Somewhere between two drafts of the same contract, a number moved by $130 billion, and the concrete did not notice.

In the version of the deal described earlier this summer, Nvidia was prepared to guarantee roughly $250 billion of obligations attached to the data center campus OpenAI wants in Pike County, Ohio. In the version now reported to be close to signing, the guarantee is under $120 billion, and it covers only the first phase of the project — about 5 gigawatts of a site planned for 10. The Wall Street Journal reported the revision on Friday; Reuters and The Information have since described the same restructuring, and neither company has confirmed the terms.

Nvidia did not get poorer between drafts. Its shareholders got louder. When the $250 billion figure became public, the stock fell about 5 percent in a single session. The fix was not to reduce the amount of steel, silicon, or electricity involved. The fix was to reduce the portion of the obligation that has to be described out loud, and to defer the rest to a later conversation about the second 5 gigawatts.

It is worth being precise about what a guarantee is, because the headline number invites confusion. Nvidia is not writing a check for $120 billion. It is co-signing. The backstop covers OpenAI's lease payments and the construction debt that SB Energy — the SoftBank subsidiary developing the site, in which OpenAI is itself an investor — needs to raise in order to build. The purpose is to make lenders comfortable enough to price the debt cheaply. A contingent liability costs nothing on the day it is signed. It costs everything on the day it is called, and only then.

So the structure, read from the outside, is this. A chip company guarantees the lease of a customer that will use the building to house that chip company's chips, on a campus financed by lenders whose willingness to lend is a function of the chip company's signature. The circularity is not a flaw anyone is hiding. It is the mechanism. Nvidia is also in talks to put roughly $3 billion of equity into SB Energy, half on signing and half at an initial public offering expected to raise about $5 billion as soon as next month, according to The Information. The chips themselves are a separate financing question, one that has been described as potentially reaching $350 billion.

The physical facts are unglamorous and enormous. The site is federal land: the Department of Energy's Portsmouth complex, an old uranium enrichment property. Ground was broken in March. The first phase, roughly 800 megawatts, is expected to begin operating in 2028. Fully built, the campus would draw power on the order of a mid-sized American state and cost something near $500 billion at today's prices for chips, labor, transformers, and skilled electricians — prices that are elevated precisely because everyone is building at once. The federal government controls the power supply to the site. Japan is contributing funding under a recent trade agreement. OpenAI is still negotiating a binding lease for the full 10 gigawatts.

Against that sits the counterparty. OpenAI is valued at $852 billion and does not make money. This is not a slur; it is the premise of the investment case, which holds that the losses are the cost of buying a position in something enormous. But a lease is not an investment thesis. A lease is a schedule of payments due on specific dates regardless of whether the thesis has come true yet, which is exactly why a guarantee was required in the first place.

The reduction from $250 billion to $120 billion is being read in the market as prudence. Read it instead as a measurement. It tells you the size of the promise that Nvidia's shareholders will tolerate seeing in one place — not the size of the promise the project requires. The remaining 5 gigawatts still needs financing. The chips still need financing. Those conversations have been moved, not resolved.

When a builder splits a commitment into a visible half and a later half, the later half is where the interesting terms end up. Nobody reads the second document. It is filed on a Friday.

Russia's Ministry of Defense said it destroyed 822 Ukrainian drones over the course of Saturday night and Sunday. Moscow's mayor, Sergei Sobyanin, said 600 of them were headed for the capital and that 201 were downed over the Moscow region alone. Russian officials called it the largest-scale Ukrainian attack of 2026, and for once the superlative appears to be arithmetic rather than rhetoric.

The dead, counted by both governments, came to at least nineteen.

In Russia: five people killed in the southwestern Rostov region, where more than 150 drones struck three towns, damaged homes and a railway station, and set a forest on fire. An 83-year-old man killed when a drone hit a private house in the Moscow region. A woman killed when a civilian bus was struck in Belgorod. One death reported in Kursk. A Wildberries warehouse in Koledino, about 25 kilometers from central Moscow, burned; three people were injured there.

In Ukraine: two killed and fourteen wounded at the ArcelorMittal steel plant in Kryvyi Rih, President Volodymyr Zelenskyy's hometown, with another person killed in the city on Sunday evening. A 66-year-old man and a 64-year-old woman killed by a guided aerial bomb that hit their house in the Zaporizhzhia region. One death each reported in Sumy, Donetsk, and Kherson. In Kyiv, air raid sirens sounded four times overnight and at least six people were hurt, including a child.

Russia's defense ministry said it had targeted the metallurgical plant at Kryvyi Rih and several military-industrial sites in Kyiv, including a facility producing components for Ukraine's Flamingo missile. Ukraine says its strikes inside Russia are aimed at military, industrial, and logistical infrastructure and are a response to four years of bombardment.

Zelenskyy put the week's total at more than 1,550 attack drones, nearly 1,560 guided aerial bombs, and 62 missiles across thirteen regions. The United Nations counted 437 Ukrainian civilians killed in July, the highest monthly figure since May 2022 — a statistic that ought to complicate the widely held belief that this war has settled into a stable, low-intensity grind. It has not settled. It has industrialized.

The spillover arrived on schedule. After the overnight raid, Romania's defense ministry said a Spanish F-18 flying a NATO mission shot down a drone that had crossed into Romanian airspace near the borders with Ukraine and Moldova. That is a NATO aircraft firing a weapon over NATO territory at a machine neither belligerent will claim, and it has now happened often enough that it produces a two-line statement rather than an emergency session.

What has changed in the last several months is not the intent on either side but the throughput. Eight hundred drones is a manufacturing statistic before it is a military one. Both countries have built the capacity to lose hundreds of airframes a night and do it again tomorrow, which means the ceiling on this exchange is set by factories rather than by generals. That is the part worth watching, and it is the part neither ministry publishes.

The Democratic Republic of the Congo's National Public Health Institute reported Sunday that the country's seventeenth Ebola outbreak has killed 2,325 people, passing the 2,299 deaths of the 2018–2020 epidemic to become the deadliest in Congolese history. Confirmed infections stand at 4,945, including 101 detected in the previous twenty-four hours. The outbreak passed the earlier record for total cases back in late July.

The number that should stop you is not the cumulative total. It is the trend in the case fatality ratio, which has risen from roughly 20 percent in early June to 46 percent now. Nearly one in two people with a confirmed infection is dying. A case fatality ratio that climbs during an outbreak usually means one of two things: the surveillance is finding fewer of the mild cases, or the sick are arriving at care too late and in worse condition. Both are indictments of the response rather than of the virus.

WHO Director-General Tedros Adhanom Ghebreyesus said on Wednesday that at the current pace the outbreak is on track to eclipse the West African epidemic of 2014–2016, which killed more than 11,000 people and produced, as its one durable legacy, a licensed Ebola vaccine.

That vaccine is Merck's Ervebo, approved in 2019, and it does not target this virus. The Congolese outbreak — declared on May 15 and elevated to a Public Health Emergency of International Concern on May 16 — is caused by *Orthoebolavirus bundibugyoense*, the Bundibugyo species, for which no vaccine and no treatment has ever been licensed. Ervebo was built against the Zaire species and has been deployed successfully against it many times in this same country.

So the WHO's Technical Advisory Group on Candidate Vaccine Prioritization met on July 31 and, reviewing non-human primate and ferret challenge data along with pseudovirus neutralization studies, recommended that Ervebo be put into a Phase 3 ring-vaccination trial in the current outbreak without a preceding Phase 2. The group found no safety concerns in the ring-vaccination context. It was explicit that efficacy against Bundibugyo transmission and disease remains unproven, that protection against symptomatic infection may be limited, and that a Bundibugyo-specific vaccine remains the preferred option. Two candidates are in early human trials; a third is in development. None of them will be ready for this outbreak.

The trial is designed by the WHO and will be run by a consortium including Congo's National Institute of Biomedical Research, CEPI, Belgium's Institute of Tropical Medicine, Médecins Sans Frontières, and Africa CDC. Gavi will supply doses from a stockpile of roughly 500,000. It could begin in Ituri province by mid-September — about six weeks from the advisory group's recommendation, against a normal clinical timeline of five to twelve months.

Six weeks is genuinely fast. It is also the wrong unit of comparison, because the outbreak has been running since May and the case fatality ratio more than doubled while the paperwork moved. The honest framing is that the world's Ebola preparedness consists of one licensed vaccine for one of six species, and when a different species surfaces in a remote conflict zone along the South Sudanese, Ugandan, and Rwandan borders, the plan is to hope for cross-protection and find out in public.

The experts say Ervebo may not stop infection but might reduce severe disease and death. Given a 46 percent case fatality ratio, that is not a consolation prize. It is the entire objective.

On Sunday, President Donald Trump posted that the annual joint military exercises with South Korea were expensive and sent "a signal that is totally inappropriate and hostile" toward North Korea, which he described as "unthreatening and respectful" during his time in office. He cited his "very good relationship" with Kim Jong Un. He conceded it was too late to cancel and said he had directed Defense Secretary Pete Hegseth to "substantially reduce" the drills.

On Monday, South Korea's defense ministry said Ulchi Freedom Shield had begun as planned.

The exercise runs eleven days, August 17 through 27, and involves 18,000 South Korean troops. It is a fixture of the Korean summer, built around the defense of the South against a nuclear-armed neighbor whose forces have spent the past two years acquiring combat experience in Russia's war on Ukraine. North Korea's foreign ministry denounced the drills on Friday, said American, Japanese, and South Korean cooperation was hardening into a nuclear alliance, and promised to answer a new level of threat with a new level of deterrent.

This is not the first time. After the 2018 Singapore summit, Trump called the same exercises expensive and provocative, and some were curtailed.

What is worth noticing is the sequencing. A field exercise of this size is not a switch. It is months of movement orders, fuel contracts, reservist call-ups, and airspace reservations, most of them executed by a sovereign government that is not the one posting. By the time the instruction to reduce arrives, the reduction that is physically available is at the margins: a cancelled flight, a shortened phase, a headquarters exercise substituted for a field one. Seoul can honor the letter of it and still run the eleven days.

So two things are true at once, and both governments can point at the one they prefer. Washington can say the exercises were scaled back in deference to a diplomatic opening. Seoul can say Ulchi Freedom Shield proceeded. Pyongyang gets to choose which version to be insulted by, and has historically chosen both.

The cost of this arrangement is not borne today. It is borne the next time an American commitment on the peninsula has to be legible to an ally in a hurry. An alliance runs on the assumption that the schedule and the statement describe the same event. When they routinely don't, the ally begins planning around the schedule and ignoring the statement — which is, in the long run, a much larger reduction than the one that was ordered.

Coming into last week, roughly two-thirds of futures traders expected the Federal Open Market Committee to raise rates in September. By Friday morning that had flipped: just under 70 percent were pricing no change, and 30 percent a hike — the lowest implied odds since the June 17 decision.

The data did the flipping. July retail sales fell 0.6 percent, the sharpest monthly decline since May of last year, and that was with a World Cup tailwind. Payrolls disappointed. Inflation came in elevated but cooling. The University of Michigan survey showed sentiment falling among both households and businesses. Taken together, a week of ominous readings for the American economy was received by financial markets as good news, because it lowered the probability of a rate increase — a reaction that tells you precisely which variable this market is trading.

Here is the part that is being underreported. The implied probability of a hike by year end is still around 90 percent.

A 30 percent September and a 90 percent December are not in conflict, but holding both means the market is not forecasting a pause. It is forecasting a delay. The tightening cycle remains live; only the calendar has moved. Traders who buy this rally on the strength of the September number are buying an asset whose discount rate they themselves expect to rise before the year is out.

The cross-asset evidence supports the split-screen reading. The ten-year Treasury closed at 4.68 percent on August 14, sitting just under the 4.74 percent level technicians have flagged as significant, while equities have been trading as though borrowing costs were a rounding error. The dollar index is stuck near 99.5 at the bottom of its monthly range; the euro pushed to a two-month high toward 1.1600, helped along by growing acceptance that the ECB delivers one final 25-basis-point hike in September. The bond market and the equity market are telling different stories and neither has conceded.

Two events will adjudicate. Wednesday brings the minutes of the July 28–29 FOMC meeting, which are likely to read hawkish and will show how far support for a hike extended beyond the three dissenters — Beth Hammack, Neel Kashkari, and Lorie Logan. Those minutes are also two soft data weeks out of date, which limits how much they should move anything. Then Jackson Hole, August 27 through 29, where Chair Kevin Warsh speaks into a market that has already placed its bet. Warsh declined after the July meeting to spell out the committee's reaction function, which is a choice with consequences: when the Chair withholds the framework, every data release becomes a referendum.

The asymmetry favors caution. A dovish Warsh confirms what is priced and buys a limited rally, capped by that 90 percent year-end figure. A hawkish Warsh pushes September odds back toward 60 or 70 percent, breaks the ten-year above 4.74, and compresses multiples hardest in long-duration growth, speculative technology, utilities, and REITs. The downside scenario has more room to run than the upside one, for the simple reason that the good news is already in the price.

And there is the consumer, who is about 70 percent of GDP and who stopped spending in July. Markets are currently rooting for a slowdown large enough to stay the Fed's hand and small enough not to touch corporate earnings. That is a narrow target, and nobody has ever hit it on purpose.

Two jurisdictions moved on data centers within a month of each other, in opposite directions, and between them they define the entire live argument about who pays for the AI build-out.

New York went first. Governor Kathy Hochul's July executive order made the state the first in the country to impose a moratorium on large data center construction: a one-year pause on any project drawing 50 megawatts or more. During the freeze, the Department of Environmental Conservation will not issue discretionary permits that have not already been deemed complete, and officials must produce a Generic Environmental Impact Statement setting consistent standards. The pause lifts when the standards are finalized, not on a fixed date — which means the industry cannot plan around it.

The 50-megawatt threshold is a scalpel. It captures hyperscale training campuses and leaves ordinary enterprise and colocation sites untouched. Behind it sits a queue of more than 12 gigawatts of large-load interconnection requests waiting on the New York grid operator, per its May figures.

The less-covered half is the money. Hochul has said she will pursue legislation repealing the sales tax exemptions large data centers currently enjoy — quietly one of the most valuable state-level subsidies available to an operator choosing a site. Freezing construction is a permitting decision. Repealing the exemption is a statement that the state no longer believes these facilities pay for themselves. The legislature is separately working a 20-megawatt bill, which would widen the net considerably.

New South Wales took the other door. Its new framework, effective Monday, promises assessment of a data center application within 75 days — in a sector where multi-year approvals are routine — provided the builder meets environmental and efficiency standards, imposes no net cost on consumers, funds additional energy and water supply, and shows power purchase agreements with renewable providers. There are more than 60 data centers operating or under construction in the state and 19 more in the pipeline. A water pricing review by the Independent Pricing and Regulatory Tribunal will follow, and a new Office of AI opens the same day.

The framework's weakness is stated plainly in the design: it sets no penalty for builders who decline the offer. It is a bargain, not a rule. Compliance buys speed. Non-compliance buys the old slow queue.

Both governments are responding to the same arithmetic, and it is not primarily about carbon accounting rules. The Financial Times examined 60 data centers under development in the United States by Amazon, Microsoft, Google, and Meta and estimated that in operation they could emit roughly 101.5 million tons of carbon dioxide a year. About three-quarters of the utilities serving those 60 sites are planning or building new gas generation; 17 percent said explicitly that new capacity was needed for hyperscale customers. Amazon reported a 16 percent rise in greenhouse gas emissions between 2024 and 2025, with emissions from purchased power up 34 percent. Microsoft reported a 25 percent increase. Alphabet, 18 percent.

The two policies test different theories of leverage. Albany assumes the industry needs New York and will pay for access. Sydney assumes the industry has options and will trade concessions for speed. Whoever is wrong will find out in the interconnection queue, which is the only ledger in this business that cannot be restructured on a Friday.

President Trump posted a photograph on Sunday of a large patch of dead grass on the National Mall near the World War II Memorial and attributed it to vandals. He offered no evidence for the attribution.

The lawn in the photograph is where a large stage was built for his July 4 celebration of the nation's 250th anniversary, an event he has said without support drew more than 400,000 people before a storm delayed his remarks. Images captured by a streaming webcam mounted atop the Washington Monument show a bandstand and seating areas covering nearly the whole of the affected lawn. Associated Press photographs from the following days show crews removing the structures and brown, flattened turf underneath.

The photographs do not, on their own, establish the cause of every dead patch. Grass dies for many reasons in a Washington August. What they establish is that the president's own event occupied the ground he is now describing as a crime scene, and that this was documented continuously from a fixed camera 555 feet above it.

This is a pattern with a short and specific history. Earlier this month, the office of U.S. Attorney Jeanine Pirro concluded that damage to the Lincoln Memorial Reflecting Pool was caused by poor construction rather than vandalism. Trump attacked her publicly for it. On August 1 he posted a different image of the same stretch of lawn appearing to show the numbers "86 47" pressed into the grass — a sequence federal officials have characterized as a threat against him — and offered it as a rebuttal: "a pure case of VANDALISM, that included the grass." He allowed that "there may have been some contractor difficulty," before adding that "the major damage was caused by VANDALS!" Interior Secretary Doug Burgum has described a pattern of vandalism on the Mall and cited seven arrests.

The numbers are not visible in Sunday's photograph.

It is tempting to file this under trivia, and the National Park Service reseeds the Mall constantly; the turf will recover long before the argument does. But the mechanism on display is the same one running through harder stories this week. A claim is made. A physical record contradicts it. The record is not disputed, because disputing it would require engaging with it; instead the claim is simply restated with more capital letters, and the burden quietly transfers to anyone who wants to go and look.

Most people will not go and look. That is what the method is for. It works considerably better on a defense posture, a permitting decision, or a set of financing terms than it does on a lawn photographed every ninety seconds from a national monument — which is exactly why the lawn is worth writing down.

The Pochaina book market sat by a metro station in Kyiv's Obolonskyi district and was never only a book market. It was part bookshop, part flea market, part accidental archive: rare and antiquarian volumes, vinyl, posters, toys, tools, Soviet household objects that nobody was preserving on purpose. It was founded in 1997. When developers moved to clear the site for a shopping center, the independent vendors organized and beat them, arguing in a 2020 petition that the market was a cultural institution supporting thousands of jobs.

Russian ballistic missiles hit it overnight on Saturday. A Kyiv city council member, Taras Kryvoruchko, said three missiles struck the trading rows. "This time, it seems, the story has come to an end," he wrote. "It is painful to see the disappearance of a place that was part of the lives of many Kyivans."

Firefighters were still working the site at midday Sunday. The Emergency Service's spokesman, Pavlo Petrov, said the blaze was extremely difficult to extinguish because of the volume of flammable material — which is a technical way of saying that a book market burns like a book market. Air raid alerts sounded repeatedly during the operation, forcing crews to withdraw to a safe distance and then come back. Three people were hospitalized, two in moderate condition.

Oleksandr Krasovitsky, who founded the publishing house Folio and works with several traders at Pochaina and with warehouses nearby, told the BBC he believes the market is lost for good. He put the number of traders working there at about 500. "Everyone who was there lost everything, except for the section of the market selling second-hand books." He estimated Folio's own losses at 1.5 to 3 million hryvnia and said damage to nearby warehouses may push it higher. A vendor named Luba said she lost about 500,000 hryvnia in stock. A stationery seller named Tetiana lost her stall. One owner of three years' standing described shutters torn out and shelves down and said he could not yet assess what he had lost.

Ukrainian officials have not said Russia deliberately targeted the market, and Russia's defense ministry said its overnight strikes on Kyiv were aimed at facilities supplying the military. What can be documented is that this is at least the third fire to hit Pochaina or the adjoining rows since mid-June: pavilions were damaged in June, and on July 30 falling missile debris ignited stalls and destroyed collections of antiques, vinyl, and Ukrainian household artifacts. It can also be documented that Russian strikes have repeatedly hit Ukrainian printing houses and publishing warehouses, and that Ukrainian literature has been removed from libraries in occupied territory.

The cultural loss will be written about in the register of symbolism, and it deserves some of that. But the concrete loss is a five-hundred-tenant small business district with no insurance market, destroyed in one night, whose inventory was largely irreplaceable not because it was precious but because it was accumulated — the specific unrepeatable sediment of thirty years of people bringing in what was in their apartments.

Some vendors in the damaged but unburned section say they will reopen. That is the detail to hold onto, and also the one that should not be used to soften the accounting. A market that reopens with a hundred traders instead of five hundred has not survived. It has been reduced, and the reduction is the point.

Every boom invents an instrument, and the instrument is always the one that lets the boom keep counting.

The railroads had the land grant. The 1920s had the investment trust that owned other investment trusts. The late 1990s had vendor financing, in which Lucent and Nortel lent competitive local exchange carriers the money to buy Lucent and Nortel equipment, and everyone booked the equipment as revenue and the loan as an asset until the carriers stopped existing.

The AI build-out has the guarantee, and it is the most elegant of the lot, because it does not require you to part with anything.

Consider the shape of the Ohio arrangement now nearing signature. A chipmaker agrees to backstop a lease and construction debt for a campus that will be filled with its own chips, leased by a customer that is one of its largest buyers and an investor in the developer. The lenders lend because of the signature. The developer builds because of the lenders. The customer signs the lease because it can, given that its obligation is co-signed. And the chipmaker records the eventual chip sales as revenue, with the contingent liability sitting in a footnote, measured in a different accounting language than the earnings it makes possible.

The cash does not move until something breaks. That is the whole appeal. A guarantee is the only way to inject certainty into a system without injecting capital into it.

And notice what happened when the number got too visible. The proposed backstop was around $250 billion; the stock dropped about 5 percent; the backstop is now under $120 billion, covering the first phase. The building is unchanged. The chips are unchanged. The gigawatts are unchanged. What was reduced was the quantity of promise held in a single, quotable place. The remaining exposure did not evaporate — it was moved into a later negotiation, alongside a separate chip financing discussion that has been described as potentially reaching $350 billion. This is not risk reduction. It is risk arrangement, and the arrangement is optimized for how the risk reads rather than for what it is.

The lazy version of this accusation is wrong. Nobody is committing fraud. Contingent liabilities are disclosed, guarantees are legitimate credit-enhancement tools, and the business at the other end is real: the compute is scarce, the customers want it, and someone has to underwrite construction that takes four years to produce a dollar.

The problem is narrower and worse. A guarantee prices the probability that the guarantor will be strong when the guarantee is called. But the guarantor is strong *because of the demand the guarantee creates.* The correlation is not a tail risk; it is the design. If AI capital expenditure slows enough that a data center tenant cannot make lease payments, that is the identical world in which the chipmaker's earnings have fallen and its ability to absorb $120 billion of somebody else's obligations is at its weakest. The backstop is most likely to be called precisely when it is least likely to hold.

Every boom instrument shares that property. The investment trust was solvent while the market rose. The vendor loan was an asset until the borrower's business model, which was also the lender's business model, turned out not to work.

None of this means the campus should not be built. Ten gigawatts of American compute on federal land is a serious national asset, and if I thought the projects were fake I would say so. It means only that we should be honest about which claim the market is actually pricing when it prices these deals. It is not pricing electricity or chips or model capability. It is pricing a signature, and the value of a signature is a function of the very thing it was signed to guarantee.

We will find out what it is worth on a day when nobody wants to.

The weekend's numbers, before the eulogies. *Spider-Man: Brand New Day* took $70 million in its third weekend, down 51 percent, for a domestic total of $785.8 million — the second-best third weekend in history and, as of Sunday, the fourth-highest-grossing film ever in North America. Christopher Nolan's *The Odyssey*, in week five, added $23.2 million to reach $504.6 million domestic and will pass *The Dark Knight* to become his biggest. *The End of Oak Street*, a PG-13 original with Anne Hathaway, Ewan McGregor, and dinosaurs, opened third with $21 million against an $80 million budget and a B CinemaScore. *PAW Patrol: The Dino Movie* opened fourth with $20.5 million against a $40 million budget and had already made $69 million worldwide.

The discourse writes itself and it is wrong. The audience did not reject originality this weekend. The audience gave half a billion dollars to a three-hour Homer adaptation.

*The Odyssey* is the fact that breaks the standard lament. It is not IP in any sense a studio marketing department recognizes; the underlying property is 2,700 years old and has no fandom to activate, no post-credits sting, no shared universe. It made $504 million domestically and became its director's biggest international release ever. What it had was a filmmaker whose name functions as a genre, a format argument that turned exhibition into an event, and a marketing campaign that started well over a year before release and treated the audience as capable of anticipation.

*The End of Oak Street* had two enormously likable movie stars and dinosaurs — a combination that should be, on paper, indestructible — and it opened to $21 million because almost nobody outside the trades knew what it was. A B CinemaScore suggests the people who did show up found it fine and will not evangelize. Reviewers who liked it, and there were some, are now writing the same sad paragraph: good movie, no chance.

That is not an audience failure. That is a distribution and marketing failure, and it is structural. The studios have not stopped making original films; they have stopped funding the eighteen-month campaign that makes an unfamiliar title into an appointment. Those campaigns now go exclusively to titles that already have awareness, which is a circular allocation that guarantees its own conclusion. Then the underperformance of the unmarketed original is cited as evidence that audiences only want sequels, and the money moves further toward the sequels.

Meanwhile the actually interesting data point of the weekend was in fifth and sixth place: *KATSEYE: WILD HEARTS* took $4 million on 724 screens, and *SIX: The Musical Live!* took $3 million on 1,462. Event cinema — proshot concerts and theatre captures — keeps quietly clearing per-screen numbers that would make a mid-budget drama weep. Audiences are not refusing to leave the house. They are refusing to leave it for something they have no reason to believe is worth the trip.

So here is the argument, and I would like the letters. Originality is not commercially dead. *Unpromoted* originality is dead, and it has been for a while, and the studios have decided this is a law of nature rather than a budget line they control. Nolan's real innovation is not the format. It is that he is the last director permitted to spend like a franchise on a film that isn't one.

Give *The End of Oak Street* the *Odyssey* campaign and we would be having a different conversation this morning. Nobody will run that experiment, because the result would be inconvenient for everyone who has already written the postmortem.