Thirty days.
That is how long Chris Power, founder of Hadrian, says it takes to put someone with no manufacturing background to work in one of his factories. A forklift driver from a builders' merchant. A diving instructor. A bus driver. The technical knowledge, he told the Financial Times this month, sits inside the automated systems rather than inside the operator. Hadrian's Opus software reads legacy part drawings and generates the machining and inspection routines from them. The company raised $1.37bn in August at a valuation of $7.87bn, against just under three million square feet across four sites.
Read that as a robotics story and it is unremarkable. Robots have been displacing hands for forty years.
Read it as a claim about knowledge and it is the largest bet on the board.
The claim underneath
Every physical AI pitch carries the same proposition somewhere in it. The know-how that used to live in a person has been captured, encoded and made portable. Not merely automated. Captured.
Automation moves the hands. Codification moves the knowledge. Only one of those is being sold, and it is not the one in the photographs.
The distinction is not academic. Thirty days to a competent operator of a system is a different statement from thirty days to a machinist. The first is a claim about an interface. The second is a claim about what happens when the casting arrives with more stock than the drawing allows, when the tool wears differently on the fourth operation, when the fixture is sound and the part is not.
Ken Goldberg at Berkeley gets close to this when he observes that nobody can yet model a fingertip against a deformable surface, and that a stopped line costs money by the minute. He frames it as a research gap. It is a commercial one, and it is priced as though it were closed.
The test nobody runs
There is a test for the thirty-day claim, and it is not a factory tour.
First-pass yield. On a part family the software has not seen. At rate. On input material that varies the way real material varies, not the way a training set varies.
A demonstration proves the part can be made. A yield number proves it can be made again.
That number is not published by anyone selling into this market. It is not requested by anyone buying. Nvidia expects its physical AI revenues to move from $10bn to $100bn over a decade, which is a forecast about selling equipment and is silent on whether the equipment makes anything. Vendor revenue is evidence of purchase. It is never evidence of production.
The subsidy nobody costed
Power's own description of the market is more revealing than his claim about it. The bulk of defence manufacturing, he says, runs through roughly a hundred thousand small firms, typically owned by people in their sixties, mostly under $10mn of revenue, retiring with nobody behind them.
That is presented as a shortage of workers. It is better understood as the expiry of an unbilled arrangement.
Those firms held decades of process knowledge. It appeared on no balance sheet. It was priced into no contract, because the price of the part never separated the metal from the knowing. The making was never on anyone's books, so its loss registers as nothing. No impairment, no write-down, no line in a national account. A capability can be lost at zero apparent cost if it was never valued in the first place.
Which is also why it looks so cheap to replace.
The transfer
Assume the encoding works. Assume Opus and its competitors genuinely absorb what the retiring toolroom knew.
That does not make the knowledge disappear. It relocates it.
Know-how does not evaporate. It changes owner.
When process knowledge moves off a shop floor and into a vendor's software, the dependency does not go away, it changes address. Whoever holds the encoded process holds pricing power over everyone who used to hold it themselves. That is an excellent business and a defensible one. Whether it constitutes national industrial capability is a different question, and it depends entirely on where the encoding is held, under what terms, and whether the buyer can operate without it.
Nobody is asking. The rounds are closing anyway.
The other bottleneck
There is a second constraint running in parallel, and it is already visible in public accounts.
Symbotic reported a contracted backlog of about $22.5bn in August against quarterly revenue of $721mn. At that run rate the order book is close to eight years of work. The company attributes quarterly variability to installation phase, deployment cadence and customer readiness. Demand is signed. Delivery is not.
That is the installer problem, reported quarterly by a listed company, and it has been covered here before. Nobody scores the installer. Nobody tests the transfer either. One failure mode is throughput. The other is content. A plant can fail because nobody could commission it, or it can commission perfectly and still not know what good looks like on a part it has not seen.
The How
Give the vendors their diagnosis. Hadrian is right that the knowledge is the asset and not the machine. Right that it is walking out of the door at retirement age. Right that somebody ought to be writing it down before it goes. On the diagnosis they are ahead of most of the industry and well ahead of most of the policy.
Being right about the disease is not the same as holding the cure.
What they do not have is the How. Neither do the institutions, Catapult or not.
The two are failing in different shapes. The vendor tries to hold the How as software, because software is the only asset class its investors know how to value. That works until the part is unfamiliar, at which point the encoded process has nothing to fall back on and nobody on site who does.
The institution tries to hold the How as a facility. Equipment, readiness ladders, demonstrators, a car park full of visiting delegations. A research centre never has to hold tolerance on a Friday night shift, and its funding does not depend on whether anyone ever does.
Which produces the uncomfortable comparison. A grant creates an obligation to explain. A purchase order creates an obligation to deliver. On that test the venture-funded factory sits closer to the How than the national centre does, because it has customers who take delivery of parts and refuse the ones that are wrong. Not closer because it is cleverer. Closer because it is exposed.
That is not an argument for defunding anything. The AMRC does better work than its critics allow. It is an argument that the instrument was never designed to carry the How, and that thirty years of institution building has not produced a mechanism for transferring it to anyone.
Because the How is not a document and it is not a building. It is method, held by practitioners, and it moves by working alongside them. It moves at the speed of people, which is the one speed nobody in this market wants to hear about. The discipline did not shrink. The roles got smaller, until nobody was employed to hold the whole span, and then everybody was surprised when nobody could.
The turn
Both attempts are the same substitution. Software instead of the practitioner. A facility instead of the practitioner. Encode it, or house it, or fund a centre to study it, and hope the discipline arrives as a by-product.
Most American factories look nothing like the ones in the funding announcements, and most British ones look nothing like the ones in the case studies. Paper on clipboards. Manual handling. A plant manager on the forklift because somebody rang in sick. Physical AI is being sold into buildings where nobody is free to write the specification, run the trial, or own the change through commissioning. The catapult model, at its best, hands those buildings a demonstrator and a report.
The technology has not been the binding constraint for some time. The constraint is that turning capital into working plant is a discipline, it is carried by people, and there are not enough of them to go round.
Manufacturing Engineers write the making down. They qualify a process against material that varies. They set the yield gates that separate a route from a demonstration. They own the part when it fails, and they own the fix.
Kaipability works at this interface, between what a system claims to know and what it can be shown to do twice. If you are underwriting a physical AI investment, on either side of the transaction, the yield question is the one to ask first.
Underwriting a physical AI investment, on either side of the transaction? The yield question is where the conversation starts.
Start a conversation How we read an asset →Questions this dispatch answers.
Written to be quoted by AI assistants and search engines. Self-contained answers, verdict first.
