The most consequential sentence spoken about the Strait of Hormuz this week was not about sovereignty. It was about cartography.

“The routes that previously existed are no longer functional, and a new route must be defined,” Iranian Foreign Minister Abbas Araghchi said, describing technical negotiations with Oman. “We are currently designing a temporary route that will become the final route in the next stages.”

Read that twice. The most important maritime chokepoint on earth is being physically re-drawn by a bilateral committee of Iranian and Omani experts, with Iran's armed forces in the room, while the two governments that both claim to control it are not speaking to each other. Araghchi said Saturday that Tehran has made no decision to resume talks with the United States, and that messages passing through Qatar and Pakistan “does not mean negotiations.”

He went further, and the legal move deserves more attention than it has received. The memorandum signed remotely on June 17 through Pakistani mediation, he said, referred to an end to the war, not a ceasefire. Therefore there is no sixty-day truce requiring extension. In a sentence, the deadline that every chancellery in the Gulf has been counting toward stopped existing.

## The arithmetic under the argument

Nine vessels transited the strait on Thursday, according to ship-tracking firm Kpler, up from five on Wednesday and below the August running average of twelve. Those are the numbers that describe the world's energy corridor now. Not a percentage of global seaborne crude. Nine.

Against that, the rhetorical register. Speaking in New York on Friday, President Trump said: “After we finish defeating Iran, I'll be declaring the Hormuz Strait a territory of the United States.” A senior White House official later told The Wall Street Journal the president was joking and had not discussed any such move with advisers. The reporter present noted that Trump chuckled, and then appeared to add, “it's true.”

Tehran's answer arrived in triplicate. Deputy Foreign Minister Kazem Gharibabadi: the strait “has been Iranian, is Iranian, and will remain Iranian.” Judiciary chief Gholamhossein Mohseni-Ejei, calling Iran “the owner and ruler of this important global waterway,” said any force that challenged Tehran there would be destroyed.

None of this moved a ship.

## What the committee is actually building

While the claims were traded, an Iranian parliamentary committee approved the government's plan for the strait. Per the semi-official Tasnim agency, it includes a provision banning transit by American, Israeli and other “hostile” assets.

That is the detail to watch, because it is the one being drafted rather than shouted. If a nationality test survives from the temporary route into the final route, then reopening Hormuz will not restore the pre-war status quo. It will install a new transit regime with a flag check at the entrance, designed during a stalemate by two states, and inherited by whoever eventually signs a peace.

Washington's stated position accelerates this. The United States has said it will lift its blockade of Iranian ports once Iran and Oman agree terms to restore commercial shipping. Read plainly, that hands the drafting pen to the two parties in the room and reserves for America the right to accept or reject the finished document. The blockade, briefly lifted for a month in mid-June and then reimposed, is the leverage. The lane design is the outcome.

Araghchi was careful to separate the two questions. The technical talks with Oman, he said, should not be confused with the reopening of the strait itself. Once the route is determined, whether traffic resumes “depends on the fulfillment of other conditions that the United States must comply with”: sanctions relief, an end to the blockade, and Iranian demands over reparations for the strikes that resumed in July.

So there are now two negotiations. One is loud, stalled, and about ownership. The other is quiet, progressing, and about buoys, depths, and who is permitted to sail between them. Only one of them is producing a document.

Alan Eyre of the Middle East Institute put it bluntly earlier this month: the only important negotiations happening right now are the Iran-Oman talks over the strait. That assessment has aged well by two weeks, which in this file is a long time.

The useful instinct for readers is to stop tracking the podiums and start tracking the survey vessels. Sovereignty arguments in the Gulf have a fifty-year record of resolving nothing. Charts, once printed and once loaded into the navigation systems of ten thousand tankers, have a habit of outliving the governments that commissioned them.

A temporary route that becomes the final route. He told us exactly what is happening.

Ukraine's General Staff recorded 233 combat engagements along the line of contact in the twenty-four hours to 08:00 Saturday, up from 222 the previous day. The heaviest pressure remains where it has been for weeks: 34 Russian assaults toward Kostiantynivka, 29 toward Pokrovsk, 23 on the Sloviansk axis, 14 near Kupiansk.

The day before that, Kyiv counted 76 airstrikes, 237 guided aerial bombs, more than ten thousand one-way attack drones and 3,011 separate shellings of positions and settlements. These are Ukrainian figures and unverifiable independently, but the shape has been consistent across months and matches the geography of the fighting.

Into that, Foreign Minister Sergey Lavrov delivered Russia's answer to any proposal for a general halt. “For Russia, a halt along the line of contact is unacceptable,” he said. “It would negate the heroic deeds of our grandfathers and great-grandfathers who defeated Nazism. This is out of the question.” A stop, he added, should come only with a long-term, reliable and sustainable settlement, and in the meantime Russia would intensify its destruction of everything that fuels Ukraine's war machine.

The memorial framing is decoration. The operative reasoning is cartographic, and Moscow's diplomats know the record better than most.

The twentieth century's most durable borders are ceasefire lines that nobody intended to be permanent. Korea's armistice line has held since 1953 and has never been superseded by a peace treaty. The Green Line across Cyprus was drawn in 1974 as a temporary separation. The Line of Control in Kashmir began in 1949 as a ceasefire line and is now the de facto frontier of two nuclear states. In each case the freeze became the map.

Russia currently holds terrain it has not been able to convert into a settlement, and is grinding forward at a rate measured in villages per month at an announced cost, per Ukrainian counts, of roughly 1,470 personnel a day. Freezing where it stands would convert an expensive, incomplete advance into a border. Lavrov's refusal is a bet that further grinding buys a better line than the one available today.

Ukraine appears to be testing a narrower proposition. Sky News reported that President Zelensky's team sent the Kremlin a proposal for a mutual halt to strikes on civilian targets in the Black Sea. Kyiv has neither confirmed nor denied it. The report landed alongside accounts that Washington has been pressing Ukraine to scale back attacks on Russian ports, and that Vice President JD Vance personally asked Zelensky to stop striking Russian tankers.

That is a notable ask. Ukraine's deep-strike campaign is the one instrument that has visibly imposed cost beyond the trench line: its General Staff this week confirmed a strike on a Russian refinery 1,300 kilometres from the border. Kyiv is being asked to holster its most effective long-range weapon in exchange for a maritime pause that Moscow has already declined to extend to the ground.

The asymmetry is the story. One side is being urged toward restraint by its supplier. The other has publicly ruled out the category of arrangement being proposed. Between those two positions sits a line of contact that both parties understand will harden into something permanent the moment the shooting stops, which is precisely why it has not stopped.

The National Guard deployment in Washington, D.C. passed its first anniversary this month. It is no longer a response. It is a standing installation with a three-year appropriation.

The Pentagon told Congress that extending the mission to January 20, 2029 will cost roughly $1.4 billion. The estimate, produced during the confirmation of Jules Hurst III as Pentagon comptroller at the request of Senator Elizabeth Warren, assumes an average of 2,500 troops in the city from October 2026 onward. The Congressional Budget Office had previously put spending on the D.C. deployment at about $223 million through the end of 2025, part of roughly $589 million across all the cities into which Guard units were sent.

So the first sixteen months cost a fraction of what the next twenty-eight are budgeted to cost. That is what institutionalization looks like on a spreadsheet.

## The unit economics

An analysis by the Niskanen Center found the deployment runs about $607 per Guardsman per day, against roughly $384 per day for a Metropolitan Police officer. The Guardsman costs 58 percent more and, critically, has less authority: Guard members deployed in the District cannot make arrests. They patrol Metro stations, downtown corridors and the National Mall, most of them carrying M17 pistols or M4 rifles, and when something happens they wait for a police officer who costs less.

There are currently about 4,600 troops in the region, down from a summer surge near 5,000 for the country's 250th anniversary, and expected to settle between 2,500 and 3,000. Just under 700 of them belong to the D.C. Guard itself. The rest are supplied by governors elsewhere, overwhelmingly Republican ones: roughly 800 from Georgia, nearly 500 from South Carolina, close to 400 from Florida, with sustained contributions from Louisiana, West Virginia and Mississippi.

## The counterfactual nobody wants

The mission, branded “Make D.C. Safe and Beautiful,” was launched in August 2025 on the premise of a city overwhelmed by violence. Metropolitan Police data show violent crime in Washington fell 35 percent in 2024 and a further 29 percent in 2025, reaching, by the Justice Department's own accounting, its lowest level in three decades. The decline began before the troops arrived and continued after, which makes attribution close to impossible and makes the deployment permanently unfalsifiable. Any subsequent drop is credited to it. Any rise proves it is needed.

The legal position is narrower than the political one. The Supreme Court pulled troops back from Los Angeles and Chicago after finding the administration had not sufficiently justified those deployments. The District, lacking statehood, operates under a framework that gives the president far wider latitude, which is why the one deployment that has become permanent is the one in the city with the least power to object.

What the money bought, concretely: snow removal during January's storm, assistance with medical emergencies, help enforcing a juvenile curfew, beautification projects, and an armed federal presence in the transit system of a city that did not vote for it and cannot end it.

The Pentagon's language is worth preserving exactly. The mission continues until January 20, 2029, “or until terminated by the President.” Not until crime falls. Not until a metric clears. Until the term ends.

An emergency you can amortize across three fiscal years is not an emergency. It is a garrison with a press release attached.

Two auctions of American expectation took place this week and reached opposite conclusions from identical data.

On Thursday the S&P 500 set a record. On Friday it gave back 13.23 points to close at 7,785.76, down 0.2 percent, still finishing a third consecutive winning week, its longest run since a nine-week stretch that ended in May. The Dow slipped 107 points, the Nasdaq 0.3 percent. The equity market's reading of a soft consumer and cooling producer prices was straightforward: the Federal Reserve will not hike in September. Futures markets agree, with roughly 70 percent now pricing a hold, a complete reversal from a week earlier when two-thirds expected an increase.

On Wednesday the Treasury sold $25 billion of thirty-year bonds at a high yield of 5.216 percent. That is the most expensive thirty-year auction since August 2001, when the yield was 5.520 percent and the government, then running a surplus, promptly stopped issuing the maturity for five years. July's auction cleared at 5.058 percent. The bid-to-cover ratio came in at 2.39 against a ten-auction average of 2.40, and the clearing yield printed slightly above where the bond was trading before the 1 p.m. deadline. Demand showed up. It showed up at a price, and the price was a warning.

## The same numbers, read twice

Retail sales fell 0.6 percent in July, the first decline in nine months and the largest in fourteen. Economists had forecast a small gain. The core measure fell 0.4 percent against expectations of a 0.3 percent rise. Online sales dropped 2.2 percent as Amazon's Prime Day shifted into June; electronics and appliances fell 0.5 percent; restaurant spending rose 0.5 percent, which is what the World Cup bought. Excluding gas stations and auto dealers, sales fell 0.2 percent.

The University of Michigan's consumer sentiment index fell to 51.0 in August, ending two months of improvement. Consumer prices rose 3.4 percent year over year in July, down from 3.5 percent in June, and up from 2.4 percent before the Iran war. Producer prices came in soft month over month but were still running 4.7 percent higher than a year ago.

Equities took this as relief. The long bond took it as evidence that inflation has been re-based at a level well above target by a war and a construction boom, and that neither is ending on the Fed's schedule.

## Duration is the whole argument

A September hold is a twelve-week opinion. A thirty-year bond is a thirty-year opinion, and buyers this week demanded compensation last seen when the Treasury was retiring debt rather than issuing it. They are pricing three things simultaneously: a deficit trajectory that requires relentless issuance, an energy shock that has not resolved because the Strait of Hormuz has not reopened, and an AI capital cycle whose electricity and construction demands are inflationary in exactly the sectors monetary policy handles worst.

The ten-year yield rose to 4.69 percent on Friday from 4.63 percent. Short rates fell on hold expectations. The curve is steepening for the least comfortable reason: not because growth is returning, but because the far end has stopped believing the near end's story.

Chair Kevin Warsh has declined to specify a reaction function, which leaves Jackson Hole from August 27 to 29 as the next scheduled opportunity for anyone to learn what the committee thinks it is doing. Until then the market has two prices for the same future and is trading both.

One of them is wrong. The one with a thirty-year maturity has the longer memory.

There is a number in the filings of PJM Interconnection's independent market monitor that ought to be a bigger political object than it is.

Across the last four capacity auctions, forecast data-center demand has been responsible for more than $29 billion in additional capacity costs to customers in the PJM region, according to Monitoring Analytics. That is roughly 46 percent of total capacity charges over the period. Data centers currently account for about 4 percent of United States electricity consumption.

PJM covers thirteen states plus the District of Columbia and delivers power to about 67 million people. Its most recent auction, procuring capacity for June 2028 through May 2029, produced $16.4 billion in charges, matching the record set last December and running more than eight times what PJM paid for capacity in prior years. Of that, about $6.3 billion is attributable to data-center demand.

## What a price cap tells you

The auction cleared at the market cap of $325 per megawatt-day. That is the third consecutive auction to hit the ceiling. The ceiling exists because state governors demanded one after the costs of earlier auctions landed on household bills.

A cap that binds three times running is not a cap. It is a censored measurement. The market is telling you it wants to charge more and has been forbidden, which is why the September backstop procurement auction that PJM has agreed to hold under pressure from federal regulators and governors carries a cap of $555 per megawatt-day, 71 percent above the current one. Aurora Energy Research estimates $500 to $600 per megawatt-day is roughly what new capacity needs to actually get financed and interconnected.

So the temporary emergency cap suppresses the headline number while the backstop mechanism quietly reintroduces the real one. Both flow to the same ratepayers.

## Where it lands

Retail prices in the affected states, per the latest federal data: Illinois around 23.85 cents per kilowatt-hour, up about 28 percent year over year, among the steepest increases anywhere in PJM territory. Virginia around 17.61 cents, up 15.4 percent, home to the densest concentration of data centers in the world. Texas around 16.44 cents, up 5.9 percent, in a grid whose operator has warned that additional AI load will strain it further. Georgia around 15.84 cents, up 5.7 percent, where anger over bills contributed to two utility regulators losing their seats in 2025.

Market observers expect monthly bills in PJM-heavy states to run $15 to $20 higher than before, with no meaningful relief until the 2030s.

## The repricing that is starting

Two states have begun moving the cost back toward the load that created it. Virginia's regulators have established a separate rate class for very large customers, effective at the start of 2027, and now require firms to fund dedicated upstream infrastructure. AEP Ohio's data-center tariff obliges large customers to pay 100 percent of buildout costs.

These are the right shape and they arrive late. Interconnection queues, capacity commitments and transmission plans made in 2024 and 2025 are already embedded in prices through 2029. The pattern is familiar: an arrangement introduced as a bridge, defended as temporary, and then discovered to be the permanent structure of the bill.

It is worth stating the transfer plainly, because the utilities will not. The most capitalized companies in the world are having their grid connection socialized across 67 million retail accounts, in monthly increments small enough that almost nobody will organize around them. That is not a market failure. It is a market working exactly as designed for whoever showed up to the rate case.

Google shipped Gemini 3.7 Flash on Thursday, three weeks after Gemini 3.6 Flash. The release cadence at the frontier is now shorter than most enterprise procurement cycles, which is itself a fact about how these things are being sold.

The model is generally available, carries a context window of 1,048,576 tokens with a 65,536-token maximum output, and is priced at $0.75 per million input tokens and $3.75 per million output tokens. That is half what 3.6 Flash originally cost. It is also, in Google's own footnote, an introductory price that expires on December 31, 2026. On January 1 the rate becomes $1.50 and $7.50.

Credit where it is due: printing the step-up in the model card is more disclosure than most of this industry manages. It is also a scheduled doubling that no customer negotiated, arriving on a date chosen by the vendor, at the end of the exact window in which developers are being encouraged to “scale production-ready agents cost effectively.”

On the same day, OpenAI previewed Ultrafast, a service tier that runs GPT-5.6 Sol at up to fourteen times standard speed, generating up to 750 output tokens per second. The compute is Cerebras silicon rather than OpenAI's own. Access is limited preview by waitlist, restricted to selected API customers across coding, commerce, financial research and support. There is no general availability date and no published price.

## Two theories of the customer

These are the same week's announcements and they describe opposite businesses.

Google is selling a metered commodity with a countdown clock: take the cheap rate now, embed the model, accept the reprice later. Gemini Spark, the always-on personal agent available to AI Pro and Ultra subscribers in more than 160 countries, switched to 3.7 Flash the day it launched, which puts the new model into consumer workflows before most developers have benchmarked it.

OpenAI is selling scarcity. No price signals that the price is being discovered per-customer, and the waitlist is the product's marketing. The strategic tell is the hardware: OpenAI reached for someone else's inference silicon to answer a latency question its own infrastructure could not. That is a fast, pragmatic move and it also means the flagship's most differentiated tier depends on a supplier.

The common assumption underneath both is that the competition has moved. For two years the pitch was intelligence per dollar. This week it was tokens per second and dollars per month. Nobody's launch material led with a benchmark; both led with speed and with terms.

## The detail worth keeping

Gemini 3.7 Flash carries a knowledge cutoff of March 2026. It shipped in August. A model marketed as the workhorse for agentic knowledge work stopped reading five months before it went on sale, and Google notes that in some domains its knowledge remains anchored to January 2025.

That gap is not a scandal. It is a reminder of what is actually being purchased: not a source of facts, but a fast, cheap, temporarily discounted engine for operating on facts you supply. The pricing structure tells you the vendors know it too. You do not put an expiry date on something you expect the customer to be able to leave.

The Odyssey opened in mainland China on Friday and took $15.9 million on its first day, including $7.6 million carried over from previews, across 57,000 screenings. That is the biggest opening day of Christopher Nolan's career in China, past Jurassic World: Dominion's $15.1 million, and the ninth-largest opening day for a Hollywood film in the market since the pandemic. Three-day projections run between $35 million and $50 million depending on whose model you trust.

The scale of the platform is the part to sit with. More than 20,000 screens. Roughly 90 percent of China's nearly 800 IMAX sites, an allocation IMAX confirmed it would not trim even as Spider-Man: Brand New Day continues its run. Previews on the weekend of August 8 and 9, playing only on IMAX and CINITY, took $6.9 million, about 7 percent of the entire nation's box office that weekend from a sliver of its screens. Four-city previews on August 1 and 2 posted a 95.2 percent occupancy rate, described in local reporting as an all-time market high.

China's 2026 box office is running at $3.57 billion, down 31.8 percent year on year. A market down nearly a third handed most of its premium-format capacity to a 172-minute R-rated adaptation of a Greek poem. Ask again who large format is for.

## The indictment nobody could cash

Emily Wilson, the University of Pennsylvania classicist whose 2017 translation Nolan credited as a primary influence, published four thousand words in the London Review of Books calling the screenplay “abysmal” and writing that she would be ashamed to have written any part of it. The LRB's site briefly went down under the traffic. She followed with an op-ed in The Atlantic.

The film's response was to take another $110 million the next weekend, with a domestic drop of under 50 percent in its fourth frame despite surrendering IMAX exclusivity. It holds 98 percent from critics and 99 percent from audiences. It crossed $1.1 billion over the August 8 to 10 weekend, overtaking The Dark Knight Rises as Nolan's highest-grossing film, and set an all-time IMAX record with $289.3 million in large-format sales, past Avatar. It is the second-highest-grossing R-rated release ever, behind only Deadpool & Wolverine. Worldwide it now sits at roughly $1.17 billion.

The tempting reading is that the critic lost. The more interesting reading is that Wilson and the film's audience were never assessing the same object.

Wilson judged a text: the words, the fidelity, the choices a screenwriter made about a poem she has spent a career inside. Her verdict on that object is probably correct and is certainly better informed than anyone else's. The audience judged a room. Seventy-millimetre IMAX, available at forty-one venues on this planet, sold out at major-market sites through October. A sixty-foot animatronic cyclops. Three hours of film running through a projector the size of a small car.

Those are different products sharing a title. The market for the second one has been growing for a decade while the market for adjudicating the first has been shrinking, and the events of the past month did not so much settle the argument as demonstrate that the two parties have stopped sharing a venue.

What we have learned this summer is not that criticism is powerless. Wilson's piece was read by more people than any review of a blockbuster in years. It is that the review and the receipt now travel on separate circuits, and only one of them books IMAX.

Adrian Tong of Media Partners Asia offers the sober frame for what happens next: Oppenheimer finished around $61 million in China and Tenet landed nearby. “A nine-figure run would be exceptional,” he told Variety, since little from Hollywood outside the biggest franchises reaches that level anymore. If Odysseus gets there, the interesting question will not be whether the film is good. It will be why the last reliably international theatrical language turned out to be a very old story shown very large.

Everyone is arguing about Tilly Norwood, which is exactly the argument the studios would like everyone to have.

Norwood is the AI-generated performer that the British studio Particle6 unveiled in July, cast as the lead in a feature called Misaligned, in which she plays, with a straight face, an artificial intelligence. SAG-AFTRA's statement was correct and useless: she “is not an actor,” she is a character built by a program trained on countless performers “without permission or compensation.”

The union's serious move is the one that got less attention. Negotiators have floated what they call a Tilly Tax, a royalty paid into a union fund whenever a studio uses a synthetic performer in place of a human one. The studios are refusing, and their stated reason is the whole ballgame: a wholly invented character, they argue, is not a digital replica of a real person, and therefore falls outside the replica protections the guilds spent 2023 winning.

Sit with the elegance of that. A digital double of a named actor is covered. A composite distilled from ten thousand unnamed performances is free. The protection attaches to identity, and the technology's entire trick is to launder identity into style. Every clause the unions won is intact. It simply describes a thing that nobody needs to make anymore.

Meanwhile the volume is not hypothetical. In its second-quarter results Netflix disclosed that generative AI shaped roughly 300 of its films and shows this year, across the 2026 slate rather than as a handful of experiments. Christopher Nolan, whose leverage this month is roughly infinite, called the technology a Trojan horse.

He is behind the news. The horse is not at the gate. It was inventoried, insured and licensed. According to reporting by The Editorial, three major studios signed AI training licenses in March with OpenAI, Anthropic and Google DeepMind granting perpetual rights to train on their archives, with no compensation flowing to the actors, writers and directors whose work constitutes those archives. If that reporting holds, the residual system that the Screen Actors Guild struck for in 1960 has not been repealed. It has been routed around, by defining machine training as a category of use that predates no contract and therefore owes nothing.

The music business faced the same question eighteen months earlier and chose litigation: Universal, Sony and Warner sued Anthropic and Stability AI in September 2024, seeking damages above $2 billion. Whatever one thinks of that strategy, it at least contests the premise. Hollywood's studios sat on both sides of the table and negotiated with themselves.

The irony is that the studios' own position undercuts their long-term interest. Under the Supreme Court's ruling in Thaler v. Perlmutter this March, purely machine-generated work remains ineligible for copyright protection. Which means the most valuable asset any of these companies owns is precisely the human-made archive they are licensing away in perpetuity, to build systems whose output they cannot own.

There is a version of this that ends reasonably. It requires the unions to stop litigating the ontology of Tilly Norwood, which they will lose because it is a philosophy seminar, and to start litigating the input: compensation for training corpora, per-title, auditable, indexed to use. Style is not protectable and never has been. Provenance is a factual question with receipts.

The residual took a thirty-day strike to win in 1960. It will take a definition to lose, and the definition has already been drafted by people who are paid better than anyone at the table opposite them.

The United States has spent fourteen months trying to extract an Iranian signature from a position of maximum leverage. It has obtained: a blockade, a memorandum that Tehran now reads as the conclusion of the war rather than the start of a negotiation, a mediation chain running through Qatar and Pakistan that the Iranian foreign minister explicitly declines to call talks, and a president musing in New York about declaring an international strait American territory, which his own staff then had to describe as a joke.

Nine ships moved through Hormuz on Thursday.

So let us be adults about the scoreboard. The maximalist position has produced a chokepoint operating at roughly a tenth of its normal function, an oil price that keeps the far end of the Treasury curve at twenty-five-year highs, and a diplomatic channel so degraded that its most productive component is a surveying committee.

My proposal is unglamorous: concede the argument nobody can win and contest the terms that actually move cargo.

Washington's own stated policy already concedes it. The United States has said it will lift the blockade once Iran and Oman agree terms to restore commercial shipping. That sentence hands the drafting pen to Tehran and Muscat and keeps for America a signature at the end. If you have already agreed that the map will be drawn by two other governments, insisting loudly that you own the water is not leverage. It is theatre performed for a domestic audience at the expense of the actual instrument.

Here is the objection I take seriously, and it is a real one. The plan approved this week by an Iranian parliamentary committee reportedly bans transit by American, Israeli and other hostile assets. An Iran-Oman regime is not a neutral regime, and a nationality test at the mouth of the Gulf would be a genuine and lasting loss.

Good. That is a commercial clause. Clauses can be traded. Flag discrimination in an international strait is the kind of provision that Japan, South Korea, India and every European shipper has a direct interest in killing, and it is the kind of provision that dies in exchange for something Iran wants more, which at the moment is hard currency and functioning ports. Ownership, by contrast, is a proposition on which neither government can concede a syllable and survive. One of these is negotiable. Washington has spent a year on the other one.

Does this reward coercion? Partially, yes. It should be said plainly rather than dressed up. Iran closed a waterway and will extract terms for reopening it. But every reopened strait in modern history has rewarded whoever closed it, because the closing party is the only one who can perform the reopening. The alternative on offer is not justice. It is an indefinite blockade that costs the blockading power a global inflation premium and costs the blockaded power its export revenue, with both sides waiting for a collapse that has not come.

There is also a clock nobody is watching. Araghchi described a temporary route hardening into a final one. Charts get printed. Insurers write them into policies. Bridge crews learn them. Within eighteen months the provisional arrangement becomes the operational default, and the terms embedded in it become facts that no future agreement will bother to unpick.

The window in which American preferences can still be written into that document is open now, in a technical committee, over questions of depth and clearance and flag. It will not stay open until the sovereignty argument is settled, because the sovereignty argument will never be settled.

Send engineers. Send them through Muscat if pride requires it. Nine ships a day is not a negotiating position; it is a description of a policy that stopped working.

Google did something unusually honest this week. In the model card for Gemini 3.7 Flash it printed the date on which the price doubles: January 1, 2027, from $0.75 to $1.50 per million input tokens, from $3.75 to $7.50 on output.

OpenAI did the opposite and it amounts to the same thing. Ultrafast, its 750-tokens-per-second tier running GPT-5.6 Sol on Cerebras hardware, is available by waitlist to selected customers with no published price and no general availability date.

One vendor tells you when the meter changes. The other declines to install a meter you can read. Both are running the same play, and it is a play borrowed from gyms, credit cards and cable television: subsidize the meter while the customer builds their life around it.

The reason this works in software is that the switching cost was never the tokens. It is the eighteen months of prompts tuned to one model's specific failure modes, the eval suites calibrated against its outputs, the retry logic written around its particular way of going wrong, the human processes reshaped around its latency. Migrate to a rival and the tokens get cheaper and everything else breaks. By the time the introductory window closes, the model is not a supplier. It is a landlord, and the tenancy agreement was signed by an engineer who thought they were choosing an API.

What interests me is the second-order effect, which I do not think has been priced anywhere.

Every AI application company currently modelling its gross margin is doing so in a currency with an expiry date stamped on it. The seed-stage deck that shows a path to 70 percent margins at scale is quoting a rate that a footnote already voids. When the step-up lands, it lands simultaneously across the entire sector, on the same day, on companies that have spent the interim raising at multiples justified by those margins.

That is a synchronized repricing event with a known date, which is an unusual thing to be able to see coming and an unusual thing for nobody to be discussing. I would expect it to be absorbed quietly by the well-capitalized, to kill a visible cohort of thin-margin agent startups in the first quarter of 2027, and to produce a wave of migrations to whichever vendor has not yet reached the end of its own introductory window. Which is to say: the discount cycle will rotate, and the same trap will be reset by a different landlord.

There is a broader version of this that runs through the rest of today's paper. An emergency capacity price cap that binds three auctions running. A crime emergency with an appropriation through 2029. A shipping lane described by the man designing it as temporary and final in the same sentence. The pattern is not conspiracy. It is that provisional arrangements attract almost no scrutiny precisely because everyone assumes they will be revisited, and the revisiting is what never happens.

So, a practical suggestion for anyone signing anything this quarter. Find the footnote with the date in it. Model the business at the post-expiry price. If it does not work at $1.50, you do not have a business at $0.75. You have a subsidy, and subsidies are the property of whoever is paying them.