The White House will seat Tim Cook, Jeff Bezos, Sundar Pichai, Sam Altman, Elon Musk and Jensen Huang at a state dinner for Xi Jinping on Thursday evening. Reuters reports they are there in part to press for a loosening of export controls on advanced chips, and that people familiar with American planning expect the controls to hold. Analysts briefed on the agenda expect the tangible output of the summit to be an extension of a trade truce and a promise to study the dangers of artificial intelligence.

That will be the story on most front pages Friday. It is the smaller story.

The larger one closed at four o'clock Wednesday afternoon in a market with no seating chart. The benchmark ten-year Treasury note settled at 5.11 percent. AFP called it the highest since 2007. The mortgage desks called it a level last seen in 2006. We are not going to resolve that disagreement for you; we note only that the professionals who watch this number for a living had to reach back roughly two decades to find a comparison, and that they reached back to slightly different years.

The move did not come from the Middle East, though Brent rose 3.86 percent to $103.08 on Tehran's refusal to soften at the United Nations. It came from a strong September purchasing managers' reading and from Federal Reserve Governor Michael Barr, who told an audience in Chicago that before last week's quarter-point increase the central bank had been "out of position," and that "in my base case, further policy adjustments are likely to be needed." The Fed raised its benchmark to 3.75 to 4 percent last week, its first increase in more than three years. Sixteen of eighteen policymakers indicated at least one more is coming. By Wednesday's close, traders were pricing roughly a seventy percent chance of another move on October 28. The Dow fell 0.68 percent, the S&P 500 0.75 percent, the Nasdaq 1.13 percent. The thirty-year mortgage reached 7.26 percent, the highest of the year.

Now put that against the plan the dinner guests are executing.

Aggregate capital expenditure by the five largest hyperscalers is projected to approach $800 billion this year and to exceed $1 trillion annually from 2027 through 2030. Between 2020 and 2024 those five companies issued, on average, about $35 billion of debt a year between them. In 2025 the figure was $93 billion. So far in 2026 it is roughly $132 billion, including a single multi-tranche offering of about $53 billion — among the largest corporate bond sales ever recorded — and a century bond, a hundred-year instrument, sold to finance equipment that is functionally obsolete in three.

Goldman Sachs Research counts nearly $500 billion of AI-related debt issuance across the wider ecosystem this year and estimates that about a third of capital expenditure will be debt-financed. Apply that share to a $1 trillion annual programme and you get roughly $330 billion of new borrowing a year, every year, through the end of the decade.

Here is the arithmetic nobody read out at dinner. The plans were drawn when the risk-free rate had a four in front of it. Call the move a point. A point on $330 billion is about $3.3 billion a year in additional interest — for one year's issuance. Stack five vintages of that by 2030 and the rate move alone, before a single dollar of credit spread, has added something in the neighbourhood of $16 billion a year to the running cost of the buildout. That is not a forecast. It is multiplication, and it is the most conservative version available, because it assumes the ten-year goes no higher and spreads never widen.

What makes this the lead rather than a note in the business section is who can do anything about it. Not the six men at the table. Their ask is jurisdictional: licences, tariffs, rare earths, a notification mechanism for AI risk. Every item on that list is within the gift of the two governments in the room. The number that actually determines whether a $1 trillion annual programme pencils is set by an auction the White House does not run, attended by buyers who were not invited, who did not applaud, and who spent Wednesday demanding an extra eleven basis points to keep lending the United States money for ten years.

There is a version of Thursday night in which Trump grants something on chips and the coverage records a win. It would not touch this. Export licences change who may sell. They do not change what money costs.

The next real negotiation is on October 27 and 28, and none of the guests are on the invitation list.

A newspaper that only publishes claims is asking to be believed. A newspaper that publishes the date its claims become checkable is asking to be audited. We prefer the second arrangement. Here is Thursday's ledger.

Thursday, September 24 — tonight. Whether the Xi state dinner produces any announced change to US export controls on advanced semiconductors. People familiar with American planning told Reuters it will not. *Claim tested by: Friday morning.*

Tuesday, September 29. Iranian security chief Mohsen Rezaei said Wednesday that Washington has five days to meet Tehran's conditions for reopening the Strait of Hormuz. Five days from that statement is the 29th. Either the strait opens, the deadline is extended, or it was never a deadline. *Claim tested by: the shipping insurance rate on the 30th.*

Wednesday, September 30, 9:00 a.m. ET. Tickets go on sale worldwide for HYBE Cine Fest. We are including this because we argued in the Culture section that the theatrical business has been rescued by fixed-time gatherings rather than by films, and the sell-out speed is the cheapest available test of that claim.

October 24–31. The Cine Fest itself, across more than sixty territories, including two live BTS broadcasts from South America.

October 27–28. The Federal Open Market Committee. As of Wednesday's close, traders priced roughly a seventy percent chance of another quarter-point increase. Sixteen of eighteen policymakers signalled at least one more this year. If they do not move, the lead story's arithmetic gets cheaper and our framing was early.

November 10. The original expiry of the Busan trade truce, which capped US tariffs on Chinese goods at about twenty percent after mutual rates briefly topped one hundred.

November 18–19. APEC in Shenzhen. The next scheduled point at which the two governments have to be in the same room.

January 10, 2027. The new expiry, after Treasury Secretary Scott Bessent said Wednesday that he and Vice Premier He Lifeng had agreed to extend. Sixty-one days bought. Note what that date implies about how much either side expects to settle on Thursday.

During 2027. Edgewater Research projects Anthropic's DRAM demand at 40 to 50 billion chips on a one-gigabit-equivalent basis, comparable to Nvidia's current consumption, with about a fifth bought directly from memory makers. In the same year, hyperscaler capital expenditure is projected to pass $1 trillion annually. Both numbers are forecasts by named firms, and both are wrong in some direction.

2036. The $4.6 billion of QTS senior secured notes that financed a Microsoft campus in Fayetteville, Georgia come due. Between now and then, the servers inside that building will be replaced perhaps six times. The bond will not be.

If we are wrong on any of these, the correction runs on the front page under the same masthead. Write to us when the date arrives.

UNITED NATIONS — Two things happened on Wednesday that will not appear in the same story anywhere else.

The first was loud. Iranian President Masoud Pezeshkian stood at the General Assembly rostrum, held up photographs of children killed in a strike on a primary school in Minab, accused the United States of terrorism and of contempt for international law, and said Tehran would never surrender. The American delegation walked out at the start of the speech. It was a direct answer to Donald Trump, who had used the same podium on Tuesday to say he could "annihilate" Iran if negotiations failed, seven months into a war he began alongside Benjamin Netanyahu.

Hours later, Mohsen Rezaei, Iran's security chief, attached a number to it. Washington, he said, has five days to meet Tehran's conditions for reopening the Strait of Hormuz.

That is the first hard deadline either side has put in public since the strait closed. It counts down to roughly Tuesday, September 29.

It is worth being precise about what a five-day ultimatum from Iran is and is not. It is not leverage in the ordinary sense; the party issuing it is the one absorbing airstrikes. What it does is convert an open-ended blockade into an event with a date, which forces the other side to either answer or visibly decline. Oil traders read it the second way. Brent settled up 3.86 percent at $103.08, a near four-percent move that outran a surprise three-million-barrel build in American crude inventories and partial signs of recovering Gulf supply. The market is no longer pricing the war. It is pricing the calendar.

Secretary of State Marco Rubio, asked about terms, said a deal with Iran "requires prolonged effort" and that military options remain available. A senior Iranian official said the two sides are still far apart but that diplomacy must continue. Both statements followed the first shuttle talks between Washington and Tehran in months, which produced no movement in either public position. The dueling speeches bracketed those talks, which is one way of saying the talks were bracketed by their own failure.

The second thing that happened Wednesday was silent.

The Pentagon updated the Defense Casualty Analysis System, the official database of American military deaths, to show nineteen fatalities connected to the war with Iran. The previous figure was eighteen. There was no announcement, no name released, no briefing. The number simply changed. The same database lists more than 820 additional service members in its other casualty categories.

One is a number a government publishes when asked. The other is a number a government publishes because the system publishes it.

The five-day clock will attract the coverage, because a deadline is a story shape and an ultimatum has a natural second paragraph. But the nineteenth entry is the one that tells you what this week cost, and it arrived without a podium, a photograph, or a walkout.

WASHINGTON — The most consequential thing to come out of the Trump–Xi summit was announced on Fox News the night before it started.

Treasury Secretary Scott Bessent said Wednesday, after talks with Chinese Vice Premier He Lifeng, that the two sides had agreed to extend until January 10 the arrangement pausing a trade war that briefly pushed mutual tariffs above one hundred percent. The underlying deal, struck in Busan last November, capped American tariffs on Chinese goods at roughly twenty percent and was due to lapse on November 10.

So the headline deliverable of Thursday's state visit was banked before the state visit. Sixty-one days, purchased in advance, so that the dinner would not have to produce them.

That tells you what to expect from the dinner. Reuters describes a lavish summit "heavy on symbolism and light on substance." Analysts briefed on the agenda expect tangible outcomes to be limited to the truce extension and an agreement to examine the risks posed by artificial intelligence.

What is genuinely live, in rough order of how badly each side wants it:

Critical minerals. Washington wants secure rare-earth and magnet supply. Beijing committed to restore flows under the Busan truce; a senior US official said last week that Chinese performance had "not been up to" that commitment. Beijing, for its part, wants assurance that export controls will stop being used as a general-purpose weapon — a request it is making while using its own the same way.

Agriculture and non-tariff barriers. US Trade Representative Jamieson Greer said this month there will be "some announcements." Trump wants a win in front of the midterms, and soybeans photograph well.

Artificial intelligence. The two sides have agreed to launch a formal US–China AI dialogue, and Washington has proposed a notification mechanism for national-security risks — a hotline, essentially, for frontier models. Talks are expected to cover both open- and closed-weight systems, a live issue because Chinese open-weight models have been winning American customers on price against closed systems from OpenAI and Anthropic.

Chips, which are the actual ask. Any American concession is expected to be limited to licensing some already-restricted parts. Frontier accelerators are not expected to move. This is the thing Jensen Huang and Sam Altman are at dinner for, and it is the thing least likely to be granted.

Also on the list and unlikely to be resolved: Taiwan arms sales, secondary sanctions tied to Iran, the war in Ukraine, Chinese EV plants on American soil, the unfinished Board of Trade agreed in May, and the stalled parallel Board of Investment. Trump is expected to press for the release of Americans detained in China.

The next checkpoint is APEC in Shenzhen, November 18 and 19. The deadline that matters is January 10. A summit whose principal function is to keep a clock from running out is not a reset. It is a payment on an instalment plan neither party has agreed on the price of.

WASHINGTON — Vice President JD Vance and Centers for Medicare and Medicaid Services Administrator Mehmet Oz announced Tuesday that the administration is removing 760,000 people from Affordable Care Act exchanges, cancelling roughly 315,000 enrollments on the grounds that the enrollees were signed up fraudulently or do not exist. The claimed saving is $2.2 billion. Another 419,000 enrollments have been flagged for additional verification. New agents and brokers are barred from signing anyone up for six months.

The administration's strongest argument is real, and it deserves to be stated properly before it is picked apart. The Government Accountability Office reported last year that the federal marketplace approved subsidised coverage for most of the fictitious applicants it submitted during a covert test. Broker-driven fraud — enrolling people in plans without their knowledge to harvest the federal premium subsidy — has been documented repeatedly. CMS officials say that is the mechanism they found, and that the losses were not primarily about claims spending.

Then Oz explained how the accounts were identified. Among the signals: the enrollees had never filed a claim and could not be reached.

This is where the announcement fell apart in public, and it fell apart fast. A customer who pays into an insurance pool and never draws from it is not an anomaly. That customer is the business model. Insurers spend enormous sums acquiring exactly that person. Treating a clean claims history as evidence of non-existence inverts the entire logic of risk pooling, and within hours it was being mocked by people who sell insurance for a living.

The deeper problem is the collision between the two halves of the government's own case. If the fraud was committed by brokers enrolling real people without their knowledge — which is what CMS says it found — then those people are not fictitious. They are real, they have addresses, and the cancellation notice arrives at one of them. The accounting recovers $2.2 billion from a ledger. The letter arrives in a mailbox.

Context matters for how that letter will land. Premiums doubled or tripled for many enrollees after Congress allowed the enhanced subsidies to lapse, prompting millions to downgrade plans or leave the exchanges entirely. A Gallup survey in June found that fewer than half of Americans can consistently afford health care — the first time that has been true in five years. Affordability is the central issue of the autumn midterms, and this announcement is the administration's answer to it: not a price, a purge.

The number to watch is not 760,000. It is whatever fraction of that 760,000 turns out to be a person who calls a pharmacy in January and discovers they are not covered. CMS has not said how a wrongly cancelled enrollee gets reinstated, or how long it takes. We have asked.

Until that process is described, "never filed a claim" remains the government's phrase, not ours, and it is the sentence this policy will be remembered by.

In April, QTS Data Centers, owned by Blackstone, sold $4.6 billion of 5.700 percent senior secured notes due 2036 to build a campus in Fayetteville, Georgia that will house servers for Microsoft. The law firm advising the banks called it the first non-recourse single-asset project bond ever sold by a data-centre project company in the Rule 144A market. Peak demand reached roughly $12.5 billion. Moody's rated it Baa2, two notches into investment grade.

Read that sentence again with attention to two words. *Non-recourse.* *Project.*

Non-recourse means that if the campus underperforms, the lender's claim runs to the campus and not to Blackstone. Project means the bond is a bet on a building and a contract, not on a company. Investors bought $4.6 billion of it, and demanded nearly three times that, because there was a Microsoft logo in the description.

On June 10, Citigroup credit analysts Daniel Sorid and Mathew Jacob noted that the spread on the QTS bond had widened by more than thirty basis points. Nothing had happened to Microsoft. What had begun to happen was that the market started pricing the structure instead of the tenant — which is what project finance investors are supposed to do, and which two years of AI enthusiasm had trained them out of.

The scale behind that one deal is the reason it matters. Goldman Sachs Research counts nearly $500 billion of AI-related debt issuance so far in 2026 and about $194 billion from the hyperscalers alone, with roughly a third of total capital expenditure expected to be debt-financed. Estimates for full-year AI-related issuance across chipmakers, developers and utilities run from $300 billion to $570 billion. Against that sits about $4.5 trillion of private-market dry powder across credit, infrastructure, real estate and private equity — which is the real reason nobody expects the financing to stop.

The capital stack has sorted itself into a sequence. Private credit funds construction, because it can absorb development risk and move fast. Public corporate bonds fund the large operators. Securitisation — ABS and CMBS — takes over once a campus is built and leased and throws off predictable cash flow. Each layer is individually sensible. Together they mean the risk ends up somewhere other than where the coverage looks.

The Bank for International Settlements has described the mechanism plainly. A hyperscaler sets up a joint venture or special-purpose vehicle with outside sponsors, takes a minority stake, commits to long-term leases or capacity offtake, sometimes adds guarantees — and the debt is serviced by lease payments and held by private credit funds. Economically the company has swapped capital expenditure for a multi-year operating expense. Accounting-wise, most of the borrowing sits off its balance sheet. Vanguard puts the consequence bluntly: public balance sheets and public bond indexes no longer capture the full financing picture, and the ultimate distribution of risk is harder to observe.

Then there is the duration problem, which no structure fixes. A data centre is a fifteen-year-plus asset. The GPUs inside it are refreshed on a cycle measured in low single-digit years. Lenders are underwriting a building against a revenue stream that depends on the tenant continuing to want to put new silicon in it — and on the operator finding the capital for each retrofit. If the upgrades stop keeping pace with what customers need, the concrete outlives its own relevance while the notes are still outstanding.

The QTS paper matures in 2036. By then the machines that justified it will have been hauled out and replaced perhaps six times.

Whoever still holds that bond in year eleven will be the person who finds out whether "non-recourse" was a technicality or the whole deal.

The research note that moved through Seoul this week is short and its implication is not.

Edgewater Research estimates that Anthropic's DRAM demand in 2027 will reach 40 billion to 50 billion chips on a one-gigabit-equivalent basis — a volume comparable to what Nvidia consumes today. Roughly twenty percent of it is expected to be bought straight from the memory manufacturers, with that share rising after 2028. Samsung Electronics, SK hynix and Micron have all signed on as strategic infrastructure partners. OpenAI agreed with Samsung and SK hynix last October to expand advanced memory supply for its Stargate programme.

Strip out the jargon and what is being described is a supply chain losing a layer. The traditional route ran memory maker to chip vendor to cloud provider to server builder to customer. The model developers are now walking past three of those steps and negotiating tonnage directly, the way a car manufacturer buys steel.

That is not a procurement optimisation. It is a change of species. A company that signs multi-year take-or-pay commitments for physical components denominated in tens of billions of units has acquired inventory risk, working-capital requirements and a balance sheet that looks industrial. It is also, as of Wednesday, a company financing that working capital against a ten-year Treasury at 5.11 percent.

The chip side is moving the same direction. Harrison Kim, general manager of OpenAI Korea, said this week that Samsung Foundry may produce at least some of OpenAI's next-generation processors — which would mean double-sourcing accelerators across Samsung and TSMC, a decision only volume forces. OpenAI already has an agreement with Broadcom for ten gigawatts of custom accelerators, and Broadcom and Samsung announced a foundry, memory and packaging partnership valued at over $200 billion through 2030. Sam Altman has spent recent weeks touring TSMC, Foxconn, Samsung and SK hynix, seeking both funding and priority in the order book. Reports suggest both major labs are also taking smaller 20 to 30 megawatt data-centre deployments simply to get capacity sooner, while the flagship projects run late.

JPMorgan expects custom ASIC shipments to grow from roughly seven million units this year to twelve million in 2027 and fourteen million in 2028. Analysts estimate around seventy percent of all memory produced in 2026 will go into data centres, with finished-goods inventory at Samsung and SK hynix sitting low.

There is a dissenting read, and it is worth holding onto. Some investors think the capacity being added to serve these commitments will arrive into oversupply and gut memory pricing power. Every direct-buy agreement signed today is a bet by both parties on which side of that turn they land. The model companies are betting that memory stays scarce enough to be worth locking. The manufacturers are betting the same commitments protect them when it isn't.

One of those two groups is wrong, and the contracts run through 2028.

Rentrak confirmed this month that the 2026 North American summer became the biggest of all time: $4.765 billion, edging the 2013 record of $4.756 billion by $9.3 million.

Before anybody frames it, the asterisks. The margin is two tenths of one percent. The 2026 season ran 130 days against 2013's 123. The figures are unadjusted for inflation, which means a ticket in 2013 cost meaningfully less than a ticket in August. On a per-day, per-dollar basis this was not the biggest summer ever and everyone quoting the record knows it.

Having said that, something genuinely happened, and it is more interesting than the record.

All four summer months cleared a billion dollars for the first time: May at $1.069 billion, June $1.079 billion, July $1.186 billion, August $1.276 billion — the highest-grossing August ever, and only the third August above a billion unadjusted. The drivers were *Spider-Man: Brand New Day*, Christopher Nolan's *The Odyssey*, and *Toy Story 5*.

Nolan's film is the one to sit with. At $588.9 million since its July 17 opening it is the second-biggest earner of the summer, 12.4 percent of the entire season, and the highest-grossing "original" of the year — original meaning not a sequel, prequel, spin-off or reboot, a definition it satisfies by being an adaptation of a poem from roughly 700 BCE. The industry's most celebrated act of originality this year was reaching back twenty-seven centuries for intellectual property so old the rights had lapsed.

Now the receipt nobody circulated. *Hadestown: The Musical* took $20.7 million and became the highest-grossing capture of a stage performance ever released theatrically. That is a filmed play, in a cinema, outgrossing most of the year's independent releases.

Put it next to the touring numbers. Pollstar's worldwide top 100 grossed $6.12 billion through Q3, the first time the measure has passed six billion at that point, up 5.3 percent on last year. BTS alone took $465.8 million from 38 shows in front of 2,314,140 people. Bad Bunny took $438.6 million from 3.1 million tickets, including 639,656 across ten nights in Madrid. The Weeknd took $261.4 million from 35 stadium shows; the four Paris dates drew 299,578 people and $47.3 million.

And from October 24 to 31, HYBE takes that business into cinemas in more than sixty territories — two live BTS broadcasts from South America, concert films from Seventeen and Tomorrow X Together, and a 3D "cheering party" where the audience is instructed to bring light sticks and sing.

That is the whole thesis. The product that saved the exhibition business is not the film. It is the appointment. A room, a fixed time, other people, something you cannot pause. *Hadestown* and the HYBE broadcasts and, honestly, an IMAX Nolan opening weekend are the same commodity sold at three price points.

Which is why the festival argument about generative video keeps missing. Venice can screen as many synthetic features as it likes. Nothing coming out of a model competes with the thing the theatres are actually selling, because the scarce good was never the images. It was the appointment, and 2,314,140 people bought one.

The guest list for Thursday's state dinner was reported in advance: Tim Cook, Jeff Bezos, Sundar Pichai, Sam Altman, Elon Musk, Jensen Huang. No leak, no scandal, no anonymous source with a grudge. The White House published it because the White House considers it an achievement, and it is one. Getting those six into a single room is harder than getting a bill through the Senate and considerably more consequential.

I want to be careful here, because the obvious column is the cheap one. The obvious column says: plutocrats, access, corruption, gilded age. That column is lazy and, worse, it is wrong about the mechanism. Nobody at that table is bribing anybody. They are doing something more effective and entirely above board. They are being *present* at the moment a decision is made, which is the only form of influence that has ever reliably worked.

So the objection is not that the dinner is corrupt. The objection is that it is efficient.

American policy on advanced semiconductors — who may sell what to whom, under what licence, at what threshold of capability — is the most economically significant regulatory question of the decade. It will be discussed Thursday evening, over a meal, by a group whose composition was determined by a protocol office. There is no docket. There is no transcript. There is no comment period. There is no adversarial party in the room, because the only entity with a directly opposed interest is the Chinese state, and it is a guest.

Meanwhile, consider who is financing the thing being negotiated. Pension funds hold the century bond one of these companies issued. Insurers hold the $4.6 billion of non-recourse paper on a Georgia data campus. Private credit funds — which is to say, ultimately, endowments and retirees — hold the off-balance-sheet debt on vehicles most of them cannot name. And a household that locked a mortgage Wednesday did so at 7.26 percent, partly because the bond market is now absorbing several hundred billion dollars a year of supply from this buildout. Each of those people is a creditor to the project. None of them has a place setting.

I am not proposing that Treasury seat a retiree from Toledo next to Jensen Huang. I am proposing something much smaller and entirely achievable, and I would like readers to tell me why it is naive.

Publish the asks. Not the guest list — we already have that. Before dessert, each corporate attendee files a one-page statement of the specific regulatory change they requested: which control, which threshold, which licence, which country. It goes on a federal website within twenty-four hours. No arguments, no justification, just the ask.

This is less disclosure than a lobbyist filing already requires, which is the point. We have built an elaborate transparency regime for the person who takes a staffer to lunch and no regime at all for the six people who have the president's attention over the entrée. The gap is not a loophole. It is the main road, and we paved it.

Write to me if you think the dinner is fine. I will print the best version of that argument, because I am not certain I am right, only that nobody is currently required to tell us what was asked for.

In a piece about Venice screening *Be Brave* — the Renzo Rosso documentary that uses generative AI to recreate its subject's early life — a Film Inquiry critic wrote this, in parentheses, about the film under discussion:

"I have not seen the film, nor will I ever see it. Film critics should boycott watching and reviewing films made with generative AI."

I have sympathy for the fear underneath that sentence and none at all for the sentence.

Start with the tactical problem. A boycott is leverage, and leverage requires that your absence cost someone something. When the Writers Guild strikes, production stops. When critics decline to attend, the film opens anyway, the festival programmes it anyway — Venice ran *Be Brave* on September 10 and opened with Danny Boyle's *Ink*, which also uses generative AI — and the only thing that changes is that the discourse about these films is now conducted exclusively by people with a financial interest in them. Venice also hosted Sonilo, an AI sound and music company, and the Reply AI Film Festival screened roughly sixty synthetic shorts nearby. The industry is not waiting for our permission. It is noting our absence and booking the panel anyway.

Then the deeper problem, which is about what criticism is for.

The only authority a critic has ever possessed is the claim to have looked. Not taste — taste is contestable and should be. Not influence — that left with the newspaper ad budget. Just the unglamorous fact of having sat through the thing and being willing to describe it accurately. Surrender that and you are a person with opinions about a film you have not seen, which describes most of the internet and none of a job.

And look what the refusal costs in this specific case. Boyle went to Toronto and defended the AI in *Ink* as small and responsibly used. He did not say which departments. He did not say which processes. That evasion is the actual story of AI in film right now — not that it is being used, but that nobody will specify where — and you cannot report it from outside the cinema. You catch it by watching the film, noticing the shot that moves wrong, and asking the question with the timecode in your hand. A boycott hands that work to nobody.

Here is what I would rather we demanded, and it is more aggressive than the boycott, not less.

A disclosure line in the credits, by department. Not "AI tools were used in this production," which is a legal shrug. Specifics: environments, de-aging, crowd replication, dialogue replacement, score, foley. If the film declines to disclose, the review opens with that refusal and treats it as information about the filmmakers, because it is. And when a film does disclose and the work is good, we say the work is good, because otherwise the standard is not a standard, it is a mood.

The boycott protects the critic's conscience. The credit line protects the audience. Only one of those is the job.

I expect letters. Send them.