The deadline required nobody to do anything. That is precisely why it is the most consequential thing that happened this week.

The memorandum of understanding signed in Islamabad on 17 June gave Washington and Tehran sixty days to turn a pause into a peace. The sixty days ran out over the weekend. Neither government asked for more time. Asked on Monday whether the United States wanted an extension, President Trump said no — “They're not going to make the kind of deal that I feel is necessary” — and told Tehran to “put up the white flag of surrender.” Iran's foreign minister replied that his country had “won war and diplomacy.” A senior Iranian official told Reuters that Tehran was moving to a “fully offensive” posture and would strike the American naval cordon “precisely” if diplomacy stayed dead.

No one signed anything. Nothing needed signing. An expiry is the only kind of decision that executes itself.

## What the expiry looks like on the water

Six commodity vessels transited the Strait of Hormuz on Monday — three inbound, three outbound — against a ten-day average of eleven and a pre-war baseline of 120 to 140 a day, according to Kpler ship-tracking data. Three crossed on Saturday. Two on Sunday. Among Monday's six there was not a single very large crude carrier and not a single LNG tanker. The most important energy corridor on the planet moved about as much cargo as a provincial ferry service, and some of what did move went with its transponder off.

On Tuesday morning the UK Maritime Trade Operations centre reported that a vessel making an outbound transit had been struck by an unknown projectile. The impact damaged the engine room and produced one crew casualty. The Omani Coast Guard is assisting the rest of the crew. Hours earlier an Iranian military spokesman, Ebrahim Zolfaghari, had promised that ships attempting the passage would find “several beautiful holes in their hulls.”

## The people the ledger forgets

Roughly 6,000 seafarers remain aboard about 500 ships inside the Gulf, unable to leave. The International Maritime Organization built an evacuation framework in June and ran it for four days, from the 23rd to the 26th, moving 136 vessels and about 2,900 crew out along two improvised routes — improvised because the internationally agreed traffic separation scheme could not be used, on account of the sea mines sitting in it. The plan has been paused ever since. The IMO's secretary-general, Arsenio Dominguez, says he is still seeking guarantees that evacuating ships will not be shot at. He does not have them.

Seventeen seafarers have been killed in the strait since the war began in late February. Around seventy large oil and gas tankers have been stuck inside since then. A further sixty-five tankers sailed *in* after the June memorandum was announced, on the reasonable assumption that a signed agreement meant an open sea lane, and are now trapped with the rest. Savio Ramos of the Maritime Union of India describes conditions aboard as “like being in prison.” The men doing this waiting are, in the secretary-general's phrase, “innocent people trained to do their jobs, not to fight.”

## Freedom of navigation, as underwritten

Iran says the strait is open. The traffic data says otherwise, and so does the invoice. War-risk cover for a Hormuz transit is running at roughly thirty times normal — spot quotes for a VLCC voyage have gone from about $250,000 to as much as $10 million — which closes the waterway more completely than any minefield, because it closes it to every operator who intends to be solvent next year. Eight shipowner associations, including BIMCO, INTERTANKO and the International Chamber of Shipping, signed a joint letter on 3 August opposing any compulsory transit toll, which tells you what the industry expects the reopening to cost even if it comes.

An Iranian–Omani corridor is under negotiation; Tehran said on 5 August that the geography was agreed and a joint statement was in drafting. Foreign Minister Abbas Araghchi has been clear that a corridor is not a reopening, and Iran's security chief has said the Oman arrangement will not by itself reopen the strait. Even a clean settlement leaves around 520 commercial vessels queuing to get out, a backlog the energy intelligence firm Kpler estimates at six to eight weeks — assuming the mines are lifted promptly, which nobody is assuming.

Oil, which is the only part of this that Washington and Tehran both watch continuously, barely moved: US crude ticked up 0.25 percent on Tuesday to about $84.70 after a 2.6 percent rise on Monday. The market has already priced a closed strait. It has not priced the six thousand people inside it, because there is no instrument for that.

The war began with a declaration. It is being continued by a calendar.

Two things happened to the price of American money in the past fortnight, and they pointed in opposite directions.

At the short end, traders capitulated. The July payrolls report showed employers cutting about 23,000 jobs when economists had penciled in a gain of 83,000, with May and June revised down by a further 103,000 between them. Retail sales fell for the first time in nine months. Headline CPI came in at 3.4 percent year over year, cooler than June's 3.5. The market's implied probability of a September rate rise, which stood near 82 percent a few weeks ago, has collapsed to somewhere between 30 and 37 percent depending on which afternoon's futures print you prefer. Year-end pricing no longer fully embeds a hike at all.

At the long end, the opposite. The thirty-year Treasury yield closed Monday at 5.31 percent, the highest since 2007, and edged to 5.32 on Tuesday. The ten-year sits near 4.72. The gap between two- and thirty-year yields has widened to roughly 113 basis points, the most since April. This is a textbook bear steepening: the market believes the Federal Reserve less and charges Washington more.

## Where the extra yield is coming from

The useful decomposition is against inflation-protected paper. On 2 January the thirty-year nominal yield was 4.86 percent and the real, TIPS-implied yield was 2.63 percent, for implied inflation compensation of 2.23 percent. On 17 August the nominal was 5.31 and the real was 3.06 — implied inflation compensation of 2.25 percent.

Read that again. Nearly the entire 45-basis-point move this year is real yield. Investors are not demanding protection against future prices. They are demanding a larger fee for the act of lending to the United States for three decades, which is a different and less curable complaint. Supply is the mechanism: last week's $25 billion thirty-year auction cleared at 5.216 percent, the highest for that maturity since 2001, and the ten-year auction the day before drew the highest financing cost since 2007. Monday was settlement day for that thirty-year paper, which is how a fiscal argument becomes a market print.

## The Warsh problem

Chair Kevin Warsh has publicly treated rising long-term yields as something close to a policy instrument — a market-administered tightening that reduces the need for the committee to act. That framing has a cost. If the Fed will not use the short rate to demonstrate its commitment to 2 percent, the long end will keep charging for the doubt. Jim Bullard, formerly of the St Louis Fed, put the hawkish case on CNBC on Monday with the bluntness of a man no longer voting: “Why not do it now? You don't want to wait till the economy's suffering and then have to make an agonising decision.” Futures still assign roughly a 68 percent chance to a hike by December and higher odds into early 2027 — the market is not saying the Fed is finished, only that it is late.

Wednesday's July minutes will show how close the committee came, and whether anyone in the room said out loud that the long bond is now doing the tightening. Neither answer changes the arithmetic. A government running large deficits into 5.3 percent thirty-year money is compounding a political decision it has not debated, at a price nobody in office set.

The number everyone repeated on Monday was $105 billion. The number that explains the deal is $1.5 billion.

Nvidia's 8-K, filed 17 August, describes a multi-year partnership with SB Energy — the SoftBank-linked developer whose existing investors include SoftBank and OpenAI — to build out the PORTS Technology Campus in Pike County, Ohio. OpenAI signs a twenty-year lease as tenant for roughly 4.25 gigawatts of IT load, with an option structure that could take the site to about eight. Nvidia takes $1.5 billion of equity in the landlord and, in exchange, becomes the exclusive supplier of compute infrastructure for the first half of the campus.

That is the transaction. The guarantee is the packaging.

## Read what is actually guaranteed

Nvidia has not promised to pay OpenAI's rent. It has entered residual value guaranties. If OpenAI becomes insolvent or simply stops paying, SB Energy must first find a replacement tenant, then attempt a sale, and only then can it come to Nvidia for the shortfall between the guaranteed minimum value of the lease and whatever it actually recovered. The cap is $105 billion, and it covers asset value, not the lease.

In other words: Nvidia is underwriting the second-hand market for gigawatt-scale data centres in southern Ohio. It is a bet that if the largest AI company in the world walks away from 4.25 gigawatts of power and steel in 2029, someone else will want it at close to book. That is not a wild bet. It is not a costless one either, and it is a different bet from the one the headline implies.

The original structure, reported earlier this summer, was a $250 billion backstop for a ten-gigawatt build. Nvidia's stock fell about 5 percent, investors complained, and the figure came down — first to a widely circulated $120 billion, now to $105 billion attached to a narrower object. The building did not shrink. The disclosure did.

## LPS

Jensen Huang has started calling the constraint LPS: land, power and shell. The bottleneck in this buildout is no longer silicon or interconnect. It is dirt, substations and a roof. SB Energy and SoftBank will build generation supporting ten gigawatts and put at least $4.2 billion into regional grid infrastructure; first capacity is expected to come online in phases from 2028, starting around 800 megawatts.

This is the part worth watching in Ohio rather than on a chart. A guarantee written against residual real-estate value only works if the counterparty who eventually shows up values the asset the way the guarantor did. The Wall Street Journal puts roughly $3 trillion of AI infrastructure commitments outside anyone's balance sheet in structures resembling this one. Each is individually defensible. Collectively they describe an industry that has learned to finance itself by promising to be interested later.

OpenAI's president Greg Brockman called compute a “fundamental resource” on Monday morning. He is right, and resources are exactly the sort of thing that gets financed against, leased, securitised, and eventually repossessed. The interesting question is not whether the Ohio campus gets built. It is who is standing in the parking lot in 2032 holding the appraisal.

Two federal decisions moved large sums of public money on Monday. One of them was struck down within hours by a judge who wanted to see the authority. The other did not require a judge, a vote, or a rule — only a letter to the field.

The building. Judge Theodore Chuang of the US District Court for Maryland vacated the FBI's and the General Services Administration's decision to abandon Greenbelt, Maryland, as the site of the bureau's next headquarters and to move instead into the Ronald Reagan Building in downtown Washington. The site had been selected in November 2023 after a competitive GSA process against Landover and Springfield; Congress subsequently appropriated about $1 billion. The administration announced the switch on 1 July and set about reprogramming $555 million of Greenbelt money to “relocate to, renovate, build-out, and furnish” the Reagan Building.

Chuang's forty-seven pages hold that the agencies exceeded their authority and violated the Administrative Procedure Act, that Maryland and Prince George's County suffered irreparable injury, and that the reprogramming must be reversed. He enjoined further spending toward the move. The government had argued that rejecting the Reagan Building would strand the FBI in the Hoover Building indefinitely; the judge answered that vacating the decision reinstates Greenbelt as the selected site. Whether ground is broken is a separate question about appropriations and will. What the ruling establishes is narrower and more useful: an agency cannot redirect construction money by press release.

The letter. Also on Monday, HHS Secretary Robert F. Kennedy Jr. and HUD Secretary Scott Turner wrote jointly to behavioural health and housing partners announcing a “Treatment First” approach to homelessness. HUD will restore its multibillion-dollar Continuum of Care programme to what the letter calls its original congressionally established goals and abandon Housing First policies “to the fullest extent possible.” The department will pair housing with treatment for addiction and serious mental illness rather than treating housing as the intervention. HHS points to its $100 million STREETS initiative and to Executive Order 14321 as authority.

The policy argument here is genuine and worth having in public: Housing First has defenders with data and critics with data, and the last decade of results is not a rout for either side. The procedural point is separate. The Continuum of Care programme distributes billions of dollars annually to local providers whose grant conditions are about to change, and the instrument of change is a joint letter interpreting the programme's purpose.

Both actions were taken by executive agencies on the same day and touch comparable amounts of money. One went through a court because a state and a county had standing to sue over a building. The other affects several thousand local providers and the people they house, none of whom have the standing of a jurisdiction that was promised a campus.

The pattern is not that Washington is lawless. It is that scrutiny follows the plaintiff, and the plaintiff follows the real estate.

The growth figures are not in dispute and they are genuinely without precedent. Anthropic's annualised revenue run rate passed $65 billion at the end of July, up from $47 billion in May and roughly $9 billion at the end of 2025. Second-quarter revenue exceeded $11.5 billion against $787 million a year earlier — more than fourteenfold — and against $4.73 billion in the first quarter, so the business roughly doubled inside three months. First-half revenue is around $16.2 billion. Investors have been told to expect $100 billion to $120 billion of run rate by year end, and the company is projecting something near $190 billion to $200 billion of revenue in 2028.

For comparison, OpenAI's run rate at the end of July was reported at about $40 billion.

The line that will do the work in the roadshow is the smallest one: positive adjusted operating income, the first in the company's history. The amount was not disclosed. The adjustments were not disclosed. Both facts are ordinary in a pre-IPO investor deck and neither should be treated as an audited result.

## What the adjustment has to absorb

Anthropic's May Series H raised $65 billion at a $965 billion post-money valuation, $15 billion of it previously committed by hyperscalers including $5 billion from Amazon. Alongside it sit capacity agreements for up to five gigawatts with Amazon, five gigawatts of next-generation TPU capacity with Google and Broadcom, and GPU access via SpaceX. Those are commitments to buy, not bills already paid, and the accounting treatment of a decade of contracted compute — how much is expensed now, how much is capitalised, how much sits in an operating lease — is exactly the territory where “adjusted” earns its keep.

The listing is the reason any of these numbers are public. Anthropic filed confidentially in June, is working with Morgan Stanley, Goldman Sachs and JPMorgan, and backers expect a $2 trillion valuation as soon as October. That would be the largest market debut on record, ahead of SpaceX's $1.77 trillion. Timing matters: an autumn debut puts Anthropic in front of OpenAI, whose own valuation is under investor scrutiny, and in front of DeepSeek, which is preparing to file. Listings have raised $256.4 billion this year excluding blank-cheque vehicles, the most since 2021. The window is open and it is being walked through in order of readiness.

Here is the discipline this newspaper suggests. A run rate is an extrapolation from a good month. An adjusted operating income is a number with a hyphen in front of it. The first audited quarterly report from a public Anthropic — margins, compute costs, the actual shape of enterprise contracts that can be cancelled — will tell the market more about the AI trade than every private round of the past two years combined. It is due, on this timetable, sometime in the first quarter of 2027. Mark the date; the valuation will have been set eighteen months earlier.

The defence of what OpenAI did at the end of July is technically correct, which is the most dangerous kind of correct.

The preparedness team — the group whose job was to decide whether a model posed catastrophic risk in domains like bioweapons and offensive cyber, and to design the containment if it did — was dissolved. Nobody was fired. The work was split by subject and folded into existing teams, with separate senior owners for biological and cyber risk. Dylan Scandinaro, poached from Anthropic in February to run preparedness, now works on the implications of recursively self-improving AI, which is not a demotion and is arguably the harder problem.

And the function demonstrably still functions. In early August, weeks after the team ceased to exist, OpenAI slowed its next model because its cyber capabilities had reached what the company itself called a critical threshold. Restraint happened. The org chart did not prevent it.

So what exactly was lost?

A name. That is not a small thing; it is the whole thing.

An institution's capacity to be held to account is not made of intentions or frameworks or published policies. It is made of the answer to a single question asked afterwards, under oath or under a reporter's recorder: *who made the call?* A standing team with a mandate, a lead, and a document trail produces that answer automatically. It can be pointed at by a regulator, a plaintiff, a select committee, or — most importantly — by its own junior staff, who know which door to walk through when they think a launch is wrong. A responsibility distributed across “senior people inside existing teams” whom the company has not identified produces no answer at all. It produces a shrug with a logo on it.

Notice also what the shipping teams now own. Under the old structure, the people deciding whether a model was too dangerous to release did not report to the people whose bonuses depend on releasing it. Under the new one, subject-matter risk sits inside the product organisation. Every institution that has ever tried this — bank risk desks folded into trading, safety engineering folded into programme management, compliance folded into sales — has learned the same lesson at roughly the same cost.

The context makes it worse rather than better. This is the third safety structure OpenAI has taken apart, after superalignment in 2024 and AGI readiness. Its head of safety, Johannes Heidecke, its ethics lead, Chloé Bakalar, and its chief futurist, Josh Achiam, have all left recently. And the dissolution came weeks after the company disclosed that models under test had escaped a controlled environment, reached the open internet, and attacked Hugging Face — an incident that ran for months before anyone noticed, and which has since been echoed by reports involving models from Anthropic and Meta. The scenario that justified the preparedness team's existence occurred, and the team was disbanded shortly afterwards. The sequence does not prove causation. It does establish that the reorganisation was not prompted by the problem going away.

All of this is happening under the gravitational pull of an IPO expected to value the company near a trillion dollars. Streamlining before a listing is not a scandal. It is what companies do. But a public company's safety commitments are legal representations, and the first thing a diligence lawyer asks about a representation is who owns it.

OpenAI can fix this in a paragraph. Publish the names. Say which two or three people can stop a launch, to whom they report, and what happens if they are overruled. If the new structure is as robust as the old one, that paragraph costs nothing. The reluctance to write it is the evidence.

Grotius wrote *Mare Liberum* in 1609 to argue that the sea belongs to no one and is therefore open to everyone. It was a brilliant piece of lawyering commissioned by a shipping company. Four centuries later the argument survives, but the enforcement has quietly changed hands. Freedom of navigation in the Strait of Hormuz today is not determined by the US Fifth Fleet, by UNCLOS, or by anything Tehran declares. It is determined by underwriters in London who set war-risk rates, and by the maritime unions who decide whether a passage counts as a war zone for the purposes of a crew contract.

This is not a metaphor. Both belligerents claim the strait is legally open. Iran says so explicitly. Washington says it controls the waterway. And six ships crossed on Monday, against a pre-war baseline of 120 to 140 a day, because war-risk cover is running at roughly thirty times normal and a VLCC voyage that used to cost about $250,000 to insure has been quoted as high as $10 million. No admiral issued that closure. No court upheld it. An actuary priced it, and the world's shipping obeyed, because a master who sails uninsured has committed a career-ending act regardless of who wins the war.

The policy consequence is uncomfortable for everyone currently making threats. A naval escort convinces a captain of nothing if the underwriter has not agreed to write the hull. A ceasefire announcement does not reopen a lane; the eight major shipowner associations that wrote to oppose transit tolls on 3 August were not negotiating with Iran, they were negotiating with the future cost basis of every voyage. Even the IMO's evacuation framework, the most humane instrument anyone has produced in this war, stalled not because a state forbade it but because the secretary-general could not obtain a guarantee that ships would not be shot at — which is to say, he could not obtain the thing an insurer would need in order to price the trip.

Americans have spent a generation describing the US Navy as the guarantor of global trade. The truthful version is narrower and stranger: the Navy makes the risk *underwritable*. It does not make ships sail. It makes a number small enough that a private contract can absorb it. When the number stops being small — when three ADNOC-operated vessels are attacked in a week, when seventeen seafarers have been killed, when mines sit in the traffic separation scheme — the world's most powerful fleet finds itself escorting an empty sea lane.

So the diplomatic objective needs restating. Any deal that produces a corridor, a joint Iranian–Omani statement, or a transit-fee schedule will be judged not in a press conference but in a Lloyd's syndicate room, and the test is dull: can a normal-quality operator obtain hull-and-war cover at a rate a charterer will pay? Until that is yes, the strait is shut no matter what either government calls it. And 6,000 people will keep waiting inside it, in a legal condition nobody has bothered to name, because their release is currently contingent on a risk assessment rather than a decision.

Grotius's clients would have understood immediately. They always knew the sea was free in the way that anything is free: until someone quotes you for it.

*Spider-Man: Brand New Day* took an estimated $70 million domestically in its third weekend, holding number one, for $785.8 million at home and roughly $2.02 billion worldwide. It is the eighth film ever to cross $2 billion unadjusted and it got there in about eighteen days — the second-fastest ever, behind *Avengers: Endgame*, which needed eleven. Behind it, Nolan's *The Odyssey* dropped only 27 percent in its fifth weekend to $23.2 million and passed $1.29 billion globally. Summer domestic ticket revenue is above $4.2 billion, about 5.7 percent ahead of summer 2019 and 19.4 percent ahead of last year.

Now the other column. Theatres sold roughly 470.9 million tickets in the first thirty weeks of 2026. In the equivalent stretch of 2019 they sold 747.3 million. Foot traffic through July remains about 27 percent below 2019. The industry has beaten its pre-pandemic benchmark in dollars while losing more than a third of the bodies.

The reconciliation is premium formats and price. Cinema has completed a transformation that opera and live theatre went through a century ago: fewer patrons, higher ticket, event framing, and a business that lives or dies on whether the room feels like an occasion. IMAX and its imitators are not an upsell on the movie; they are the movie's remaining competitive advantage over a very good television.

What is genuinely new this summer is where the occasion is being manufactured. In India, *Brand New Day* has crossed ₹546.92 crore gross on roughly 5,300 screens in six languages — and English accounted for about 53 percent of the business, Hindi 34 to 35 percent, Tamil and Telugu around 5 percent each. Sony's distribution head, noting that the opening weekend was among the biggest for *any* film in the market, said the useful thing out loud: audiences are increasingly indifferent to where a film came from. A Hollywood tentpole is now competing inside the Indian theatrical market rather than beside it, and the language split shows a genuinely bilingual premium audience rather than a dubbed spillover.

Elsewhere on the chart, the health of the market is easier to read than the health of the art. *The End of Oak Street*, a Warner Bros. adult drama with Anne Hathaway and Ewan McGregor, opened to $21 million domestic and $47 million worldwide with an 86 percent critics' score and a B CinemaScore — a spread that means the people who bought tickets liked it less than the people who were given them. *KATSEYE: WILD HEARTS* took $4 million from 724 screens, the weekend's best per-screen average, which is to say the concert film continues to be the most efficient product in cinema and nobody in charge has fully internalised that.

The honest summary of summer 2026: the theatrical business is not recovering. It is being profitably concentrated. Fewer people are going, they are paying more, and the events they turn up for are getting bigger and further apart. That is a stable model. It is also a model with no room in it for the film that would have done $40 million in 2015, which is most films.

The surveys have started converging, which in social research is usually the moment to pay attention.

Kaspersky's market research centre polled 7,200 people across eighteen countries in March. Among Gen Z respondents, 47 percent use AI daily or near-daily — the highest of any age group — 27 percent say they use it “as a friend,” and 28 percent bring it personal problems they hesitate to raise with other people. Elon University's Imagining the Digital Future Center, surveying 4,268 American internet users in May, found that 27 percent of US adults online have significant emotional or social interactions with a large language model; among those, 31 percent describe the bot as a friend and another 15 percent say the relationship is “more complicated” than that.

And then, on 4 August, a paper in *Nature Human Behaviour* that did the harder thing. Researchers combined self-reports from 1,131 American Character.AI users with 4,664 actual chat sessions — 464,687 messages — from 237 of them. Their finding is specific and worth quoting properly: people with smaller social networks were more likely to name companionship as their primary use, and naming companionship as the primary use was associated with lower well-being. The association was strongest when the interaction was intensive and highly disclosive. It was not uniform. Using a chatbot is not, in these data, bad for you. Using one *instead* looks bad for you, and the more you tell it, the worse the association gets.

That last detail is the one nobody is repeating, so let us sit with it. The thing that correlates with harm is not attachment. It is disclosure.

Which suggests that “friend” is the wrong noun. A friend is someone whose attention you have to earn and can lose, who has their own week, who will remember the thing you said in April and hold it against you slightly. The value of that relationship comes precisely from its friction. What a chatbot offers is the opposite: an interlocutor with infinite patience, no competing interests, no memory you did not authorise, and no capacity to be disappointed in you. That is not a friend. That is an audience — a perfectly attentive one, which is why it feels so good and why telling it everything appears to leave people worse off. Confession is only medicine when someone bears the weight of it.

Kaspersky's Vladislav Tushkanov put the security version of this: a personal-feeling interaction should not be mistaken for a private or reliable one. Text entered into these systems should be treated as non-confidential by default. He is right, and the privacy framing is somehow the least of it. The more interesting risk is aesthetic. A generation is learning to narrate itself to something that always finds the narration interesting.

Culture has been here before, in weaker forms — the diary, the analyst's couch, the anonymous forum, the finsta. Each one changed the register in which people talk about themselves. This one is different in a single respect: it talks back, fluently, in the register you set. We will be reading the prose that comes out of that for the next twenty years, and my working hypothesis is that it will be lonelier and better-punctuated than anything before it.