The deal was signed. Oil slid to a three-month low. And the ships still did not move.

In the hours after Washington and Tehran announced a reopening of the Strait of Hormuz, insurers held their position. Tankers waited on a seventy-two hour de-mining clearance window before they would cross. Owners asked for fresh war-risk quotes. Reinsurers wanted to see the survey before they would price the next voyage. The headlines said the strait was open. The underwriting said it was not.

I keep coming back to that gap because it is doing more work than the press conference. It is the most honest read on what a deal actually buys you in 2026.

A political announcement can move a futures contract in seconds. It cannot, by itself, move a hull. What moves the hull is a chain of small, unglamorous verifications: a minesweeper report, a P&I club sign-off, a flag-state notice, a charter party amended to reflect new routing. Each link is a person with a job and a liability. None of them care about the press conference. They care about whether they can defend their decision the next morning if something goes wrong.

This is what the Hormuz lag reveals, and it travels further than oil.

Access is not throughput

Diplomatic permission reopens a route. It does not restore trust on that route. Trust, in any serious supply chain, is not a sentiment. It is an underwriting function. Somebody, somewhere, is pricing the residual risk and writing it onto a balance sheet. Until that person is willing to sign at a rate the buyer can absorb, the lane stays closed for practical purposes even if it is open on paper.

What looked like an oil story is really a clock. The clock starts when the announcement lands. It ends when an underwriter is willing to put a number next to the shipment. The space between those two moments is where every supply chain in the world actually lives.

Most of the time we do not see this clock because the lanes are quiet and the rates are stable. A crisis exposes it. So does a buildout.

The Taiwan counterpoint

A few thousand miles east, a different kind of deal has been forming, and it is structured against this exact lag.

Taiwan's research arm has been working with U.S. defense firms on drone development, and the recent expansion of cooperation with Anduril is not framed as a procurement contract. It is framed around AI applications, drone manufacturing, supply chain integration, and local production. The language matters. The deal is not "we will sell you airframes." The deal is "we will stand up a verifiable production loop on your territory, with components whose provenance you can audit, and an integration path your engineers can rehearse."

That is a different shape of agreement. It is designed so that the headline and the operational reality converge faster, because the operational reality is part of what is being signed.

The Hormuz deal moved a price. The Taiwan partnership is trying to move a capability, and the partners are doing the verification work inside the deal rather than waiting for insurers to do it afterward.

This contrast is what makes the mechanism visible. It is not that political deals are empty. It is that the ones which translate quickly into throughput are the ones where verification is built into the structure, not bolted on after the announcement.

What this means for AI supply chains

Now look at the way AI capacity gets announced.

A hyperscaler signs a multi-year compute commitment with a sovereign buyer. A chip vendor agrees to a fab partnership in a new jurisdiction. A model lab signs a data residency deal with a national champion. A government publishes an MOU on cross-border AI infrastructure. The market reads each of these the way it first read the Hormuz deal: as throughput.

My read is that most of them are closer to the press conference than to the minesweeper report.

The AI supply chain has the same structural feature as the tanker lane. There is a long chain of small verifications between the announcement and a workload that actually runs at the promised cost, latency, and provenance. Component sourcing. Power interconnect. Cooling commissioning. Network peering. Model weight custody. Export-control compliance on the specific SKUs delivered, not the SKUs promised. Integration tests against the buyer's actual code and data, not a demo environment. Each one has an owner. Each owner is, in effect, an underwriter.

When any one of them hesitates, the announced capacity is not capacity. It is optionality on capacity, priced at a risk premium that does not show up in the press release.

This is the same calendar gap I wrote about when robot half-marathons and procurement bans collided in the same week. The capability is real. The permission is real. The interval between them is a capital allocation variable, and most boards still treat it as a rounding error.

The design move

The useful move here is not skepticism. It is a small change in how AI capacity decisions are gated.

Stop conditioning capacity decisions on contractual milestones. Condition them on operationally verifiable ones. Specifically, the ones a third party would be willing to underwrite.

A few examples of what that looks like in practice:

A compute commitment is real when an independent auditor can confirm a live integration test on the buyer's actual workload at the promised price-per-token, not when the term sheet is signed. A fab partnership is real when sample wafers from the new line clear the buyer's qualification regime, not when the groundbreaking photo is taken. A sovereign AI agreement is real when a regulated workload runs end to end inside the new jurisdiction with documented data lineage, not when the joint statement is read.

None of this is exotic. Shippers, reinsurers, and trade financiers have run this playbook for a century. The new part is applying it to AI infrastructure with the same seriousness, instead of letting the announcement do the work the verification has not yet done.

There is a quiet discipline in this. The political climate is not in your control. The shape of the next deal is not in your control. What is in your control is the criterion you use to commit your own capital, your own roadmap, and your own customer promises against somebody else's announcement. If you tighten that criterion, the noise outside matters less.

One design rule

Here is the rule I would put on the wall.

Before your next AI capacity decision moves to a yes, ask whether an independent party would be willing to underwrite the flow you are betting on. Not the strategy. Not the relationship. The flow. The specific shipment, the specific workload, the specific integration, at the specific date, under the specific conditions.

If an insurer or an auditor would still hesitate, the announcement has not earned safety yet, and your decision is paying a risk premium nobody on your team has named.

The ships outside Hormuz were not waiting on politics. They were waiting on somebody willing to sign. That is the question worth carrying into every AI infrastructure deal that lands on your desk this quarter.

The deal is what they announce. The throughput is what they can underwrite.

Sources