Every story we have run this week has been about a valve. A strait you can close. A pipeline you can strike. A refinery you can burn to the crankshaft and, in doing so, take diesel off the world market until the end of the month. The lesson of July 2026 has been that whoever owns the physical bottleneck owns the price of everything downstream, and that the bottleneck can always, eventually, be reopened. Fix the compressor. Reflag the tanker. Wait out the fire.

Sunday breaks that pattern.

On July 27, according to technical documentation reviewed by researchers and the company's own timeline, the Beijing startup Moonshot AI intends to publish the complete weights of Kimi K3 — a 2.8-trillion-parameter model it unveiled on July 16 and which, on independent read-throughs of its benchmarks, lands second in the world, behind only Anthropic's Claude Fable 5 and, on some tests, OpenAI's GPT-5.6 Sol. It is, by parameter count, roughly seventy-five percent larger than DeepSeek's V4. It carries a million-token context window and native vision. And unlike a tanker or a turbine, once its weights are sitting on a drive in Frankfurt or Fremont, nobody can take them back.

That is the whole story, and it is worth saying plainly because the industry keeps trying to make it complicated. For three years the American strategy toward Chinese AI has been a chokepoint strategy. Export controls on the most advanced chips. Restrictions on the fabrication tools that make the chips. The theory was Hormuz applied to silicon: hold the narrow passage, and you meter what gets through. That theory assumed the valuable thing was the compute. What Moonshot is about to demonstrate is that the valuable thing was always the file — and a file, once released under a permissive license, obeys no strait.

Read the specifications carefully and the release is a piece of strategy as much as engineering. Moonshot is younger and thinner-resourced than the American labs; three days after K3 went live on kimi.com it stopped taking new subscribers because demand had overrun its own hardware. It does not have the compute to serve the world. So it is doing the one thing that turns that weakness into leverage: it is giving the model away, and thereby renting every other company's data center as its distribution network. Databricks, the hyperscalers, anyone with idle GPUs becomes, on Sunday, part of Kimi's delivery system. The AP put it bluntly — the California titans are being made to sweat by the startups that publish rather than hoard.

There is a caveat that matters, and we will not bury it. Until Sunday, the weights do not exist outside Moonshot. The 2.8-trillion figure, the benchmark scores, the second-place ranking: all of it is, for now, a specification sheet published by the company that stands to gain from believing it. No outside researcher has been able to reproduce a single number. As one analyst noted, July 27 is the day the largest-open-model claim stops being a PDF and becomes something anyone can download and check. It is entirely possible the reproduction disappoints. Newspapers that forget this end up printing press releases with a dateline.

But set the exact ranking aside, because the ranking is not the news. The news is the mechanism. For a decade the governing metaphor of AI policy has been the arms race, and arms races are won by controlling supply. What open weights do is convert the race into something closer to a book — and you cannot un-publish a book that a hundred thousand people have already photocopied. Thinking Machines shipped Inkling. Zhipu shipped GLM-5.2. The gap between the closed frontier and the open one, which used to be measured in a year, is now measured in weeks. Sunday it may close to a matter of a single weekend.

We have spent this week reporting the addresses of physical leverage — the ships, the grids, the pumps. Today we report the one address that has no lock on the door. The people who spent three years building a wall around the compute are about to learn that the thing they were walling off had already been copied to the other side. We will be watching the download counter on Sunday. So, we suspect, will they.

The Federal Reserve was built to fight a fever it can feel — too much money chasing too many goods, cooled by making money more expensive. What it faces going into Tuesday's meeting is not a fever. It is a tax, levied at the pump and the pipeline by people the committee cannot summon to Washington.

On Thursday, front-month Brent settled at $100.69, up 7.04 percent on the day — the benchmark's biggest one-day gain since July 13 and its highest close since May 22. The trigger was Yemen's Houthis claiming attacks on two Saudi tankers in the Red Sea, opening a second front on top of the near-total halt of traffic through the Strait of Hormuz. Brent is now up nearly 40 percent since the Iran war began in February, with almost the entire move arriving this month. West Texas Intermediate closed at $92.19.

Here is the arithmetic the committee cannot escape. Chair Kevin Warsh's Fed has held the funds rate at 3.50 to 3.75 percent all year, and its own July Monetary Policy Report concedes inflation "remains elevated" in part because of "supply shocks that have driven price increases in certain sectors, including energy." Governor Lisa Cook put the headline number at 3.7 percent, nearly double the 2 percent target. A rate hike does nothing to the price of a barrel that is expensive because a strait is closed. It can only strangle the demand side of an economy already paying an energy surcharge — which is to say, it can turn a supply shock into a recession while barely touching the price that started it.

The market has decided the Fed will do it anyway. CME's FedWatch now prices a September hike at roughly 82 percent, up from below 53 a week ago; Kalshi traders moved from about 30 to 48 percent over the same stretch. Next week's meeting is still expected to hold, but the probability of a July surprise has jumped from under 12 percent to nearly 38. Goldman Sachs is telling clients Brent could top $120 in the fourth quarter and average $100 through 2027 if Hormuz stays disrupted.

The consensus economist forecast, per FactSet, still says no hike this year and a half-point of cuts in 2027. That is the tell. The professionals who model the economy think the Fed should not tighten into an oil shock. The professionals who trade the funds rate think it will anyway, because a central bank that spent a decade rebuilding its inflation credibility cannot be seen shrugging at 3.7 percent while gasoline leads the evening news.

That is the corner. Warsh can protect the real economy or he can protect the institution's reputation, and this month those two mandates point in opposite directions. Whatever the dot plot says Wednesday, the deciding vote is being cast in the Strait of Hormuz, and nobody on the committee holds it.

For most of the year the wars in the Gulf and in Ukraine were narrated as separate emergencies with separate maps. This week they merged into a single line on a trader's screen, because both have arrived at the same tactic: stop attacking the front line and start attacking the machinery that refines the fuel.

In the Strait of Hormuz, the Islamic Revolutionary Guard Corps declared early Thursday that the waterway would remain "completely closed" for the duration of U.S. military operations, warning that no tanker would be allowed to enter or leave and that any vessel transiting without Iranian coordination would meet the same fate as the ships already burning. It was the tenth-plus consecutive night of U.S. airstrikes, and they have not loosened Tehran's grip on a passage that carried about a fifth of the world's traded crude in peacetime. Lloyd's List Intelligence counted just 53 transits in the week through July 20, down 66 percent; tanker and gas-carrier crossings fell to 30 from 90. The interim deal signed in June has collapsed.

Two thousand kilometers north, the same logic runs in reverse. Ukraine's General Staff says its drones have struck sixteen major Russian refineries and fuel terminals since January, disabling more than thirty percent of the country's refining capacity — a campaign President Zelensky calls "fully justified" retaliation, telling Russians they "must feel that their state is waging this war." Moscow's answer, announced July 8 by Deputy Prime Minister Alexander Novak at a meeting chaired by Putin, was a full ban on diesel exports through July 31, plus emergency imports of petroleum products and a decree letting refineries drop to dirtier Euro-3 fuel. Russia is the world's second-largest diesel exporter after the United States. Taking its barrels off the market tightens middle distillates worldwide precisely as the Gulf seizes up.

Stack the two and the picture is stark: the supply side of the global energy market is now being managed not by OPEC quotas or shale drillers but by targeting officers on two continents choosing which compressor station or crude unit to disable next. Bab el-Mandeb and the Red Sea have become a third pressure point, with Goldman warning of further upside if the Suez route also degrades.

The reader in Ohio filling a tank does not experience this as two wars. She experiences it as one number that keeps climbing, set by decisions made in Tehran, Kyiv, and Sana'a, none of whom will ever appear on her ballot.

The number that will decide more House seats than any speech this fall is printed in foot-high digits on a plastic sign at the edge of every American town, and on Thursday AAA said it read $4.09 for a gallon of regular — up fifteen cents in a single week, with most states now above four dollars.

The cause is not mysterious. Crude in the nineties, driven by Hormuz and now the Red Sea, feeds the pump on a two-week delay, and the second half of summer is when it bites. Prices had fallen as low as $3.79 on July 7; they have climbed steadily since as Washington and Tehran trade nightly fire. A year ago the average was $3.14.

But the average lies, because the pain is not evenly distributed and neither is the politics. California drivers are paying $5.50, Hawaii $5.42, Washington above $5.00 — the West Coast hit hardest, analysts note, by its outsized reliance on Middle Eastern imports and years of refinery closures, on top of the nation's highest fuel taxes. Meanwhile Indiana sits at $3.35, Mississippi and Texas near $3.57. Premium is worse everywhere; the national premium average has crossed $4.90, and diesel is within 66 cents of its all-time record.

The political arithmetic is where this stops being a commodity story. A recent Harris poll found 95 percent of Americans believe the country is in an affordability crisis. A CNN survey in May found 77 percent — including a majority of Republicans — blaming the president for rising costs. The White House keeps insisting prices will fall as the military degrades Iran's ability to hit shipping in the strait; ten-plus nights of strikes have not reopened it, and the president has himself suggested prices may stay elevated. The June memorandum that briefly cracked the market lower is dead.

Here is the cruelty of the map for the party in power. The cheapest gas in America sits in states it already holds. The most expensive sits on a coast it has largely written off. The seats that flip in November — the exurban Pennsylvania and Michigan and Arizona districts averaging $4.15 to $4.27 — are precisely where a fifteen-cent weekly jump reads as a verdict on competence. No amount of degrade-and-defend messaging survives contact with a $70 fill-up, and every driver in a swing district is being handed that receipt twice a week.

Ask what an AI model costs and you used to get a benchmark score. Ask now and you get a rate card, because the industry has quietly changed what it is selling. In a matter of weeks this summer, Anthropic shipped Claude Fable 5, OpenAI rolled out its GPT-5.6 family — Sol, Terra, Luna — xAI released Grok 4.5, and Meta unveiled Muse Spark 1.1 alongside a metered Model API. The trade press has a name for the shift: the model is becoming a runtime, the chat window a control plane, and "answering" is giving way to "executing."

What that means in a boardroom is that competition has stopped being a contest of intelligence and become a contest of unit economics. Grok 4.5 arrived with aggressive pricing aimed squarely at the agentic-work layer. Meta wants to sell metered intelligence outright. And into that price war walks Kimi K3, quoting roughly 30 cents per million input tokens with a cache hit and $15 per million output including reasoning — figures that undercut the American frontier by an order of magnitude. Anthropic's Fable tier runs many multiples higher.

The strategic logic is the same one reshaping every commoditizing market: when the product becomes fungible, the winner is whoever delivers the cheapest reliable unit of completed work. Benchmarks measure whether a model can do a task. Rate cards measure whether you can afford to run it ten thousand times an hour inside an autonomous agent loop, which is where the actual money now lives. On that axis a model that is 95 percent as capable at 10 percent of the cost does not lose. It wins the workloads that matter and leaves the flagship labs the shrinking premium tier.

This is why the closed labs are staging access, locking provider-model-version pairs, and racing toward computer-use and orchestration features that a raw open-weight file cannot easily replicate. Capability alone no longer defends a price. The question every enterprise is now asking — which model wins for my actual workload, not which model tops the leaderboard — is the question the entire market is reorganizing around, and it is a question that rewards the cheap and the good over the merely spectacular.

A nearly three-hour, R-rated adaptation of a 2,700-year-old poem opened last weekend to $123.5 million domestic and $263.7 million worldwide — Christopher Nolan's biggest global launch ever, and the year's largest live-action opening. By Wednesday the six-day cume stood at $181.6 million, running 54 percent ahead of where Oppenheimer sat at the same point. IMAX and premium large-format screens accounted for 52 percent of the opening weekend; the 41 theaters worldwide running true IMAX 70mm film are sold out for weeks. Presales for the second weekend, one exhibitor says, rival some films' entire opening.

Every studio executive who watched this and concluded "sign more auteurs" misread the mechanism, and the misreading is the most interesting thing in Hollywood right now. The Odyssey did not work because it was expensive or because Nolan is a brand. It worked because Universal manufactured scarcity that streaming physically cannot replicate — a 100-day theatrical window, a format that does not exist on any couch, an experience engineered to be unavailable at home until roughly November. The scarcity is the product. Flood the market with $250 million director-driven event films and you destroy the very thing you are trying to sell.

The sharper analysts have started calling it the barbell. At one end, cheap pre-validated internet material: "Backrooms" turned about $10 million into $350 million, and studios now pay seven figures for horror memes with no script attached, because the audience already exists for free. At the other end, the $250 million auteur event that only works because it cannot be watched properly anywhere else. Both ends solve the same problem in opposite directions — one buys an audience that already exists, the other manufactures scarcity that streaming can't touch. What dies is the comfortable middle: the $150-to-$200 million franchise sequel with a work-for-hire director, a familiar logo, and no reason on earth to leave the couch.

The domestic box office is on track to clear $10 billion this year for the first time since the pandemic. Read that alongside a Fed cornered by oil and a lab about to give away a two-trillion-parameter model, and the through-line of this whole issue comes into focus. Everything that can be copied is being copied, and its price is collapsing toward the cost of a download. What holds its value is what cannot be reproduced — a strait, a refinery, a 70mm print running at 24 frames a second in a room you had to drive to. The couch has almost everything now. Homer is betting there is still something worth leaving it for. This week, at least, he is collecting.

There is a particular kind of policy failure that comes from winning the wrong argument brilliantly. American export controls on advanced AI chips are a masterpiece of that genre. They are precise, aggressive, technically sophisticated, and enforced with genuine will. They are also, as of this Sunday, aimed at the wrong object.

The theory was elegant. Frontier AI needs enormous compute; enormous compute needs the most advanced chips; the most advanced chips come from a handful of fabs using a handful of tools. Control that chain and you control the frontier. It is Hormuz logic applied to silicon — hold the narrow passage and meter what gets through. And for a certain window it worked, or appeared to, because everyone agreed the scarce resource was the training run.

But a trained model is not a training run. It is a file. A very large file, but a file, and files do not respect the physics that make chokepoints work. You cannot mine a download. You cannot board it off the coast of Oman. When Moonshot publishes Kimi K3's weights on July 27 under a permissive license, the most sophisticated export-control regime in history will be standing guard over the factory while the product walks out the front door and copies itself onto a hundred thousand drives, in every jurisdiction, permanently.

I want to be careful about what I am and am not claiming. I am not saying open weights are automatically good, or that a Chinese lab releasing a frontier model is a gift to anyone. There are real proliferation risks in putting a two-trillion-parameter system into every hand — that is precisely the point. What I am saying is that the risk was never controllable by the mechanism we chose, and pretending otherwise wasted three years we could have spent on approaches that survive contact with reality: interpretability, liability regimes, standards for what deployed systems may do, defenses that assume the model is already out because the model is always, eventually, already out.

The chip controls did one thing well. They convinced Chinese labs that hoarding was a losing game and that their best route to influence was to give their models away and rent the world's idle compute as a distribution network. We did not slow the frontier. We changed its business model, in the direction of maximum, irreversible diffusion. That is not the outcome the policy was designed to produce. It is the outcome it produced anyway.

Sunday is the day the specification sheet becomes a fact on a hard drive. When it does, the honest response is not another tightening of the same valve. It is to admit we were guarding a strait while the cargo learned to teleport, and to start building for the world we actually live in — the one where the most powerful tools cannot be recalled, only governed. We should have started three years ago. We can start Monday.

The most dishonest sentence in American politics this month is short and keeps getting repeated: prices will come down as we degrade Iran's ability to strike shipping. It has the shape of a plan and the content of a wish, and the market has stopped believing it. So, increasingly, have voters, who possess an inconvenient talent for reading a number on a sign.

The number is $4.09, up fifteen cents in a week, and it is climbing because a strait is closed and a second front has opened in the Red Sea, and no bombing run yet has reopened either. This is worth stating without partisanship because both parties are about to lie about it in opposite directions. The administration will insist force is working and relief is imminent. The opposition will imply a different president could conjure cheap gas from a war zone by wanting it more. Both are selling the same fantasy: that American policy sets a price being set in the Strait of Hormuz by people who do not take Washington's calls.

Here is the uncomfortable thing I think is true. There is no version of this conflict that is cheap at the pump, and the sooner someone in power says so, the more credibility they keep for when it ends. A supply shock is a supply shock. You can cushion it — release reserves, cut the gas tax, send checks to the households a $70 fill-up actually breaks — or you can pretend military progress is refinery output. The cushioning is unglamorous and finite and admits the war has a domestic cost. The pretending is free until the next AAA report, which arrives every Thursday like a bill.

What makes this politically lethal rather than merely expensive is the map. The cheapest gas in America sits in safe states; the most punishing weekly jumps land in the exurban swing districts that decide control of the House. Ninety-five percent of Americans already tell pollsters they are living through an affordability crisis. You do not talk a voter out of that with a briefing on degraded Iranian capabilities. She is not experiencing a foreign policy. She is experiencing a receipt.

I am not arguing the strait does not matter, or that deterrence is worthless, or that there is a costless exit. I am arguing that the first politician to stop promising a discount on a war and start honestly managing its price will be the one voters trust when the discount does not come. Everyone else is running a tab that gets read aloud at every gas station in the country, twice a week, until November.