shown the window for 180 of those days, so a slipped exit lands inside it with probability 180 / A, and an exit there favors the option by at most
LR(A) = A / 180, for A of 180 days or more
FERC accepted the first cluster on February 17, 2026, 245 days before October 20, when this was written: 180 / 245 is 73.47 percent and the ratio at most 1.3611. Independent exits multiply, so the two completed exits together favor the option by at most
LR_total = (245 / 180)^2 = 1.8526
while two slipped projects would both have left that early by chance alone with probability (180 / 245)^2, 53.98 percent. Taking even odds to 95 percent needs a ratio of 20, and at these ages that takes
n = ln(20) / ln(245 / 180), which rounds up to 10 exits inside the window
Nor can these agreements supply them: their cheap windows closed in August and September 2026, so any exit they file from now on comes after the fee steps up, where the clock says nothing and only the door and the fee speak.
How much data, then, would let a signal stand out, Hasselmann's third move? With option exits drawn into the cheap window with probability 0.8 and otherwise landing there, like slipped exits, with probability p_0, the window's share of their timing, a record in which a share s of the exits are options shows a window share
w = p_0 + 0.8 x s x (1 - p_0)
and detecting an option share of one in five, with 80 percent power at a one-sided 5 percent level, takes 21 exits when p_0 is 5 percent, 87 when it is 25 percent, and 645 at ComEd's 73.5 percent, where the window has covered most of the agreements' lives; estimating the share to a standard error of 5 points takes 112, 259 and 1,541. So the sixth frozen measure counts the door each ComEd exit of the coming year uses, doors, never developers,
free share = free exits / exits, decisive only when exits are 3 or more
supported at one in three or fewer and falsified at two in three or more; with three exits it lands in the wrong band 26 percent of the time when the true share sits at the other band's edge, 10 percent when the true share is 20 or 80 percent, and 2 percent at that edge with eight.
The same discipline tells us what to make of the industry's numbers. Hiller reports that developers ask several utilities at once, so the same project is double, triple or quadruple counted, and duplicates alone could be 50 to 75 points of the 77 percent that Wood Mackenzie does not call high confidence, so that figure counts requests, not phantoms. Texas is where counting cannot be the whole story: Jim Burke, the CEO of Vistra, a power producer, calls its queue "overstated by more than 20x" and duplicates at four utilities explain at most a factor of 4 of that.
Vista's Library has an entry for the utility's own situation, which economists call screening. A seller who cannot tell its customers apart, Michael Mussa and Sherwin Rosen showed in 1978, offers a menu of versions and prices built so that each type of customer picks the version meant for it: the airline's cheap ticket with strict rules beside its expensive flexible one. Economists have applied it to the regulation of utilities since David Baron and Roger Myerson in 1982; what is rarely done is to read a utility's own terms for its largest customers as the menu, a deposit, a minimum payment, a term, collateral and an exit fee, on which a serious developer signs and a casual one does not. A second entry says when a commitment holds: Thomas Schelling's observation of 1960 that a promise is believable when breaking it costs more than keeping it.
To see what sorts, we built a small illustration, every number an assumption. Write q for a speculative request's chance of finding a tenant and being built, f for a real project's chance of failing for other reasons, S for what a project that gets its capacity on time is worth to its owner per megawatt, K for the security a requester posts and c for the cost of carrying it while it is held, and W for what it forfeits if it walks away. Each type's expected cost of signing is its chance of walking times the forfeit, plus the carry:
cost_O = (1 - q) x W + c x K
cost_R = f x W + c x K
and the menu sorts when it costs the speculator more than its option is worth and the real project less than what it expects to gain:
sorts when cost_O > q x S and cost_R < (1 - f) x S
Suppose q is one in five, f one in ten, S $1 million a megawatt and c 4 percent, two years at 2 percent. Then the speculator's request is an option worth about $200,000 a megawatt, the real project expects about $900,000, and with forfeitable security the menu sorts at forfeits between about $238,000 and $6.4 million a megawatt. A deposit that comes back screens almost no one, since W is zero and only the carry remains, and ComEd's refundable study deposit, $5,000 a megawatt, would cost the speculator a few thousand dollars a megawatt even if forfeited outright. Ten yearly payments at ComEd's 2026 price of $46,876 a megawatt, counting a dollar a year from now as worth less than a dollar today, at 10 percent a year, are worth about $288,000 now, and a ten-year guarantee for the full amount, backed by collateral of that size, would turn the speculator away. At the Karis agreement's minimum, 64 or 80 percent, the same guarantee is worth about $184,000 or $230,000 and falls short. And the fee for leaving: at its top in the first 180 days, about $55,556 a megawatt, it costs our speculator about $54,000, of which $10,000 is the cost of carrying the collateral, and sorts no one; past day 180 it sorts only above about $430 million for a 1.8 gigawatt agreement. The same guarantee backed by $1 of credit support sorts nobody, for a developer with nothing else to lose; it is a free option with a signature on it.
Two published menus make the comparison. The Virginia utility Dominion Energy Virginia's new terms for large customers, from January 1, 2027, run 14 years with an 85 percent minimum for transmission, an exit fee for the rest of the term, and collateral of $1.5 million a megawatt, reducible by up to 70 percent for established credit: at our numbers it sorts outright, at $1.5 million or at $450,000. AEP Ohio, the Ohio utility of American Electric Power, asks a minimum bill of up to 85 percent for the ramp plus eight years, collateral of half the term's minimum charges, repayment of its buildout costs if the customer cancels before service, and an exit fee held for its customers: the collateral alone costs our speculator about $134,000, a third short of its option, so the menu sorts only with the repayment on top. Schelling's test takes one line: a single-purpose company's commitment binds only when its posted security K covers the present value of the minimum charges it still owes, PV(M), less what it could recover by reassigning the capacity:
binds when K >= PV(M) - salvage
And the $1 billion of collateral is an average that says nothing about any one agreement. Proverbs puts the utility's position in one line: "He who is collateral for a stranger will suffer for it" (Proverbs 11:15, WEB). A utility that accepts a pledge it cannot collect has become the stranger's guarantor, and behind it stand its other customers.
The oldest version of the menu we know of was argued in the House of Commons and survives in Hansard, compiled in those years from the newspaper reports. In 1837 the Commons adopted a Standing Order requiring the promoters of a railway bill to deposit with a court one tenth of the money their investors had pledged before Parliament would consider the bill. When the rule was attacked in 1841, one member said that before it "men of straw," men with nothing behind them, "came before the public with ephemeral and good for nothing speculations."
Then the screen was loosened: on May 20, 1844 the Commons cut the deposit to one twentieth, 5 percent, Gladstone, then President of the Board of Trade, arguing that the order existed only to give the House a guarantee "that the scheme was a bona fide one," a genuine one. Applications for new lines ran 63 in 1843, 199 in 1844 and 562 in 1845, and investors held scrip, certificates for shares on which only a small deposit had been paid; a deposit of £1 on a £20 share, Odlyzko writes, "was really an option," because until Parliament approved the scheme the shareholder owed nothing more.
Parliament saw it and put the deposit back to one tenth on July 29, 1845, in the middle of the boom. The following January, when bankers complained that the cash deposits were draining the money market, Sir Robert Peel, the Prime Minister, told the House what a deposit is for: the deposit had been five percent, "but that being thought too small," the House, "with the view of discouraging unsound speculations," had raised it to ten.
Then the filter ran. Peel told the House on April 23, 1846 that 519 railway bills were pending, asking leave to raise about £304 million. Of the 516 bills considered that session, Gareth Campbell, the economic historian, records, 81 failed the House's entry rules, the deposit rule among them, 83 were withdrawn, 106 were rejected after a hearing, and 246 received the Royal Assent: the rules at the door took out more than one in seven before anyone looked at their merits. Set that funnel beside Exelon's: about 18 gigawatts in February, 11 in July, 4 of them signed and, by the company's account, collateralized. The two columns measure different things, so the chart is a picture, not a proof. But the shape is the shape.
What the filter could not do is the lesson. The part of the deposit lodged in court came back: T. S. Duncombe told the House in April 1846 that it "might be only borrowed," lent by a banker and returned whether the bill passed or failed. The investors' own deposits on their shares paid the promoters' costs, and in the failed schemes, Odlyzko writes, investors "typically lost all or a large part of their deposits." The number of schemes rose through the cut of 1844 and the rise of 1845 alike, and we found no study that measures how much the rule itself changed it. What it shows is a legislature reaching for the instrument a utility reaches for today.
For each dollar of new wires a regulated utility is allowed to collect, each year, the owners' allowed return with the income tax on it, the lenders' interest, depreciation and upkeep. With e the equity share of the capital, ROE the allowed return on it, t the composite tax rate, i the cost of debt, and d and m depreciation and upkeep as shares of the plant's cost, a dollar of plant costs each year
alpha = e x ROE / (1 - t) + (1 - e) x i + d + m
Using ComEd's own 2026 formula rate, posted by PJM, the regional grid operator, an allowed return of 11.5 percent on equity that is about 54 percent of the capital, with depreciation and upkeep assumed at 2.5 and 1 percent, that is about 14 cents. ComEd charges a single zonal price per megawatt of peak demand, r_0, $46,876 a year in August 2026: its yearly revenue requirement RR_0 over its customers' total peak demand L_0. New plant costing C for new load L billed for a share b of what it reserved moves that price to
r_1 = (RR_0 + alpha x C) / (L_0 + b x L)
so the new load lowers it whenever its plant costs less than
C* = b x L x r_0 / alpha
So there is a ceiling: the amount of plant the new customers' payments can carry. With ComEd's preliminary 6,700 megawatts and Chang's hypothetical minimum of three quarters of contracted size, about 5,000 megawatts at $46,876 each is about $235 million a year, which at 14 cents of yearly cost per dollar can carry about $1.7 billion of plant. ComEd's preliminary $950 million sits under that ceiling, and the rate equation gives a fall of about 8 percent in the price everyone else pays; at $1.4 billion of upgrades the price still falls about 3 percent, and at $1.9 billion it rises by close to 3 percent.
Turned around, the billed share that leaves the other customers no worse off is
b* = alpha x C / (L x r_0)
about 43 percent of reserved capacity for $950 million of plant, a number no public document yet gives.
The customer in Chang's own example would raise the zonal price by about 0.7 percent at our yearly cost, which is why our test's falsify line sits at 1 percent. A ten-year promise is shorter than the wires' life, about forty years in Chang's example and our assumption, so after year ten about three quarters of the plant's cost is left with whoever pays the zonal price then.
The shareholder's side is simpler, and smaller than the headlines suggest. The owners earn the allowed return on the part of the money they supplied, so plant costing K earns them
E = e x ROE x K
54 percent of each $1 billion at 11.5 percent, about $62 million a year. Exelon's plan funds about 40 percent of its new spending with new shares: with N shares outstanding earning EPS each, and N_new new shares, 0.4 x K / P of them, sold at a price P, the yearly lift per share is
lift = (E - EPS x N_new) / (N + N_new)
Net of the dilution, each $1 billion of plant adds about three and a half cents a share a year at the prices at which Exelon's own share sales settled in 2026, a little over 1 percent of the $2.81 to $2.91 a share it expects this year; two other ways of counting the owners' money, charging the retained equity at the allowed return or borrowing the balance at the parent, give about two and about three cents, and all three are hypothetical.
None of these is a recommendation, and only the Exelon hypothesis is scored.
Exelon, the screen as sounder growth. The hypothesis is that Exelon's regulated growth from data centers is real, late and neutral for its other customers: it will enter the plan at the plan's own 40 percent of new shares, the zonal price will not rise because of it, and the agreements will bind. Falsified if ComEd's 2027 update shows the new large loads raising the zonal price by 1 percent or more, if the year-end plan shows none of the $1 billion-plus entering it and the beyond-plan figure falling below $1 billion, or if two in three of the agreements that end in the next year leave for free.
American Electric Power, the stricter menu on paper. The hypothesis is that a menu as firm as AEP Ohio's turns away more speculative load than ComEd's; falsified if AEP's disclosures for 2026 show Ohio large-load contracts cancelled or renegotiated at a higher rate than ComEd's, or if the tariff is loosened by the appeal pending at the Supreme Court of Ohio.
Dominion Energy, the heaviest collateral. The hypothesis is that $1.5 million a megawatt over 14 years makes Dominion's commitments self-enforcing, so that cancellations after 2027 come with forfeitures the company discloses; falsified if Dominion's 2027 filings show large-load contracts cancelled with no forfeiture, or with forfeitures kept by the company rather than credited to customers.
What follows was frozen on October 18, 2026 for its first five measures and on October 20 for the sixth. The test lives in a computer program with a digital fingerprint, a code that changes if any line is altered: 3cdca99a971a77159602bcf77ba4e69849924ad6efdb8aeb4b16aacc0a0152ce, and the program refuses to run if any pass or fail band changes. It is scored from ComEd's 2027 formula rate filing, PJM's 2027 peak demand list, Exelon's annual report for 2026 with the capital plan of February 2027, and FERC's electronic library; the first five measures by June 30, 2027 and the sixth by December 31, 2027. Each measure gets one of four verdicts, supports, falsifies, inconclusive or untested; a silent record scores untested, never as nothing happened, and no measure treats a cancelled agreement as proof that its developer was speculative.
The first four measures are in the table. The third, the cost shift, is measured against the price the same costs and peak would give without the new large loads, because the plain price cannot be the test: it rose about 22 percent from January 2025 to August 2026 with every peak behind it measured before ComEd announced a single TSA, and 1 percent of the zone's yearly cost is about $9.7 million.
The fifth is where the money goes when a developer leaves: for each cancellation FERC has accepted, Karis Critical's and Monarch Rock Air's so far, whether a termination payment or forfeited collateral went to the other customers and covered any cost shifted to them. A payment with no sign of where it went, or nothing paid and no cost shifted, is inconclusive; nothing shown is untested, which is where it stands.
The sixth, frozen on October 20, is free exits. Of the ComEd agreements that end by a filing with FERC made from October 21, 2026 through October 20, 2027, it counts the share that left through a door that costs nothing and is not kept for delays, regulators or outside events: a withdrawal before FERC accepts the agreement, a termination that takes effect before the agreement does, or mutual consent with the fee waived in writing. One in three or fewer supports, the agreements bind; two in three or more falsifies, a signed agreement is in practice a free option. Fewer than three exits in all settle nothing, and a consent exit, or a notice that names no clause, with neither payment nor waiver shown is scored both ways.
| Measure | Where it stands today | Supports if | Falsifies if |
|---|---|---|---|
| T1 Load taking service under TSAs, as a share of the scheduled minimum | untested (undisclosed) | 100% or more | under 50% |
| T2 Large-load transmission capital entering the plan, and its equity share | inconclusive ($1B+ beyond the plan; equity ~40%) | $1.0 billion or more at 40% equity or less | none enters and the beyond-plan figure falls below $1.0 billion |
| T3 Effect of the new large loads on ComEd's zonal rate | untested (no large-load plant or MW identified) | 0% or below | +1% or more |
| T4 Share of the high-probability block with signed TSAs | 36% to 40%, inconclusive | 50% or more | under 30% |
| T5 Termination payments on accepted cancellations | untested (Karis Critical and Monarch: fee promised, none shown) | credited to the zone or held for customers | costs recovered from others beyond what was paid |
| T6 Free exits among ComEd TSAs ending by filings from Oct. 21, 2026 to Oct. 20, 2027 | untested (window opens Oct. 21) | one in three or fewer | two in three or more |
No measure sits in its support or falsify zone today, which is as it should be: the test can fail.
Before the scoring dates come unscored signposts: Exelon's third-quarter results on October 30, 2026, where a block that falls while the future pipeline grows by about as much is the delay's pattern and a block that falls while the pipeline holds is the screen's; FERC's decision on PJM's rules; and any ruling in Hillwood's suit that reads the credit-support terms.
Likely, we can get some of it from primary research, information gathered by asking people who know first hand, which is what Vista is all about. Four things would settle the most. The first three are public records that exist and are closed to us, or that we have not yet read; the fourth, and the practice behind all of them, is where the interviews below come in.
The people who ran these queues, discounted these forecasts, wrote these formula rates and filed these requests know what the documents cannot say, and finding and asking them is what Vista is really about. We would most like to hear from these.
As a matter of general past practice, of the large-load requests that signed a security agreement with minimum payments, what share went on to take service near their contracted size, and what made the difference? We would also ask what happened to upgrades already built or ordered when a customer left early, whether a termination fee was ever collected or waived, and how billed megawatts compared with the contract minimum in the first years.
The answer that would change our view"Fewer than half of the signed requests came near their contracted size, and those that left paid far less than the cost of the plant built for them."
As a matter of general past practice, how did planners discount speculative large-load requests in the load forecast, and how often did the discounted requests come back? We would also ask how planners found a project that had asked several utilities, and which way the forecast errors ran.
The answer that would change our view"We counted almost none of the unsigned requests, and a large share of them came back anyway."
In general, when a large-load customer leaves early, how are the termination payment and the stranded plant treated in the formula rate, and who bears the remainder? We would also ask how often a utility waives a termination fee by consent, and the difference in practice between a regulator accepting an agreement and independently finding that it protects other customers.
The answer that would change our view"In practice the termination fee is waived more often than it is paid, and the plant stays in rate base for its full life, so the remainder falls on the customers who remain."
In general, how many utilities does a developer file with for each site it builds, and what deposit or minimum payment makes a developer drop a duplicate request? We would also ask what it cost to withdraw a request, and how often a signed tenant existed when the request was filed.
The answer that would change our view"A developer files with five or more utilities for every site it builds, and no deposit a utility asks for is large enough to make it drop the duplicates before it knows which site wins."
Every conversation would draw on past, general experience only: no current employees of the companies named, nothing confidential, and every Advisor, as Vista calls the experts it interviews, screened for conflicts before a word is said. For a client engagement, Vista would combine this inquiry with those interviews: two or three Advisors, structured written conversations, and one brief, a short report, that says where they agree, where they split, and what would change the answer.
Four notices at one utility are a small thing to have read so carefully. ComEd's agreements are the smallest and best-lit instance of a problem that runs through the AI trade: sellers valued on backlogs their customers can cancel for a fee.
The grid first. Dominion Energy Virginia counted about 53.8 gigawatts of data center load under some form of contract in July 2026, 12 of them in signed service agreements. Texas is the largest and the loosest: ERCOT, the Texas grid operator, tracked about 438 gigawatts of large-load requests in May 2026, the vast majority with no study filed.
GE Vernova, a turbine maker, counted 116 gigawatts of gas equipment under contract at June 30, 2026, 63 of them slot reservations, deposits against a place in the factory's line, which its own annual report says "may not result in orders"; conversion into orders is the only door it shows.
Then the backlogs that hold up the AI trade itself. Oracle's remaining performance obligations, the contracted revenue it has yet to deliver, were $638 billion at May 31, 2026, by its annual report, and $664 billion a quarter later, by its release; Microsoft's commercial remaining performance obligations were $678 billion at June 30, 2026, up 84 percent in a year; CoreWeave, which rents AI computing, had $103.7 billion under take-or-pay contracts, paid whether used or not, of about five years.
Aircraft order books and homebuilders' backlogs are older versions of the problem, and the one place where a haircut is already applied, Boeing's accounting adjustment for orders whose financing is in doubt.
The method needs three public things, the fee schedule, dated exits and the door each exit used; here is who discloses what, and how the rest could be learned.
| Candidate | Fee schedule public? | Exits with dates public? | Door public? | How we could learn the rest |
|---|---|---|---|---|
| ComEd TSAs (Exelon) | Shape only; amounts redacted; "tens of millions" in the first 180 days, by ComEd's description | Yes, each a dated FERC filing (four so far) | A reason in each; a clause named only for Monarch | ComEd's yearly formula rate update; the Illinois commission's docket; former ComEd account managers and regulatory counsel |
| AEP Ohio | Yes, in full | Counts only: about 30,000 MW asked, 13,022.7 studied, 5,642 signed | Only "did not sign" | Data responses in the Ohio commission's case 24-508; former AEP Ohio customer staff |
| Dominion | Yes, in full, from 2027 | Stage counts each quarter; no exit list | Stage only | The yearly report the Virginia commission requires; former Dominion account managers |
| Georgia Power | Rule public; contract summaries from 2026 | Quarterly flows: entered, exited, modified | Split three ways; no reason given | The docket's attachments; the early termination letters due within 15 days; former commission staff |
| ERCOT and Oncor | Rules and collateral totals only | Request counts by quarter; "some project cancellations" | No | Public records requests; the transmission providers' interconnection engineers |
| Generator queues (LBNL, PJM) | Yes, in full: Order No. 2023 deposits and penalties | Yes, 24,221 withdrawals since 2000; the date missing for about 37 percent overall and about 14 percent in the seven grid operators | Withdrawn, deemed withdrawn, suspended; the phase at exit for about half | FERC compliance filings on penalties; developers' interconnection managers |
| GE Vernova and Siemens Energy | No; deposits named in the cash flow discussion | Quarterly bridge in gigawatts | Conversion only | The contract liability notes; slot data vendors; procurement heads and former OEM sales directors; the companies, asked |
| Oracle and Microsoft | No | No | No | The notes on leases not yet begun; former capacity planners; wholesale leasing heads and brokers; the companies, asked |
| CoreWeave | No; terms only: take-or-pay, about five years, termination for cause | None yet | None yet | The 8-K exhibits; project finance lenders; rating agency analysts |
| Boeing and Airbus | No | Monthly, in aggregate by model | Cancel against convert only | Fleet databases; lessors' reports; fleet planners and former OEM contracts managers |
| Homebuilders | The deposit is the schedule; forfeits in dollars at Toll | Quarterly rates; the 2006 to 2009 record | Financing contingency and walk-away not split | Division presidents and sales managers |
| LNG (Cheniere) | Yes, in full | Quarterly, by notice date, with the fee paid | Cancel against suspend; the fee shown as revenue | Cargo trackers and brokers; LNG traders at the buyers; Cheniere scheduling staff |
Read down the columns and the pattern is plain, and unwelcome. The backlogs worth the most, the AI cloud ones, disclose the least: no schedule, no dated exit, no door. The utilities disclose the most, because regulators make them; the turbine makers sit between. So the tool is partly a reading of public records and partly primary research, which is what Vista does: the people who see the exits know the door and the fee when no filing says it. Where the record is silent, the second route is to ask the company, and the third is to ask the regulator to require it, as Georgia's commission staff did in August 2026, when Georgia Power agreed to report every early termination notice within fifteen days.
Before the tool is pointed at a backlog whose exits nobody discloses, it should be run where history and public data allow, with the pass and fail lines written first.
The first place is the grid's own queue for new power plants, from 2017 to 2027. The fee steps there are known, and some came before a project's own cost estimate: in MISO, a Midwestern grid operator, a request in the cycles after FERC's order of December 2019 got its milestone payment back in full only if it left before the study of its costs began, and FERC's Order No. 2023 of July 2023 made that shape the default wherever FERC regulates, no penalty before the first study, then penalties that step up by stage. A project that leaves just before one of those dates cannot be answering its own cost estimate, because none has arrived, and that separates the fee from the cost news. The regulators' own refund rules then label the exits a cost shock caused: PJM returns a project's deposits when its share of upgrade costs has jumped by at least a quarter and by a set sum a megawatt, and Order No. 2023 waives its penalty on the same kind of rise. What would show the tool works: exits bunched just before the dates money goes at risk, among projects with no bad cost news; more bunching where the step is bigger; and a forecast of how much of a cohort leaves at the next step that beats a cost-based model of their kind. What would show it does not: no bunching beyond what cost news already explains. We have not run this test, and until it is run the fingerprint's claim to read a backlog is a hypothesis, not a result.
The second place is the LNG cargoes of 2020, where the fee paid is known: Cheniere's customers may cancel a cargo with notice and still owe the fixed fee, and in 2020 they did, $969 million of such revenue that year, booked the day each notice arrived, while quarterly volume fell by nearly 40 percent. A tool that reads those exits as an option's fingerprint has failed, because the fee was owed in full and booked as revenue, which is the trace of a real customer and not of an option; a tool that reads them as demand buying flexibility, and would have left Cheniere's backlog without a haircut, has passed.
The bet this inquiry makes. That a utility which makes requesters pay to hold their place can turn a flood of options into a smaller, surer book of customers, and that Exelon's version of the screen keeps the payments with the customers who would otherwise pay for the plant, so that its data center growth is both real and neutral for everyone else. A reader who takes Exelon's 11 gigawatts and $1 billion at face value is betting that the collateral is collateral, that the billed share holds above the break-even, the 43 percent of reserved capacity computed above, and that a developer who walks away pays the fee the agreement names, through a door that is not free. Behind it stands a larger bet, which this inquiry does not score: that a commitment can be read by how it ends, and that the reading works on any backlog of cancellable commitments.
The payoff. If the bet is right, each $1 billion of large-load transmission adds about three and a half cents a share a year under the plan's own funding, a little over 1 percent of this year's earnings, expected largely from 2030 with some of it sooner, and the other customers' price holds or falls. If it is wrong, the same plant earns the same return while the zone pays for it, until a regulator notices, and the growth is a transfer rather than a gain. The larger payoff is not in Exelon. It is a way to read the backlogs that hold up the AI trade: Oracle's $638 billion and Microsoft's $678 billion of remaining performance obligations at midyear 2026, CoreWeave's $103.7 billion, GE Vernova's 116 gigawatts of turbines under contract, the 53.8 gigawatts Dominion counts and the 438 Texas does. Each is a headline that counts options and real demand alike, and the exit fingerprint, if it works, a chance we put at about one in seven, and under one in twenty for reading the AI backlogs directly today, turns the exits each of them discloses, or can be made to disclose, into a haircut on that headline; a haircut on a backlog of that size is worth more than any figure in this inquiry.
Our assessment. The structure favors the bet: the Illinois tariff credits shortfalls to the zone, ComEd has promised FERC in writing that termination fees go there too, FERC's show-cause order, its June 18 question to PJM, points the same way, and Exelon's plan does not count the money yet. The evidence does not yet test it. In the two exits that are complete no fee is shown collected: one, we infer, before any fee was owed, one by consent with the fee paid or waived and neither shown; the other two notices were rejected, and the one agreement whose security is in dispute is before a court, over when the full letter of credit was due. Read by the method of this inquiry, both completed exits fall where a cheap option would leave, and where most real projects that were delayed or fell through would too; that timing can never settle it, and the next year's exits, read by the door they use and by whether they keep coming once the fee has stepped up, can say what this year's cannot. In 1846 Parliament's 10 percent, lodged in court and returned when a bill failed, did little to cool the speculation.
What settles it, and when. ComEd's 2027 formula rate update, due by May 15 and posted by PJM in June, and Exelon's 2026 annual report, scored against the five bands frozen on October 18 by June 30, 2027; every ComEd cancellation filed at FERC in the year from October 21, 2026, scored by the door it uses by December 31, 2027; the first filing that shows what a departing developer paid and where it went; before any of that, the conversations above with the people who ran the queues and wrote the rates; and, before the tool is pointed at anything larger, the two backtests above, on the interconnection queue and the 2020 cargoes, with their pass and fail lines written first.
The gentlemen of the Commons could count bills and pounds and deposits, and in April 1846 Peel counted them to the pound. What they could not count was which of the 519 schemes before them were bona fide, and the deposit was the instrument they built to make the schemes count themselves. It worked badly: Campbell judges 10 percent far too small to restrain shares bought for a fraction of their price. A hundred and eighty years later the instrument is a signature on a ten-year guarantee with a fee for leaving that is smallest at the start, and the question is the same one Peel put to the House: whether the deposit is large enough to discourage unsound speculation, and whether, this time, it stays. The Commons could at least count its deposits. The backlogs that hold up the AI trade are counted in the hundreds of billions of dollars, and no one outside the companies can yet say what it costs to leave one, or who has.
This inquiry began with Jennifer Hiller's report in The Wall Street Journal of October 13, 2026, "'Phantom' Data Centers Are Demanding More From the Power Grid Than They Need," which we read in full. The comparison at its heart, Exelon's 18 gigawatts against its 11, comes from it, and we confirmed both figures in Exelon's own investor decks, the slide presentations a company shows its investors; the Wood Mackenzie and Vistra figures above are the article's, credited where they appear. The question the article does not ask, where a departing developer's money goes, is the one this inquiry set out to answer. The inquiry grew from a report I developed with my AI research assistant and Vista's research desk; Vista Research is the firm that publishes these inquiries. The closest published work is credited in "What is new here, and what is not."
The facts are public and others found them first. Hiller's report in the Journal raised the question and the word. Chang's concurrence holds the three doubts this inquiry is built on, the unassessed terms, the rolled-in upgrades and the forty-year plant against a ten-year promise; ComEd's own filing holds the $950 million and 6,700 megawatts; and FERC's orders and ComEd's own filings hold the fee, its promise to the other customers and the eleven doors, for anyone who reads them. Campbell and Odlyzko told the railway story with the numbers.
Nor is the idea behind the fingerprint ours. Economists have long read causes from exits that bunch where the price of leaving steps, and from the way people leave: retirements timed to the steps in a pension, people leaving the unemployment rolls as their benefits run out, mortgage defaults split into those who walked away and those who could not pay. Kleven's review of 2016 gathers the first kind under the name bunching; Ganong and Noel's study of 2023 is the closest twin of our option share, since it measures how many underwater borrowers, those owing more than their homes were worth, defaulted by choice against a comparison group with no reason to. The power plant queue itself has been read this way: Sarah Johnston, Yifei Liu and Chenyu Yang, in a paper revised in August 2026, read the timing of withdrawals around the dates study results arrived and find that most follow cost news, in years when leaving the queue cost little. Analysts already read some doors by hand, as TD Cowen did in 2025 when it split Microsoft's data center pullback into leases voided, agreements allowed to lapse and deferrals. Even the free option is not our framing alone: a 2026 paper by Jing on Japan's battery queue is titled "The Free Option to Congest the Grid"; "a free option with a signature on it" is our phrase for the same thing.
What we add is the reading of the four notices by Hasselmann's method, with the patterns drawn from the agreement's terms, a dated test written before the next year's exits, and the white shoe of Hempel's raven naming what weighs nothing. From that reading come the findings: that the stated reasons cannot separate an option from a slipped project, that the timing of the exits so far cannot either, and that the doors of the next year's exits can. And around it, the fee set against the collateral per megawatt and against the screening band, the range of forfeits in which a menu sorts; the ceiling calibrated to ComEd's own formula rate and its 6,700 megawatts; the load-side agreements read as a screening menu from the buyer's side, with the published menus priced on one scale; and six thresholds frozen before the first disclosure, with untested kept apart from absent.
And the tool? Cancellation rates are already watched, every quarter, for homebuilders and for aircraft makers, and Boeing's accounting adjustment is already a haircut on its order book; nothing here invented the idea that a backlog can shrink. What may be new is narrower, and we claim only this: using the fee steps written into commercial commitments to measure how much of a backlog leaves when leaving gets more expensive, read together with the door each exit uses and whether the fee was actually collected, with the regulators' own refund rules as the labels for exits forced by cost, and tested as a dated forecast against the headline backlog and the measures already in use. Our reading of ComEd's cancellations, by the fee step, the door and the fee shown, is new so far as we can find, in that narrow sense only. Our own adversarial review, a reader set to attack the draft, put the odds to us plainly, and I pass them on the same way: about one chance in seven that this becomes a genuine contribution to investment analysis, most of it through the queue test above; under one in twenty, today, for reading the AI backlogs directly, until those companies disclose their exits. I would rather name the odds than the hope, and the queue test is how we find out.
A request for grid power by a data center developer that will never be built. Asking a utility for power costs little next to what a place in line is worth, so a developer without a tenant or financing can file at several sites and build at one. The Wall Street Journal's word for such requests is phantoms; this inquiry uses it for a type of request, never for a named developer, because a real project can fail for many reasons.
In part, and the record cannot yet say how well. Exelon's count of likely data center load fell from about 18 gigawatts in February 2026 to about 11 in July, and 4 of the 11 have signed ComEd transmission security agreements, ten-year payment guarantees that Exelon says are backed by $1 billion of collateral. Every agreement carries a fee for leaving early, which ComEd told FERC would be credited to its other customers, but in the two exits that are complete no fee appears.
Less than the coverage suggests. Each notice named a withdrawn or cancelled application for service, which fits a request held as an option and a real project that slipped alike. Two notices ended agreements, Karis Critical's just before it took effect and Monarch Rock Air's by mutual consent; FERC rejected the other two, so those agreements stand, and Hillwood's, whose letter of credit was for $1, is before a court. Both completed exits came in the first 180 days, while leaving cost the least, which cannot tell an option from a delayed project.
Klaus Hasselmann's optimal fingerprinting, the method behind climate attribution, carried to contracts: write down in advance the pattern each cause would leave, let the data count most where the patterns differ and the noise is quiet, and say how much more data it would take for a signal to stand out. For a commitment the evidence is how it ends: when the exit came relative to the date the fee steps up, through which door, and whether a fee appears afterward. The inquiry froze a test on October 20, 2026 that counts the share of ComEd's next year of exits that leave for free.
The order books, slot reservations and remaining performance obligations that hold up the AI trade are backlogs of cancellable commitments of the same kind: Oracle's $638 billion and Microsoft's $678 billion of remaining performance obligations at midyear 2026 among them. None discloses its exits the way ComEd's filings do. Reading a commitment by how it ends could turn the exits a backlog discloses into a haircut on its headline, and the inquiry names the backtests it must pass first.
Every figure in this inquiry is computed in code from public documents, with each input traced to its source in the inquiry's working files. The yearly cost per dollar of plant uses the equity share, allowed return, debt cost and composite tax rate, federal and state combined, in ComEd's 2026 to 2027 formula rate workbook, posted by PJM, with depreciation and operating cost assumed at 2.5 percent and 1 percent; the zonal rate, revenue requirement and peak load are PJM's posted figures as of August 2026 and the 2026 network service peak loads. The ceiling divides the new load's yearly payments at the zonal rate by that cost; the rate changes add the plant's yearly cost to the revenue requirement and the billed load to the peak. ComEd's 6,700 megawatts and $950 million are its own preliminary figures as quoted in Commissioner Chang's concurrence, and the 75 percent minimum is her hypothetical; the 64 and 80 percent readings are the Karis agreement's own clause. The shareholder figures multiply $1 billion by the equity share and allowed return, fund 40 percent of the capital with new shares, and divide by Exelon's six-month diluted share count of 1,027 million; the main case prices the new shares at $43.29 and $44.03, the weighted-average net prices at which Exelon's forward sale agreements settled in 2026, and the retained-equity and holding-company-debt cases are sensitivities with their assumptions stated. No current Exelon share price is used anywhere. The screening illustration prices every published menu at ComEd's 2026 rate so that they compare; its type parameters are assumptions, it discounts at 10 percent a year, and it charges posted security 2 percent a year for two years to carry. The fee per megawatt spreads ComEd's own words, tens of millions and hundreds of millions, over the 1.8 gigawatts Utility Dive reports for the Hillwood project; the amounts themselves are redacted, the agreement's own scheduled megawatts are not public, and the first-180-day verdict would flip only if the agreement covered less than about 421 megawatts. The 180-day step is known for that agreement alone, from ComEd's answers, and is assumed for the other three from the shared form, whose public text sets the fee by the year in which termination takes effect. The fingerprint reads the four notices against two patterns drawn from the agreement's terms on October 20, 2026; we knew the notice dates when we wrote them, so only the sixth measure meets the method's rule that a pattern be fixed before its data. The exits needed in "The fingerprint" are our design review's figures, recomputed in model.py from the window-share equation printed there with p_0 at 5, 25 and 73.5 percent: to detect, the normal approximation to the binomial for a one-sided test at 5 percent with 80 percent power, a power an exact binomial test first reaches at 26, 96 and 645 exits; to estimate, the exits at which the estimated share's standard error falls to 5 points, which for a count of exits is exact. At ComEd's unrounded 73.47 percent the two are 644 and 1,538. The door count's error rates are exact binomial chances, scored by the sixth measure's own rule with each exit's door and fee taken as shown. The timing step takes the option as always leaving inside the first 180 days and a slipped project's exit as timed by events unrelated to the fee, so that a completed exit inside the window, read when an agreement is a given number of days past acceptance, favors the option by at most that number over 180; the agreements took effect one to five days after acceptance, which changes no conclusion, and only the two completed exits are counted. The duplicate arithmetic takes the Journal's two, three or four utilities per project with at most one build each. Exelon's pipeline figures come from its first-quarter, second-quarter and fall 2026 investor decks, read as text extractions, with slide 28's layout taken from a rendering of the slide; Jeanne Jones's words on the July 30 call come from a third-party transcript. The FERC material was read at source in the commission's electronic library on October 20, 2026: the Karis Critical agreement and ComEd's covering letter, the order accepting it, the four cancellation notices and the orders that followed, ComEd's answers in the Hillwood docket and the commissioners' statements; the Red Energy and Monarch agreement texts were not read, and Exhibit E is redacted in every public copy. The effective dates of the Red Energy, Monarch and Hillwood agreements are those ComEd requested, printed in the Federal Register's notices of the filings; the orders accepting them were not read, and Karis Critical's was accepted as requested. The railway figures come from Hansard, the London Gazette, Campbell's peer-reviewed papers and chapter, and Odlyzko's working paper, each caveat kept. The frozen test's bands are constants in model.py with a fingerprint that stops the program if any is changed; its bookkeeping rules, beyond those in the text, are that a disclosed range is scored at both ends and scores inconclusive if the ends land in different zones, that a source missing on the scoring date leaves its measures untested until it appears, that a change of definition by Exelon or ComEd is scored with the same bands and recorded beside the score, and that measures are computed from disclosed amounts where the sources give them and otherwise from printed percentages; its scoring rules pass 92 known-answer checks, and the model as a whole 392. The first five measures were frozen twice on October 18, 2026, the second time, before anyone had read the test, to name the Karis Critical cancellation the first freeze had missed, with no pass or fail line moved. The sixth was added on October 20 with its own dates and re-frozen that evening, before anyone had read it, after the fact-check of this draft: its text had said FERC's Hillwood order called the form standardized, which was ComEd's argument, and FERC called that agreement bilateral; its first free door now names a withdrawal before acceptance or a termination taking effect before the agreement does; and its minimum reads fewer than three exits in all. No band, date or rule moved, and the program rebuilds each earlier freeze from the present test and checks that it reproduces exactly. The figures in "If it works here" are quoted from the documents named in the Sources, not computed, except three: the share of withdrawals with no date in the seven grid operators, about 14 percent, is what Heidari's 86.3 percent leaves; Cheniere's fall of nearly 40 percent sets the second quarter of 2020 against the first, from the volumes in its second-quarter release; and the gap in GE Vernova's reservation bridge, the quarter's opening reservations plus new ones less those converted, set against the reported closing figure, differs by at most a gigawatt in each of the three quarters, which is within the company's rounding. The sum of Oracle's and Microsoft's remaining performance obligations in the dek adds Oracle's $638 billion at May 31, 2026 to Microsoft's $678 billion at June 30; with Oracle's later $664 billion it is larger still. Where a figure in that section comes from a press account rather than a filing, the text says so. The calculations are in model.py and charts.py.
This inquiry is research, not investment advice. It states what the filings and models show under labeled assumptions; nothing here is a recommendation to buy, sell or hold any security, and the decision belongs to the reader. Russ W. Rosenzweig, Vista's founder, owns shares of American Electric Power, Cheniere Energy, CoreWeave, Dominion Energy, GE Vernova, Lennar, Meta Platforms, Microsoft, Oracle, Sempra, the majority owner of Oncor, and Vistra, and holds options on CoreWeave, Microsoft, Oracle and Southern Company, the owner of Georgia Power; he owns no shares of Exelon, Siemens Energy, Boeing, Airbus, D.R. Horton, PulteGroup or Toll Brothers. Holdings through mutual funds and exchange-traded funds are not counted. How this inquiry was made: written by Russ W. Rosenzweig with Vista's AI research desk. It began with the Wall Street Journal report credited in "The idea," read in full, and a report I developed with my AI research assistant and Vista's research desk; the facts were gathered from Exelon's investor decks and filings, FERC's electronic library and the Illinois commission's documents, PJM's posted rates and the parliamentary record, the models were built and every figure computed in code, each public source was checked against its original, and the draft was fact-checked and read for plainness before I read it. I read it and listened to it before it was published. Why I write with AI. Why I publish these inquiries.