Diesel crossed $6, the Houthis closed the other chokepoint, and the European Central Bank raised rates into a bond rout

The Strait of Hormuz and the Bab el-Mandeb are not substitutes for each other and they are not independent either. Hormuz is the only sea exit from the Gulf: there is no way around it, which is why a threat there prices crude directly. The Bab el-Mandeb is the southern gate of the Red Sea, and it does have an alternative — the long way round the Cape of Good Hope. That alternative is what made the 2024 Red Sea crisis a freight-cost event rather than a supply event. Cargo still arrived; it arrived later and dearer.

What makes this week different is that the two are being squeezed simultaneously by actors on the same side. A cargo leaving the Gulf has to pass Hormuz; if it is bound for Europe it then faces a choice between the Red Sea and an extra two to three weeks around Africa. Pressure on both raises the cost of the same voyage twice. And the product most exposed is the one already setting records: roughly 3.6 million barrels a day of refined products move north through the Red Sea toward Europe, with another 2.6 million heading the other way to Asia. Diesel is the marginal barrel in that flow.

The second-order effect to watch is the Suez Canal, which has an odd position here. Suez transits have recently run at their highest since the start of 2024, partly because Saudi traffic has been rerouted northbound. Egypt's canal revenue depends on the Red Sea being usable. If Perim changes underwriters' minds, Egypt loses income at the same time its food import bill rises with freight. That is a fiscal problem for a country the United States underwrites, and it is the channel by which this reaches Washington as something other than a naval question.

On the military side, the significant detail is not the island but the withdrawal. Government forces left; they were not defeated on Perim. A coastal advance that proceeds by other people declining to fight is fast and cheap, and it is also reversible in principle and very hard to reverse in practice, because the force that would have to do the reversing is the one that just walked away. Washington has threatened action if the Houthis block shipping, but it is fighting Iran with the same ships and aircraft, and a second front on the Red Sea coast is a capacity question before it is a policy one.

Start with why the 30-year auction and the buyback can disagree without either being wrong. An auction sells new paper at whatever yield clears; a buyback asks holders to hand back old paper at a price Treasury sets by accepting or rejecting offers. These are different markets in the same instrument. New thirty-year bonds are the most liquid point on the long end and the natural home for a foreign reserve manager who wants duration at 5.3 per cent; 79.5 per cent to indirect bidders is that buyer showing up. Off-the-run bonds maturing between 2037 and 2046 are the illiquid middle of the curve, held by people who want out and will only leave at a price. Strong demand for the former and a disagreement about price on the latter is an entirely coherent picture, and it is close to the narrow reading Treasury itself has offered.

The dealer number deserves a moment. Primary dealers took 2.2 per cent of the auction against an 11.5 per cent average. Dealers are the residual buyer: they absorb what nobody else wants, so a tiny dealer share normally means genuine end demand. But it also means dealer balance sheets were not expanded by this auction, and dealer balance-sheet capacity is exactly what the buyback programme exists to relieve. So the two results are consistent in a less comfortable way than the headline bid-to-cover suggests: investors wanted new long bonds, and nobody wanted to warehouse old ones.

Now the ECB, which is the part a US-only reading misses. Europe is raising rates because the same oil shock driving American diesel is driving euro-area inflation, and Lagarde said the Governing Council did not discuss where rates go next. Unanimity plus no forward guidance is about as hawkish as that institution gets. For an American holder of Treasuries the consequence is mechanical: as bund yields rise, the hedged return on buying US duration instead of European duration falls, and the marginal foreign buyer of the long end gets less marginal. That channel does not show up in any domestic indicator. It shows up as a 10-year at 4.95 per cent on a day with no American news in it.

What this does to the fiscal arithmetic is the part worth holding. CBO has net interest at about $1.0 trillion in fiscal 2026, rising to $2.1 trillion by 2036, with the 2026 deficit near $1.9 trillion and debt held by the public passing 101 per cent of GDP this year. Those projections are built on an interest-rate path. Every basis point of permanent increase at the long end raises the cost of the refinancing that has to happen regardless of what Congress does, and Treasury's response so far — buying in long bonds with cash raised in bills — shortens the average maturity of the debt into a meeting where the Fed may raise the front end. Protecting the long end has a price, and the price is paid in the bill market.

One date: 4 November, the quarterly refunding, when the $4 billion long-end floor set on 19 August is reset, raised or dropped. After yesterday, that decision is more informative than anything said between now and then.

The American policy apparatus for stopping adversaries from getting advanced computing is built around hardware. Export controls restrict chips, the bill fights in Congress are about Blackwell-class parts, and the enforcement model assumes the thing being controlled is physical, serialised and crosses a border. The misuse half of this report describes something that does not fit that model at all: a weapons cell in Yemen obtaining frontier engineering labour through an API, paying retail, with no hardware in the country. You cannot put an export licence on that transaction in the way the existing statutes contemplate. The control point is a terms-of-service enforcement decision made by a private company, which is precisely the arrangement the administration's AI framework says is adequate and which its critics say is not.

The distillation half cuts the other way and is politically more convenient, which is a reason to be careful with it. If Chinese labs are training on frontier American model outputs at the scale alleged — 151 million exchanges from one campaign — then the chip controls are leaking at the output layer, and the obvious policy response is to restrict who may access American models, not merely who may buy American chips. That is the gated-capability argument this brief has had an open thread on since 5 September. It would also be a significant commercial win for US labs, which is exactly why a claim of this kind deserves a neutral adjudicator before it becomes a premise of policy.

Three weeks of calendar matter here. Xi Jinping is reported due at the White House on 24 September, and Treasury has floated AI safety talks whose date has never been published. A report alleging that three named Chinese firms systematically exfiltrated the reasoning traces of American models, landing two weeks before that meeting, is now part of its furniture whether or not anyone intended it to be. Any Chinese response will be read against it, and any American concession will be read as made in spite of it.

The cross-connection within this issue is worth stating plainly, because it is the kind a beat-by-beat reading misses. The same week a report says a Houthi-linked cell used an American AI model to write missile guidance software, the Houthis took the island in the middle of the Bab el-Mandeb. Those are not causally linked on any evidence available — the report says no fielded weapon was found, and seizing an island does not require guided missiles. But they are the same organisation acquiring two different kinds of reach in the same week, and the policy questions they raise — who may buy frontier engineering, and who can shut a strait — are going to be answered by the same government at the same time.

One methodological note this brief owes its reader. The company's prior disclosure, covered here on 10 September, rested on a scan of 481 million transcripts whose second-stage adjudication was performed by a model from the same family as the ones under investigation. The same caution applies to this report's negative findings: a claim that nothing worse exists is only as strong as the filter that looked.

The reason an auditor registry is more consequential than its dull name suggests is that it answers a question nobody else has answered: who is allowed to say an AI system complies with anything. Every AI statute now on the books, state or foreign, eventually requires somebody to assess a system against a standard. In practice that work is done by a handful of nonprofits and consultancies operating on contract, with no accreditation, no conflict-of-interest rules and no licence to lose. AB 1405 creates the licence. Once a state can strike an auditor from a register, the auditor's incentive shifts from pleasing the client to keeping the registration, which is the mechanism that makes financial auditing work at all — imperfectly, but it is the mechanism.

That is also why the Anthropic connection in the adjacent story above is not a coincidence worth glossing over. That company has just contracted METR, an independent evaluation nonprofit, for an eight-week investigation with access beyond an incident window. SB 813 would make that kind of arrangement a recognised institutional category rather than a voluntary gesture by one firm. A lab that has already chosen to buy an outside audit has an obvious interest in everyone being required to, and an equally obvious interest in the standard being one it already meets. Neither of those makes the law worse. It does mean the law's supporters and its subjects are the same people, and a reader should not need a disclosure line to notice that.

The preemption collision is the live federal question. Congress declined to preempt state AI law twice — in the reconciliation bill and again in the NDAA. The administration's response was an executive order in December 2025 creating a DOJ AI Litigation Task Force to sue states, and a policy framework in March urging Congress to try again. As of 1 July, states had enacted 109 AI statutes and 28 data-centre statutes this year. Until a court rules, state law is the law, and California's is the operative one because no company builds a separate product for one state. What is novel about the audit laws as preemption targets is that they regulate auditors rather than AI developers, which makes the usual dormant-commerce and field-preemption arguments harder to run: a state licensing professionals inside its own borders is close to the core of what states do.

On the child-safety package, the detail with the longest reach is SB 867, on chatbots in toys. Companion-AI regulation so far has assumed a screen and an account. A toy has neither, has no login to attach parental controls to, and is bought by someone who will never read a model card. If that statute's obligations are workable it will become the template; if they are not, it will be the first test of whether this legislative approach can reach hardware at all.

The grace period is young. It was created in 2017, in the rule on retention of employment-based immigrant workers, and it codified a practice that had until then been a matter of case-by-case discretion. DHS's framing is therefore accurate as history: the prior state of affairs was no guaranteed window. What that framing leaves out is why the window was created. Before 2017, a worker laid off was in an incoherent position — technically out of status the moment employment ended, but with a pending or possible new petition, and dependent on an adjudicator's forbearance. The 60 days replaced discretion with a rule, which is generally considered an improvement in administrative law.

The practical mechanics are where the consequence sits, and they are unforgiving. H-1B portability lets a new employer file and the worker begin work on receipt of the filing, but somebody has to file, and that takes a job offer, an attorney, a labour condition application certified by the Department of Labor, and a fee. Two months is a tight timetable for that sequence in a soft hiring market. Zero days is not a timetable at all: it makes the new petition necessarily late, which means a change of status from abroad, consular processing, and a gap in employment that many employers will simply decline to bridge.

The steeper edge is unlawful presence. Status and presence are different legal facts with different consequences. Falling out of status can often be cured; accruing unlawful presence triggers the three- and ten-year bars on re-entry under INA 212(a)(9)(B), and those are not discretionary. A grace period keeps a laid-off worker on the status side of that line while they sort it out. Removing it moves the line to the last day of employment, which means an involuntary event — a reduction in force someone learns about on a Tuesday — starts a clock with a decade-long penalty attached.

Who this lands on is not evenly distributed. L-1 holders are intracompany transferees whose status is tied to one employer by definition and who have no portability to use. TN is a USMCA category, which raises a separate question about whether removing a benefit from Canadian and Mexican professionals interacts with the trade agreement at a moment when the United States is also banning Canadian imports. O-1 and E categories include people whose work is structurally intermittent. The population most often discussed — H-1B holders at large technology employers — is the one best equipped to absorb it.

I have marked this underreported because it is. It appeared this morning in the Federal Register and has so far been written up almost entirely by immigration law firms, which is the usual pattern for a rule whose subjects cannot vote. Comment closes in sixty days, which is the ordinary period and is itself the thing to notice: an agency that wanted this defensible gave the full window.

A tariff and a prohibition are not points on one scale. A 50 per cent duty is a price: some trade continues, importers and exporters split the cost, and the flow adjusts. A ban sets the quantity to zero, and zero has effects a price does not. Supply chains with a Canadian sole-source input do not get more expensive, they stop. Contracts become impossible to perform rather than unprofitable, which moves the dispute from commercial negotiation to force majeure. And the distributional consequence inverts: with a tariff, the US Treasury collects revenue; with a ban it collects nothing, so the policy costs the government money while costing importers more.

The product selection repays reading because it is not a list of what Americans buy most. Whey protein concentrate is an industrial input to protein powders, infant formula and processed food, not a consumer good, and Canada is a significant supplier. Motorcycles over 800cc is a category with one obvious political referent. Non-alcoholic beer appearing on the dairy list is a tariff-schedule artefact rather than a statement about beer. And the additions effective Monday — aluminium profiles, furniture, lighting, golf carts, fishing rod parts — are the kind of list that catches firms which checked whether they sell cars, wine or cheese, concluded they did not, and stopped reading. The annexes specify eight-digit HTS codes, so exposure is determined by classification, not by industry.

On USMCA, the point is simple and often got wrong. The agreement is not self-executing in US law. 19 U.S.C. 4512(a)(1) provides that no provision of the agreement inconsistent with US law has effect, which means a statute like Section 338 beats a treaty obligation in an American court. Canada's remedy is therefore not litigation in the United States but a Chapter 31 dispute panel or a WTO case, both of which take years and neither of which can stop goods being turned away on 29 September. Whether these measures qualify as emergency measures exempt from those obligations is exactly the kind of question those panels exist to decide slowly.

One piece of unfinished business from this brief's own record. It has twice reported an effective date of 19 August for the original three proclamations, from Holland & Knight quoting the proclamation text, against a Covington alert saying 22 August. A third source now also gives 22 August, and explains the gap: the 19 August date was suspended for three days. On the balance of evidence the duties took effect 22 August, and the earlier 19 August statements here were wrong. The Federal Register notice remains the document that would close it.