VANCOUVER — British Columbia sued OpenAI on Monday over a telephone call that would have taken ninety seconds.

"One telephone call to the RCMP could have prevented the tragedy," the complaint filed in federal court in San Francisco asserts. It names the company and its chief executive, Sam Altman, as defendants. The co-plaintiff is School District 59, Peace River South, which runs a secondary school in a logging town of two thousand people in the province's northeast corner.

On February 10, an eighteen-year-old former student named Jesse Van Rootselaar killed her mother and her half-brother at home, drove to Tumbler Ridge Secondary School, killed five students aged twelve and thirteen and an educational assistant, and then killed herself. Eight victims, in one of the worst mass shootings in Canadian history.

Eight months earlier, in June 2025, OpenAI's own systems flagged her.

That gap is the whole case. What distinguishes Monday's filing from the thirty-seven private suits already pending in the same courthouse—seven filed by Edelson PC in April, thirty more this month—is the identity of the plaintiff. Private plaintiffs allege injury. A government is asking an American judge to find that a California corporation owed a legal duty of care to a Canadian town it had never heard of.

Attorney General Niki Sharma is seeking money to rebuild the school, punitive damages, and an order requiring safeguards that "reliably refuse, terminate or de-escalate" conversations in which a user expresses intent to harm others. That second request is the consequential one. It would put a federal judge in the business of specifying a content moderation architecture, which courts have spent thirty years declining to do.

## The chain of command

The allegation that gives the case its force is not that the model said something terrible. It is who decided what to do afterward.

According to filings in the related private litigation, OpenAI's Intelligence and Investigations team reviewed the flagged conversations and concluded they "posed a credible threat of gun violence against real people." The team recommended contacting the Royal Canadian Mounted Police. The plaintiffs allege the recommendation was overruled by the company's global affairs organization—the shop responsible for public relations, policy, and political relationships, led by the political strategist Chris Lehane—and ratified by Altman. Lawyers for the victims say twelve employees pressed the company to act. OpenAI has disputed that Lehane supervised or directed the threat assessors.

The company refers matters to law enforcement, by its own stated policy, when conversations indicate "an imminent and credible risk of serious physical harm to others." It told Canadian media in February that Van Rootselaar's activity did not clear that bar. It banned her account. She opened another one.

The plaintiffs have an answer to the imminence defence, and it is the most damaging paragraph in the litigation. In November 2025, after an activist allegedly threatened the company, OpenAI locked down its San Francisco offices, circulated the man's name and photograph to thousands of employees, and called the police. There was, the complaints note, no indication of an imminent attack in that case either. "When its own people were at risk," one filing reads, "no 'privacy' interest stopped OpenAI from notifying the police."

In April, Altman published an open letter in the *Tumbler RidgeLines*, the local paper. "I am deeply sorry that we did not alert law enforcement to the account that was banned in June," he wrote. The province's complaint says the reforms he promised did not follow, and that Sharma's office could not get the company to engage.

## The forum is the argument

There is an irony in the venue. OpenAI has asked the court to dismiss the earlier Tumbler Ridge suits on the grounds that the witnesses, records and harm are all in Canada and the litigation belongs in British Columbia. British Columbia has now voluntarily walked into California. Sharma's reason is precise: the decision not to call the police was made in California, and that is where the omission occurred.

It is also where the doctrine lives. The duty to warn an identifiable potential victim of a credible threat is not an invention of the AI era; it is *Tarasoff v. Regents of the University of California*, decided by the California Supreme Court in 1976, built on this fact pattern—a professional who knew, an institution that reviewed, and a warning never sent. The province is not asking for a new law. It is asking whether a company that runs threat assessments with trained investigators has made itself into the kind of institution that has to finish the job.

Sharma has also written to Ottawa asking for Criminal Code amendments, an acknowledgement that she may lose. A loss in San Francisco would still produce a public docket, and the docket is arguably the point: the province has asked for the conversations, and Sharma told reporters she has not read them.

The ground has shifted underneath the defence. Treasury Secretary Scott Bessent, pinning the recent Hugging Face breach on OpenAI's management rather than on the autonomous agents involved, has effectively closed off the federal liability shield the frontier labs wanted.

It will be years before any of this resolves. But what changed on Monday is smaller and harder than a statute. A sovereign government has formally asserted that when an AI company's own safety staff writes down the words *credible threat*, something is legally required to happen next.

NEW YORK — The eighty-first session of the United Nations General Assembly opens its General Debate today, and by the standards of the seventh month of a war it has an unusually specific agenda: one waterway.

President Masoud Pezeshkian is due in New York today. His foreign minister, Abbas Araghchi, arrived over the weekend, spoke to Secretary-General António Guterres by phone, and told reporters that Iran remains in a "state of war" following the US-Israeli strikes. President Trump, who addresses the Debate today, told Fox News on Sunday that Iran's options were "wiping Iran out," letting its economy "rot," or a deal, and that he is in "deciding mode." Asked whether he would meet Pezeshkian on the sidelines, he said he was "probably open" to it.

Nothing has been confirmed by either side, and an Iranian lawmaker has already publicly rejected the idea of direct presidential talks. Washington withheld visas from some members of the Iranian delegation before issuing them to senior officials. The US ambassador to the UN, Mike Waltz, previewed the president's speech by saying he would continue to "call out the ridiculous."

## The mediators are more interesting than the principals

The substantive channel is not at the podium. Qatari foreign ministry spokesman Majed Al-Ansari said Doha's mediators have spent nearly two weeks travelling between capitals exchanging proposals, and have received assurances from American officials that Washington wants an agreement. Pakistan is also carrying messages. What Qatar is trying to assemble is explicitly not a peace: it is a short-term arrangement narrow enough to reopen negotiations.

Al-Ansari named the two conditions Doha regards as unavoidable. Restore navigation through Hormuz. Remove the threat of military action.

Tehran's published terms are a longer list—a ceasefire on all fronts, the reopening of Iranian ports, sanctions relief, the release of frozen assets, an end to the American naval blockade. The June memorandum that collapsed contained versions of several of them. But the asymmetry is clean: Iran wants its own ports and money back, and the thing it controls is everyone else's shipping lane. That is the trade, and everything said this week in a suit will be decoration on it.

Iran's military spent the weekend warning that its intelligence indicates a new large-scale strike is being prepared by the United States and its allies, promising "painful" retaliation across the region and stating that any regional government backing such a strike would be treated as a party to the conflict. That last clause is aimed at the Gulf states now hosting the tanker workarounds described elsewhere in this issue.

## The market already voted

The most legible read on the diplomacy is not in any communiqué. Brent fell 3.1 percent on Monday to $100.62, after touching nearly $110 last week, on nothing more than the possibility of contact in New York. WTI dropped over three percent toward $97. Ten-year Treasury yields fell with it, from a nineteen-year high of 5.041 percent last week to 4.96 percent. French ten-year yields fell more than twelve basis points.

That is a global asset market moving several hundred basis points of risk on a maybe-meeting between two men who have not agreed to meet. It tells you how much of the current price of everything is a bet on a strait.

It also tells you the size of the disappointment available. Oil was around $72 earlier this summer. The distance between the current price and that one is the option value of a handshake that, as of this morning, nobody has scheduled.

Seventeen trackable commodity vessels crossed the Strait of Hormuz over the weekend of September 19 and 20, according to Kpler. The previous weekend it was thirty-seven. In a normal month it is a river.

And yet crude clearing Hormuz is running at roughly 6.5 million barrels per day in September, the highest sustained rate since the brief June ceasefire. Both facts are true, and the reconciliation between them is a piece of maritime improvisation that has quietly become one of the largest cost shocks in the global economy.

## How it works

A shuttle tanker loads inside the Gulf and crosses the strait. Outside, off Fujairah or Sohar, it ties up alongside a larger vessel, connects hoses, and pumps its cargo across. The receiving ship—the industry calls it the daughter vessel, which is either charming or grim depending on the week—sails for Asia. The mother vessel turns around and does it again.

ADNOC invented the pattern in April, faced with a tanker shortage and owners who would not send hulls into the strait. Iraq's state marketer Somo followed with its first direct ship-to-ship tender in late August. Saudi Aramco, which has historically sold almost exclusively on term contracts, began offering September-loading Arab Medium and Arab Heavy on a transfer basis off Sohar.

Roughly 2.5 million barrels a day will move this way in September, up from 1.4 million in August—about forty percent of everything crossing Hormuz. Before the war it was a rounding error.

## What it costs

Very large crude carrier freight from the Middle East Gulf to China has been quoted at or above $30 a barrel, the highest on record by a wide margin. Cargoes loading outside Hormuz go for roughly $20. Historically freight was two to three percent of the delivered cost of a barrel. It is now over a quarter.

Sit with that. Crude itself has moved from about $72 this summer to $100.62 at Monday's close—call it forty percent. Freight has moved by an order of magnitude more, and it shows up in the delivered price of diesel in Rotterdam and Singapore rather than on any screen a retail investor watches.

The mechanism is not mysterious. Every barrel now consumes more vessel-time: a shuttle leg, an anchorage, a transfer, and a return. The same physical oil requires something like twice the ship. Tanker supply is finite and global, so the Gulf's improvisation is bidding hulls away from the Atlantic basin and raising freight for people who have never been near the Persian Gulf. This is why diesel is at a record in the American Midwest.

## Two things to watch

First, capacity. Kpler believes the transfer zones off Fujairah and Sohar are at or close to their limits. There is no obvious third location; the candidates are offshore west India and Malaysia, both of which add days. The workaround runs out of parking before it runs out of ships.

Second, concentration of risk. The arrangement works by pushing the dangerous leg onto a small group of shuttle tankers and trading houses while buyers and their vessels stay outside the strait. Iran has attacked ADNOC shuttles repeatedly. A few dozen hulls and a handful of charterers now carry the physical risk for a very large share of Asian energy supply, with AIS transponders switched off—which means the most important freight market in the world has no reliable public record.

Saudi Arabia shows how thin the redundancy has become. Attacks forced the kingdom to shut its East-West pipeline on September 13; there have been no visible loadings at Yanbu since the 16th. Its response was to push everything back through the strait it cannot secure: Saudi crude through Hormuz averaged 2.9 million barrels a day over one recent six-day stretch, against 700,000 in August.

That is not resilience. That is a single point of failure with better logistics.

The single most revealing thing the Federal Reserve did on September 16 was not raise interest rates. It was a deletion.

In July, the committee's statement said inflation was elevated "in part reflecting supply shocks that have driven price increases in certain sectors, including energy." Last week that clause was gone, replaced by a note that consumer and business spending "has been resilient." The vote was 12-0 for a quarter point, to a 3.75–4.00 percent range—the first increase since 2023—and almost everyone on the committee signalled a second before year end.

A central bank does not delete a sentence like that by accident. Deleting it is the argument. It says: we are no longer going to treat this as something that happens to us.

## What is actually happening to them

August CPI came in at 3.4 percent year over year, unchanged from July. Gasoline rose 3.9 percent in the month alone and accounted for more than a third of the total. The national average pump price hit $4.44 a gallon, up 18.5 cents in a single week and $1.30 from a year ago. Diesel set a record above $6.40, up more than seventy percent since the day before the war began. Moody's chief economist Mark Zandi put the war's cost to the average American household at $1,650 as of early September.

None of that is demand. The Fed's own projections concede the point in the most uncomfortable way available: headline PCE at 3.7 percent this year, core at 3.4, and no return to two percent until 2029. A central bank that expects to miss its target for three more years has effectively told you that the thing raising prices is not the thing its instrument touches.

Chairman Warsh's defence is credibility, and he made it plainly: "the plain fact is that inflation is too high and has been for too long," and the hike "starts to show we're serious about this." He cited three intermeeting developments—stronger activity, insufficient disinflation, and geopolitics.

The less flattering account, which CNN put bluntly last week, is that the bond market issued an ultimatum. The ten-year crossed five percent for the first time in three years and touched 5.041, a nineteen-year high. The thirty-year sits above 5.29. The Fed hiked, and the curve responded by flattening hard: the ten-year fell back to 4.963 on Monday while the two-year *rose* to 4.755, compressing the spread to 20.8 basis points, the tightest since March of last year. That is the market saying it believes the tightening and doubts the growth.

## The part that should worry you

Real average hourly earnings have fallen 0.7 percent between February and August. Wage growth has trailed inflation for five consecutive months. On a twelve-month basis, real hourly pay is down 0.3 percent. Prices sit twenty-seven percent above January 2021.

Consumer spending, which is sixty-eight percent of GDP, has nonetheless held up. It has held up because affluent households are spending out of portfolio gains rather than paychecks, and the S&P 500 closed Monday within 1.3 percent of its all-time high, carried in large part by the artificial intelligence buildout.

So the structure is this: a war raises energy prices; energy prices raise inflation; inflation exceeds wages and erodes the paycheck-financed consumer; the wealth-financed consumer keeps spending because the AI trade is up; and the AI trade's capital expenditure is, by the Fed's own account, itself an inflationary input. Officials explicitly named expanded AI investment as a potential inflationary factor in their September deliberations.

Warsh is now tightening into an economy in which one of the two things he is trying to cool—the buildout—is the only reason the other one—the consumer—is still standing.

The committee meets again on October 27 and 28. Between now and then the variable that matters most is not a data release. It is whether anything happens in New York this week that reopens a strait.

The most common explanation for the memory crisis is that artificial intelligence is eating the world's RAM. It is true and it is also the least interesting half of the story.

Three companies make nearly all the world's DRAM: Samsung at 38.5 percent of the market in the first quarter, SK Hynix at 28.8, Micron at 22.4. Those three decided, rationally, to route wafers away from commodity memory and toward high-bandwidth memory for AI servers, where margins are better and buyers sign five-year agreements. Nvidia has a memory supply arrangement with SK Hynix reported at roughly $500 billion. HBM is sold out.

The consequence is a two-market split. In the AI segment, pricing power sits entirely with the makers and volume is contracted years out. In the conventional segment—laptops, phones, ordinary servers—demand is left to scramble for whatever wafers remain. That is why consumer memory prices have roughly quadrupled in a year despite consumer demand for memory doing nothing of the sort. DDR4 spot reached $42.50 in August.

## The numbers, in order of how much they should alarm you

J.P. Morgan estimates DRAM prices will have risen more than 400 percent from the start of 2024 to the end of 2026. Conventional DRAM contract prices rose 93 to 98 percent in the March quarter and another 58 to 63 percent in June—two consecutive quarters that roughly tripled the cost of the same chip. TrendForce expects a comparatively gentle 13 to 18 percent in the third quarter, which is what passes for good news.

TrendForce also projects that DRAM and NAND will consume 47 percent of cloud service providers' hardware capital expenditure this year and 68 percent next year. Server DRAM component costs are tracked to rise 270 percent year over year by the end of 2026; enterprise SSDs 235 percent. Total cloud capex is up 98 percent this year with another 50 percent forecast for 2027.

When two-thirds of a hyperscaler's hardware bill is memory, that is no longer a component line. It is the business.

## It is in the CPI now

This connects to the interest rate story on the previous page. The consumer price index for software and accessories and the producer price index for storage devices have each risen 23 percent since the end of 2024. OVHcloud has warned customers of price increases attributable to memory. Meta raised capital expenditure guidance by $10 billion citing AI hardware and memory costs, which is a polite way of saying it got repriced.

When the Federal Reserve listed expanded AI investment among potential inflationary factors this month, this is the transmission channel it meant. A decision made in three fabrication planning meetings in Korea and Idaho is now a term in an American inflation print.

## When it ends

Not soon, and possibly badly. New DRAM fabs take at least eighteen months to reach volume and the visible pipeline lands late—Micron's Singapore plant and SK Hynix's Cheongju site in 2027, Samsung's Pyeongtaek and SK Hynix's West Lafayette in 2028. Leading-edge nodes need EUV tools with twelve-to-eighteen-month lead times. IDC has DRAM supply growth at 16 percent for 2026. Intel's chief executive has put relief at 2028.

Which sets up the trap. Every fab being funded today is underwritten by an assumption that AI server growth compounds through the rest of the decade. Dell's record $60.9 billion quarter of AI server orders says that is a reasonable assumption this month. But each bottleneck in this buildout has resolved into the next one—first GPU allocation, then grid interconnection and power, now memory—and each resolution has been accompanied by capacity commitments sized for the panic rather than the trend.

If AI server growth slows even modestly, the capacity being poured today does not arrive into a shortage. It arrives into 2028, all at once, and the same three companies that engineered this squeeze will spend a year explaining a glut.

Here is a fact that should bother anyone who works in film marketing. *The Paradise*, a Telugu action drama directed by Srikanth Odela and starring Nani, opens worldwide on Thursday with paid premieres Wednesday night. It has reportedly sold ₹103.75 crore of worldwide theatrical rights, the largest of Nani's career and, per trade reporting, the first time a Telugu star outside the top tier has crossed ₹100 crore on rights alone.

It has not released a trailer.

Not a teaser-then-trailer sequence. Not a trailer held for the last week. There isn't one. In a market where the theatrical trailer is treated as the load-bearing beam of the entire opening—cut for months, timed to the hour, judged instantly by a million-comment consensus—the film simply skipped it and let two songs do the work.

The first was "Yeshanagula," composed by Anirudh Ravichander, written by Kasarla Shyam in a Palamuru folk register, sung by Prabha Bangarigalla and Addula Jangireddy. It escaped the film entirely during Ganesh Chaturthi and became a street anthem—played off trucks, danced to in processions, absorbed into a religious festival that has nothing to do with the movie. By the time the cast performed the hook steps at the pre-release event, the song was no longer promoting the film; the film was riding the song. The second, "Aaya Sher," did the narrower job of establishing the register of Nani's character.

## What the trailer was actually for

A trailer performs two functions and we have always conflated them. It proves the film exists at a certain level of craft, and it tells you how the film will make you feel. The first requires images. The second does not.

A folk hook that thousands of people danced to in the street delivers the second with more conviction than any two-minute cut could, because it arrives without the frame of an advertisement. Nobody watches a trailer without knowing they are being sold something. Everybody who danced to "Yeshanagula" at a Ganesh pandal was, for that moment, just at a party.

That is the innovation, and it is not a Telugu-cinema curiosity. The industry built its marketing around the trailer because the trailer was the only thing it could put in front of a captive audience in a dark room. That audience no longer exists in the same form, and the thing that does travel—a fifteen-second sound with a dance attached—is not a compressed version of a film. It is a different product that happens to be attached to one.

## The wager

The risk is severe and worth being honest about. Trade estimates put the break-even at roughly a ₹105 crore worldwide share, around ₹200 crore gross. Nani's career best is about ₹119 crore worldwide *total*. The film has to nearly double the biggest thing he has ever done just to clear the line.

The advances suggest it might. Roughly 72,000 tickets and ₹7.85 crore across the first four days as of Tuesday morning, with day-one gross crossing ₹1.25 crore before block seats and ₹3.81 crore including them. Trade projections run from ₹50 crore to ₹60 crore worldwide on day one against a previous personal best opener of ₹37.80 crore. North American premiere pre-sales stood at $561,960 from 24,255 tickets across 505 locations—reportedly the fastest $100,000 any Telugu film has recorded, on a single chain, without inflated premium pricing.

It runs two hours and forty-nine minutes and carries an A certificate, which caps the audience.

So it may not work. A hundred and four crore of rights is a number that can produce a disaster as easily as a triumph, and songs sell an opening weekend rather than a second one. But the opening is already sold, and it was sold by a wedding-season folk track and a hook step. If the film holds, some very expensive marketing departments are going to spend October explaining what a trailer is for.

On September 12, Dario Amodei published an essay called *We Must Pace the Frontier*. It said: "We must slow the pace at which we improve the capabilities of AI models." It called for "industry-wide coordination" among frontier laboratories to establish "limits on the rate of unchecked AI progress." And it explained—this is the sentence that will be read aloud in a courtroom—that coordinated slowing would let each participant reduce its pace "without sacrificing commercial advantage."

By the end of that day, Sam Altman, Elon Musk and Demis Hassabis had each publicly agreed. On September 18, a class of paying subscribers sued all four companies under Section 1 of the Sherman Act.

Let me be precise, because the easy version of this column—*look at the hypocrites*—is cheap and mostly false. I do not think Amodei is lying about being frightened. The fear is sincere and some of it is warranted. That is exactly why the essay is a problem.

## The waiver he asked for and did not get

Amodei anticipated the antitrust objection and answered it. He wrote that it would be helpful for the United States government to mediate, or at least enable, cross-lab discussions, and that the government would need to issue a narrow waiver for certain kinds of safety conversations.

No such waiver exists. Congress has passed no exemption. No agency has compelled the conduct. Senator Josh Hawley said in a recent hearing that "there is no world" in which he would grant the most powerful companies in history an exemption from the antitrust laws. Treasury Secretary Bessent has made clear the administration regards AI safety as a duty of management, not a collective industry project. Cohere's Aidan Gomez called the proposed self-regulatory body "a cartel by any other name."

Then on September 15, OpenAI's chief global affairs officer Chris Lehane confirmed his company had been working with Anthropic and Google DeepMind on these questions for weeks *before* the essay, and that the firms believed they could proceed without waiting.

That is the whole case. They identified the legal barrier, requested its removal, were refused in public, and went ahead. The complaint's most elegant line is that the agreement "was proposed in public, accepted in public, and confirmed in public." This is not a smoke-filled room. It is a smoke-filled Substack.

## The category error

Here is what the plaintiffs get right, and it is subtler than antitrust pedantry. They do not object to any lab slowing down on its own. They object to the "shortcut" of substituting collective restraint for individual accountability.

That distinction is the entire moral argument. If Anthropic believes a capability is dangerous, it can decline to ship, eat the competitive cost, and be judged for the decision. That is expensive, legible and attributable. Amodei proposed instead a world where no firm bears the cost of caution and none bears the blame for a failure, because the standard was set jointly. Responsibility diffused across four companies belongs to none of them.

We have a demonstration of where this leads on the front page of this paper. In the Tumbler Ridge litigation, OpenAI's threat assessors recommended calling the police and were allegedly overruled by a communications organization. The safety function existed. It produced the correct output. It was then routed into a decision process where nobody's name was on the outcome.

A safety cartel is that failure mode, scaled to an industry and given a charter.

The suit may lose. Sherman Act cases built on public statements are hard, and the defendants will argue that agreeing with an essay is not a contract. But the labs have spent two years insisting they are uniquely positioned to understand the risks of their technology, and they are right. The remedy for being uniquely positioned is to act alone and be answerable. It is not to form a committee and call it conscience.

Strip the British Columbia complaint of its horror for one minute and look at it as an organizational chart, because that is what it is.

OpenAI built a threat assessment function. It staffed it with people trained to evaluate whether a person intends to commit violence. In June 2025, that function reviewed a set of conversations and reached a professional conclusion: credible threat of gun violence against real people. It made a recommendation: call the RCMP.

And then, according to the filings, the recommendation went *up*. Into a global affairs organization responsible for public relations, public policy, and political relationships. Where it died.

The company disputes that account of the reporting lines, and it will get its day to prove otherwise. But nobody disputes the shape of the thing, because the shape is normal. It is how almost every technology company is built. Trust and safety recommends. Policy and communications decides. The people who know sit two levels below the people who choose, and the people who choose are measured on something other than whether the recommendation was right.

This should keep executives awake, and not only at AI labs.

## The word 'imminent' is doing all the work

OpenAI's defence rests on a threshold: it notifies law enforcement when a conversation indicates an "imminent and credible risk of serious physical harm to others." Credible was satisfied. Imminent, the company concluded, was not. Eight months later, eight people were dead.

Here is the problem with imminence as a corporate standard. It is not a safety concept. It is a liability concept. Imminence is what you require when the cost you are optimizing against is the cost of a false positive: the awkward call, the wrongly flagged user, the news story about a company that reported a teenager to the police over a video game. Those costs are real, and they are borne entirely by the company. The cost of a false negative is borne by a school district in a town of two thousand people.

An organization that routes a threat assessment through its reputation management function will discover imminence thresholds. It cannot help it. That is what the function is for.

## The counterexample is in the same complaint

In November 2025, an activist allegedly threatened OpenAI. The company locked down its San Francisco offices, distributed the man's name and photograph to thousands of employees, and notified the San Francisco Police Department. On the plaintiffs' account there was no more evidence of imminence in that case than in the other one.

I do not read that as villainy. I read it as the most ordinary thing in the world: when the exposure was internal, the decision was made by people who would personally bear the consequence of being wrong, and they chose the false positive without hesitation, correctly. When the exposure was a Canadian town, the decision was made by people whose downside was a headline.

That is not a policy failure. It is a design failure, and it is fixable in one line: the authority to escalate a credible threat to law enforcement must sit with the assessors, not above them, and must not be reviewable by anyone whose job description includes the word *communications*. A safety function that has to win an argument with the press office before it can dial a phone number is not a safety function. It is documentation.

British Columbia is asking a federal judge to impose that design by injunction, which is a poor way to run an industry and may well fail on jurisdiction. Every company that would rather not be redesigned by a court has roughly one litigation cycle to redesign itself.

The cheapest safety intervention available in this industry costs nothing, requires no research, and has been sitting on the table since June 2025. It is letting the person who wrote *credible threat* pick up the phone.

A newspaper that makes claims about the near future should be gradeable. This is the second daily installment of The Wall. Each entry is a specific, falsifiable expectation with a date attached. Where we are uncertain, we say the odds rather than hiding behind a verb.

Today, September 22. Trump addresses the General Debate; Pezeshkian arrives in New York. *Our expectation: no announced bilateral meeting between the two presidents this week.* We think the probability of a confirmed Trump–Pezeshkian meeting before the Debate closes is below one in four, given that an Iranian lawmaker has already publicly rejected direct presidential talks and the Qatari channel is still working at the level of a short-term arrangement rather than a summit. If they meet, we were wrong and we will say so at the top of the paper.

September 23–24. *The Paradise* opens, with paid premieres Wednesday night. Trade projects ₹50–60 crore worldwide on day one against a break-even near ₹200 crore gross. *Our expectation: the opening lands inside that projected range, and the second-weekend hold is the number that decides whether songs can replace a trailer.* We will publish the day-one figure Friday and the weekend-two drop the following Monday, whichever direction it cuts.

Late September. Kpler's monthly Hormuz throughput. September crude clearing the strait is running near 6.5 million barrels per day with roughly 2.5 million of that transferred ship-to-ship in the Gulf of Oman. *Our expectation: October ship-to-ship volumes do not rise materially above September's, because Fujairah and Sohar are at or near anchorage capacity.* If transfer volumes jump anyway, it means a third transfer zone opened and we missed it.

Early October. Q3 DRAM contract prints. TrendForce forecast 13 to 18 percent. *Our expectation: the print comes in at or above the top of that range for server DRAM specifically, and below it for mobile.* The allocation split is the thesis; if server memory prices come in soft, the thesis is weaker than we have written it.

October 27–28. The Federal Open Market Committee. *Our expectation: a second 25 basis point hike, conditional on Brent holding above roughly $90.* Nearly every member signalled another move; the only thing that stops it is an oil collapse out of New York this week. If Brent breaks below $85 and the Fed hikes anyway, Warsh is targeting something other than energy and we will have to rewrite our read of him.

This autumn, date to be set by the docket. OpenAI's response to British Columbia, and its pending motion to move the earlier Tumbler Ridge cases out of California. *Our expectation: OpenAI moves to dismiss the province's suit on jurisdictional and forum grounds rather than contesting the June 2025 facts.* We also expect the responsive pleadings in *Buist v. Anthropic* to argue that public agreement with an essay is not a Section 1 agreement. Neither prediction requires a leak; both are visible on PACER when they land, and we will read them.

Standing, carried from yesterday. The tariff list authorized under the Russia sanctions law is due October 18. We said we had marked it. It remains marked.

If you think any of these is wrong, write and say which one and why. Letters that name a date and a number will run ahead of letters that don't.