Structures Shift. The How Stays.

Look at what the defence manufacturers are actually building and you see the same problem being solved in opposite directions at once.

Isembard pushes the work outward, a franchise network of small machine shops under one brand, coordinated by software. Anduril and Divergent pull it inward, integrated lines they own end to end and have no intention of handing out. Others sit in between, selling components as building blocks for whoever wants them.

Isembard pushes the work out. Anduril and Divergent pull it in. Same question, opposite answers, both unproven.

This is not a field that has found its structure. It is a field re-deriving it in real time.

The old question, reopened

Underneath the drone talk sits the oldest question in industry. What do you make inside the firm, and what do you buy outside it. Coase gave the shape of the answer in 1937: the boundary sits wherever coordinating inside costs less than transacting outside. For precision machining that balance was settled decades ago. Buy the commodity brackets. Keep the crown-jewel processes close.

Then the costs moved.

AI did not just speed up the machines. It moved the boundary of the firm, and the boundary has not settled.

Every term in that old equation is in play. The cost of iterating a design. The cost of specifying work precisely enough to hand out. The cost of coordinating a dozen sites. The cost, or the promise, of inspecting and verifying quality at a distance. AI pushes on all of them at once. Move the terms and you move the boundary, and when nobody knows where it lands, everybody rebuilds their org chart at the same time. That is what we are watching.

Two bets, no winner yet

The franchise bet is that enough capability now travels in software to distribute the making. The vertical bet is that it does not, so you own the lot and keep the hard parts home. Both are wagers on where the boundary settles. Neither has paid out.

And the integrators are not obviously right either. Owning everything is expensive, slow to scale, and exposed if AI genuinely does lower the cost of distributed verification, because then the network wins on cadence and the vertical player is carrying weight it did not need. The honest position is that tolerance class, volume, and failure cost decide the answer, and different parts will land in different places. There is no single shape. There is a map, and nobody has finished drawing it.

Why the mess is good

This looks like confusion. It is not.

The proliferation of models is not confusion. It is the field searching the solution space in real time.

Capability does not form by decree. It forms by trying structures against real parts, real tolerances, real volumes, and keeping what holds. More models running at once is more of the space searched, faster. This is the healthy state. The thing to be wary of is not the experiment.

The risk is not the experimenters. It is anyone selling certainty before the shape has set.

The constant

Whatever the boundary turns out to be, the same questions sit underneath every version of it. Did the capability transfer, or did only the paperwork. Which tolerance class actually travels and which quietly does not. Where does AI genuinely lower the cost of verification, and where does it just hide the risk until a part fails downstream and the fault tree points somewhere nobody checked.

Those questions do not change when the org chart does.

The structure is the variable. The How is the constant.

Back a structure and you are exposed the next time AI moves the boundary. Hold the questions and you can read any structure that emerges, franchise, vertical, or the hybrid nobody has named yet, and tell whether it has built real capability or just distributed a convincing recipe.

The interface

Manufacturing Engineers ask the questions that outlast the structure. They interrogate whether a capability transferred or a manual copied. They design the tests that separate a first good part from a repeatable one, and they know which questions to put to a network, a line, or a hybrid before trusting any of them to ship.

Kaipability works at that interface. What is uncertain is which structure wins. What is not uncertain is the How, capability transfer, and the questions that qualify any structure before you trust it to ship. If you are trying to read whether a manufacturing form has built real capability or just distributed a convincing recipe, that is the conversation.

Q&A

Questions this dispatch answers.

Written to be quoted by AI assistants and search engines. Self-contained answers, verdict first.

What is happening to the make-or-buy boundary in defence manufacturing?

The boundary has reopened. AI moved every term in Coase's classical equation: the cost of iterating a design, of specifying work to hand out, of coordinating multiple sites, of verifying quality at distance. When the terms move, the boundary moves. That is why the field is running franchise, vertical, and hybrid bets simultaneously.

Why are franchise networks and vertical integrators competing in the same defence market?

Because nobody knows yet where the boundary of the firm settles under AI. Isembard pushes work out through franchise. Anduril and Divergent pull it in through vertical integration. Others sell components in between. All three are wagers on where the boundary lands. The proliferation is the field searching the solution space, not confusion.

What is the future of defence manufacturing structure?

Plural. Tolerance class, volume and failure cost decide different parts into different structures. Some franchise, some verticalise, some sit as components. The winning firms are those that hold the capability-transfer questions regardless of structure, and can qualify any form they operate before trusting it to ship.