There is a deal, and there is a strait, and this week they have almost nothing to do with each other.

The deal is close. The United States, Iran and Oman are in the final stages of a 60-day interim arrangement to reopen the Strait of Hormuz, with negotiators reporting that they have agreed on the geographical coordinates of a new transit corridor: inbound traffic through a northern lane in Iranian waters, outbound through a southern lane in Omani waters, no tolls charged during the temporary period. Iran's deputy foreign minister says the understanding could hold for two to four months. A Gulf official put the odds of a broader U.S.-Iran agreement by Friday at exactly 50:50, which is another way of saying nobody knows.

Here is what the strait itself is doing while the diplomats talk. According to vessel-tracking firm Kpler, more than half of all crossings through Hormuz are now "dark" transits — ships moving with no reliable tracking signal, transponders switched off, running blind through the most surveilled chokepoint on Earth precisely so that no one can watch them do it. Daily tanker traffic has repeatedly hit single digits in each direction. On some days it has hit zero.

That is the story the wires keep burying under the word "reopening." A strait does not reopen because two foreign ministries agree on coordinates. It reopens when a shipowner in Athens or Singapore decides the insurance math works, and right now it does not. On Monday a cargo ship was struck by a projectile in the engine room and crew quarters while transiting northbound through Omani waters; one crew member is still missing, the vessel disabled and adrift. On Wednesday a tanker reported two explosions nine nautical miles off Kumzar. Days earlier, two tankers reported VHF threats from voices claiming to be the "Sepah Navy," warning shots, and an explosion 1.5 nautical miles off the port quarter. The Joint Maritime Information Center describes a pattern along the Limah-Khasab axis: drone overflights, targeted surveillance, projectile strikes, near-misses.

You cannot sign a coordinate agreement out of that. The reason the interim deal leaves the fleet unmoved is the reason the July memorandum before it collapsed: it does not resolve who controls the water. Iran has demanded that ships obtain permission and pay transit fees — a toll booth on the world's oil supply — and both the Gulf states and Washington have rejected it. Oman has floated a voluntary-contribution model borrowed from the Strait of Malacca. Tehran, sources tell *The National*, is now trying to trade toll-free passage for sanctions relief, which reveals the whole game: the attacks were never about the strait. They were about leverage, and leverage does not want the strait to be boringly, verifiably open.

So who is actually transiting? Analysts say the ships that moved during the last truce were overwhelmingly Gulf oil majors and "risk-tolerant independents" — the same profile likely to move under any new deal. The large international owners are waiting for evidence the arrangement can hold, and evidence is the one commodity Hormuz cannot currently manufacture. As one analyst put it, Iran "has the ability to wait this out."

Meanwhile the market has decided the deal is basically done. The Dow closed at a record 54,349 this week and opened higher again Thursday on "Hormuz optimism," oil steady, the VIX asleep at 15.75. Wall Street is pricing a reopened strait. The strait is pricing a war of attrition conducted by transponder.

When those two prices disagree this violently, one of them is wrong. Our house position, as it has been all week: bet on the inventory, not the announcement. The ships in the dark are telling you what they believe. They believe the deal is a rumor and the projectile is a fact, and they have arranged their entire risk posture around that belief. A signed statement out of Muscat will not change it. Sixty consecutive days of an unmolested tanker might.

Until then, the most honest instrument in the whole affair is a switched-off transponder. It is the sound of the real economy refusing to be quoted.

The distance is the news.

Overnight into Thursday, Ukraine struck two Russian oil refineries: the Slavneft-Yanos plant in Yaroslavl, roughly 700 km from the Ukrainian border, and — the one that matters — the Bashneft-Novoil refinery in the republic of Bashkortostan, about 1,300 km away. That is more than 800 miles. That is a Ukrainian drone flying the distance from Washington to Kansas City and finding a specific set of storage tanks at the end of it. President Volodymyr Zelensky called it part of a campaign of "long-range sanctions" aimed at the oil revenue Moscow uses to finance the war. The Yaroslavl governor confirmed the refinery caught fire after falling debris struck its storage tanks; four people were injured by shrapnel.

Russia's own numbers confirm the shift more honestly than Kyiv's do. The Defense Ministry said air defenses shot down 605 Ukrainian drones overnight, including at least five headed toward Moscow itself. You do not intercept 605 of anything unless 605 of them are coming. The deep strike is no longer a stunt; it is the operational tempo.

The cost, as always, landed on civilians facing the wrong direction. Russia's overnight barrage killed at least 11 people across Ukraine. Three died in Balakliya in the Kharkiv region when drones hit homes — two women and a man, per the local military administration. Three more were killed in the northern Sumy region, where twelve people were wounded in the city of Sumy by guided aerial bombs. Five were killed by strikes in the central Dnipropetrovsk region. In the Black Sea port of Odesa, one person died when Russia hit two dry cargo ships, one of them carrying wheat. The entire city of Kherson lost electricity after a critical-infrastructure facility was struck.

Zelensky also said Ukraine hit "two military patrol boats and shadow fleet vessels" in the Black Sea — the tankers Russia uses to move sanctioned oil past the blockade. That detail rhymes with everything else on today's front page: the war, the markets, and the Strait of Hormuz are all now fights over inventory that has learned to hide. Russia's dark tankers, Iran's dark tankers, and Ukraine's willingness to reach 800 miles to set the loading terminals on fire are the same story told in three theaters.

What changed in the last sixty minutes is not the diplomacy. It is the map of what can be hit. When a refinery 1,300 km from the front is a legitimate overnight target, there is no rear anymore. There is only range, and Ukraine keeps buying more of it.

A truce is a claim about the future. On Thursday, southern Lebanon tested the claim.

Two Israeli soldiers were killed in an explosion in the village of Majdal Zoun, Israel's military said — the first Israeli deaths since the shaky ceasefire with Hezbollah began nearly two months ago. The Israeli army had launched strikes the day before in response to what it called a Hezbollah violation of the ceasefire, and warned residents of the village of Mansouri to flee, the first such evacuation order Israel had issued inside Lebanon in weeks.

The Lebanese ledger from the same 24 hours: the country's Health Ministry said one person was killed and 12 wounded in an Israeli strike on the town of Tebnine on Wednesday, with another eight wounded in overnight strikes on Burj al-Shemali. Artillery also shelled Mansouri and Majdal Zoun, according to the state National News Agency.

The current Israel-Hezbollah war erupted on March 2, when Hezbollah fired rockets across the border two days after the U.S. and Israel launched their own attacks. The June truce paused that exchange. It did not end it. The pattern of the last two days — an Israeli strike citing a violation, an evacuation warning, retaliation, then dead soldiers — is what a ceasefire looks like when neither side has actually agreed on what "ceasefire" means. Same lesson as Hormuz, printed in a different alphabet: the signed word and the shelled coordinate keep failing to match.

Read the fundamentals cold and you would guess these were the best two reports of the quarter.

SanDisk posted fiscal Q4 revenue of $8.97 billion against a $8.4 billion estimate, non-GAAP EPS of $39.25 versus $34.37 expected, full-year sales of $20.2 billion up from $7.4 billion a year ago, and data-center revenue up more than 400% for the year and roughly doubling sequentially to $2.98 billion. CEO David Goeckeler said the company now has over four years of demand visibility and has signed five new long-term deals worth $94 billion in minimum revenue. Western Digital delivered a double beat: revenue of $3.75 billion, up 44% year over year, with gross margin surging past 54%.

The market's verdict: SanDisk fell about 13% to $1,178. Western Digital shed as much as 19% to around $420, its steepest one-day drop since April 2025. Micron slid 6%, Seagate 6%, SK Hynix down with the group.

So what went wrong? Nothing, which is the whole problem. SanDisk guided next-quarter revenue to $10.3–$10.8 billion; the midpoint landed below the roughly $11.2 billion analysts wanted. That is the entire crime. Yahoo Finance's Brian Sozzi called the selloff "absurd" given the durability of demand — and he is right about the demand and wrong about the absurdity. This is not irrationality. This is what happens to a stock that has risen more than 3,000% in twelve months, as SanDisk has, or more than 500%, as Western Digital has. At those prices, the guidance is not a forecast; it is a hurdle, and "merely excellent" clears it by going under.

Here is the mechanism worth understanding, because it will repeat. A stock priced for perfection converts good news into bad news. Every beat that is not a bigger beat is a deceleration signal. Every four-year backlog is re-read as "the growth is normalizing." The bull case — multi-year AI storage demand, contracted revenue, expanding data-center mix — is entirely intact. The bear case is narrower and, today, decisive: margins may be peaking, valuations are stretched, and sequential guidance is starting to decelerate off an impossible base.

Watch the rotation underneath. Gold has climbed more than 7% in a few days; Bitcoin is holding above $64,000. When the year's biggest AI winners get sold on strong numbers, capital does not vanish — it goes looking for the next thing that has not yet been priced for perfection. The memory trade is not broken. It is just, for the first time in a year, being asked to be worth what it costs.

It is rare to watch a policy get refunded and reissued in the same news cycle. This week you can.

The refund first. After the Supreme Court ruled 6-3 in February that the administration exceeded its authority by invoking the 1977 International Emergency Economic Powers Act to impose "reciprocal" tariffs on more than 90 countries, Customs and Border Protection disclosed in a court filing that it has now completed and disbursed about $100 billion in refunds — roughly 60% of the $166 billion collected under the struck-down regime. Nearly $129 billion in "potential and certified refunds" has been accepted for processing. Importers, in the phrasing of one outlet, struck gold.

Now the reissue. The ruling left untouched the sector-specific tariffs imposed under the 1962 Trade Expansion Act — steel, autos, copper — and the administration has spent the months since methodically rebuilding the rest on new legal footing. Last month it announced 10–12.5% levies on dozens of countries it accuses of insufficient action on forced labor. As soon as Thursday, the White House is expected to unveil a 15% tariff plus minimum import prices on polysilicon and its derivatives — wafers, solar cells, finished modules — grown out of a Section 232 national-security probe. New country-specific reciprocal tariffs of 10–41% are set to take effect August 7.

The legal resistance is organizing. On Monday, 25 states — including New York, California, Arizona and Colorado — filed a challenge calling the latest measures a pretext to re-impose exactly the duties the Court already killed. The Consumer Technology Association has warned the polysilicon rule is unworkable in practice: once the material becomes a wafer inside a chip, there is no way to trace where the original silicon came from.

Strip the politics and the shape is clean. A tariff struck down is not a tariff removed; it is a tariff that has to find a new statute. The $100 billion going back out the door is the cost of the first attempt. The polysilicon proclamation is the second. The question the states are asking the courts is whether you are allowed to lose a case and keep the outcome.

The most important AI story this week is not a model. It is a retreat, a land grab, and a guarantee — all on the same patch of Texas dirt.

Start with the retreat. Oracle and OpenAI have canceled plans to expand their flagship Stargate campus in Abilene, Texas, the first Stargate site to go live from the "$500 billion" joint venture announced last year. Two buildings launched last September, with six more due this year for roughly 1.2GW of capacity; the plan had been to grow it to 2GW. That expansion is dead. Bloomberg reports it was complicated by financing challenges and OpenAI's "frequently changing demand forecasting." Power at the site won't be ready for a year anyway — by which point OpenAI would rather deploy Nvidia's next-generation Vera Rubin chips at a new campus than the Blackwell GPUs going into Abilene.

Then the land grab. Into the space OpenAI vacated stepped Microsoft, which signed for 900MW of capacity next door — two new buildings and an on-site power plant — bringing the site's projected total to 2.1GW and making the two AI rivals literal neighbors. In an unusual defensive move, Nvidia reportedly paid a $150 million deposit to developer Crusoe to secure remaining capacity and keep a rival chip designer's products out of the campus, while shopping the space to Meta.

And then the guarantee, which is the part that should make you sit up. OpenAI is in talks with Nvidia for a backstop of up to $250 billion — Nvidia's credit standing behind the lease and construction debt for a 10-gigawatt data center in Pike County, Ohio, being developed by SoftBank and SB Energy with the Department of Energy. The chips inside are a separate negotiation. Recall that Nvidia's previously announced $100 billion investment in OpenAI never fully materialized; it contributed $30 billion to OpenAI's March funding round instead.

Put the three moves in a line and the shape is unmistakable. The company that was going to build everything, everywhere, could not finance the expansion of its own flagship, ceded the ground to its largest backer-turned-competitor, and is now asking its chief supplier to co-sign a quarter-trillion in debt so it can build somewhere else. This is not the story of AI running out of demand — data-center revenue is exploding across the sector. It is the story of AI running into the balance sheet. Compute is infinite. Credit is not.

Everyone is quoting the dollar figure, which is the least interesting number in the story.

Yes: *Spider-Man: Brand New Day* opened to $360 million domestic in final numbers, dethroning *Avengers: Endgame*'s $357.1 million to claim the biggest opening weekend in North American history. It drove the entire market to a record $430 million three-day frame — the biggest domestic weekend ever for all films — and pushed 2026's year-to-date box office past $6.19 billion, up more than 15% from last year. Sony's Tom Rothman, allowed one victory lap, took it: "records are made to be broken, but we will take great pride in this one."

But dollars are inflated by ticket prices, premium formats, that $63,000-per-screen Dolby average. The number that actually means something is the one from EntTelligence: 24 million admissions in a single weekend — the biggest opening-weekend turnout of the post-Covid era, ahead of *No Way Home*'s 20 million. Twenty-four million individual human beings decided, on the same three days, to leave the house, drive somewhere, buy a ticket, and sit in a dark room with strangers.

That is the product streaming cannot manufacture. Consider the backdrop: Netflix stock is sitting near 12-month lows around $71, and the read across the industry is that streaming has become *commonplace* rather than dominant — the thing that's always on, not the thing you plan your Saturday around. Sony is reopening the Cinerama Dome and the Hollywood Arclight in 2028. A studio does not restore cathedral cinemas because the model is dying.

Here is the taste argument underneath the ledger. The theatrical event is not competing with your television. It is competing with your inertia, and this weekend it won by 24 million. The thing people will pay a premium for, still, is the *occasion* — the shared, un-pausable, phone-in-your-pocket hour that a queue of algorithmic thumbnails structurally cannot produce. You can stream the movie in six weeks. You cannot stream the crowd. What Sony sold this weekend was not Spider-Man. It was the crowd, and it turns out the crowd is the one thing that never went on sale.

We keep describing the dark fleet as a loophole. It is not a loophole. It is the shape of the world now, and the lit, transponder-on, fully-insured economy is the shrinking exception we are nostalgic for.

Consider what today's front page actually shows. In the Strait of Hormuz, more than half of all crossings now go dark — transponders off, tracking dead — because visibility has become a liability. In the Black Sea, Russia moves sanctioned crude through a "shadow fleet" of tankers with scrambled ownership and forged papers, and Ukraine now reaches 800 miles into Bashkortostan to burn the terminals that load them. Iran runs its own ghost tankers past a U.S. blockade. Three theaters, one behavior: the highest-value transactions on Earth are increasingly conducted specifically so that they cannot be observed, priced, or governed.

We built the entire architecture of modern trade on the opposite assumption. Insurance, letters of credit, sanctions regimes, futures markets — all of it presumes a visible transaction. A ship you can see. A cargo you can price. A counterparty you can name and, if necessary, punish. The dark transit breaks that presumption at the root. And once a critical mass of the trade goes dark, the visible market stops being the market and becomes a kind of press release — a well-lit stage where the official numbers are announced while the real volumes move offstage in vessels nobody is allowed to watch.

This is why the Hormuz "deal" is theater and the switched-off transponder is truth. The negotiators are haggling over coordinates for a corridor the serious tonnage has already decided to run blind through regardless. The agreement governs the lit economy. The lit economy is no longer where the risk lives.

I am not romanticizing the smugglers. The dark fleet is uninsured, unaccountable, and periodically on fire with crew members missing. That is the point. We are watching the connective tissue of global trade migrate into a domain with no visibility, no recourse, and no referee — and calling each individual instance an evasion, when the aggregate is a regime change.

The governable transaction was a mid-century luxury, propped up by one navy that could see everything and one currency everyone had to touch. Both of those props are wobbling, and the ships have noticed before the ministries have. When your most sophisticated actors — oil majors, state shippers, sanctioned exporters alike — all independently conclude that the smart move is to disappear from the radar, they are not cheating the system. They are telling you the system's central assumption has expired.

The front page will keep being written in declarations. But the century is being written in the dark, one unlogged transit at a time. The sooner we stop treating that as an anomaly and start treating it as the baseline, the sooner our institutions might build something that works in the dark — instead of announcing, over and over, a daylight that the ships have already left.

Say the deal out loud, slowly, and listen for the moment it stops making sense.

Nvidia is in talks to guarantee up to $250 billion in debt so that OpenAI can lease a 10-gigawatt data center. The data center exists to run Nvidia chips. OpenAI is one of Nvidia's largest customers. So the supplier is standing behind a quarter-trillion in borrowing so that its own customer can afford to buy — and house, and power — the supplier's own product. The chips themselves are a *separate* negotiation, which means Nvidia's credit is backstopping the building that will hold the thing Nvidia sells to the company Nvidia is financing.

That is not a supply chain. It is an ouroboros with a balance sheet.

I want to be precise about what is and isn't alarming here, because "circular financing" gets thrown around lazily. There is nothing wrong with vendor financing per se; the world runs on it. What is unusual is the *concentration* and the *direction of the guarantee*. Normally a supplier extends credit and eats a manageable slice of risk if the customer stumbles. Here the supplier is being asked to underwrite the customer's entire capital structure — the lease, the construction debt — on a scale that rivals the GDP of a mid-sized nation, for a single project, on top of a $100 billion investment pledge from last year that, notably, *never fully materialized.* Nvidia put in $30 billion of a promised $100 billion. Now the ask is $250 billion in guarantees. The numbers are not getting more grounded. They are getting more theological.

And look at what is happening simultaneously in the physical world: OpenAI just *canceled* the expansion of its flagship Abilene campus because it couldn't finance it and its demand forecasts kept moving. The company that cannot reliably predict how much compute it needs is the company whose debt Nvidia is being asked to guarantee at nine zeros. The retreat in Texas and the guarantee in Ohio are not two stories. They are the same balance sheet, seen from two ends.

Here is the uncomfortable structural fact. When the dominant chipmaker's revenue, its customers' solvency, and the debt markets funding the buildout all rest on the same handful of names guaranteeing each other's paper, you have not diversified the AI economy. You have wired it in series. Pull one bulb and the whole string goes dark. The bull case for AI has always been that demand is real, and it is — data-center revenue is up triple digits across the sector. But real demand financed in a circle is still financed in a circle, and circles have a specific failure mode: they look self-sustaining right up until the moment they don't.

I am not predicting a crash. I am pointing at the wiring diagram and noting that we have stopped drawing it, because if we drew it honestly it would embarrass everyone in it. A market has independent buyers and sellers who can each walk away. What Nvidia and OpenAI are building is something else — a closed loop of mutual dependence dressed in the language of a deal. Call it what it is. Not a market. A circle. And ask the only question that matters about a circle this large: what, exactly, is holding it up from the outside?