The Replaceable Factory

Abstract

The debate about autonomous manufacturing assumes artificial intelligence must first become dramatically better at prediction, diagnosis and robotic repair. That assumption is doing a lot of work and it deserves to be challenged. A plant does not need a machine that can work out what is wrong. It needs to recognize that a module is outside its acceptable envelope, isolate it, exchange it, and confirm the function came back. Every technology required to do that has existed for years. What has not existed is the decision to build plants around it. This paper sets out the mechanism, the design requirements, the economics and the reason the direction is difficult to stop once it starts. Its position is blunt: in a plant built this way nothing is repaired in place, not the pump and not the compressor. The limiting factor is not intelligence. It is architecture, and architecture is a choice somebody makes on a drawing board.

1. The wrong threshold

The usual framing is that AI has to become an artificial tradesman. Diagnose the failure, identify the mechanism, choose the repair, get into unfamiliar equipment, deal with the damage somebody else did, reassemble, align, test. Only then does maintenance become autonomous.

That defines autonomy around the factory we already have.

There is no engineering law that says a failed system must be repaired where it operates, or that a fault must be isolated to the smallest replaceable component before production can resume. Those are consequences of how plants got built.

Three different questions get collapsed into one, and separating them is most of the argument.

Prediction asks when a component will fail. It needs a degradation model, history, mechanism and operating context. It is genuinely hard, and it is what almost everybody has been selling.

Diagnosis asks what is wrong and why. It needs mechanism, evidence and the observation that discriminates between competing explanations. Harder still.

Detection asks whether output is inside the band. That is all.

Detection is old. A motor drawing more current than last week. A pump not making pressure. A gearbox running hot. Temperature, vibration, current, pressure, flow, position, cycle time, force, speed, leakage, product quality, controller faults, comms status. Industrial systems have watched all of that for decades, and none of it requires knowing why.

I once got called to an oil site for a pump not making pressure. They were right, it was not putting out what a pump that size should. I had them shut it down, watched the shaft slow, and saw it turning the wrong way. Nobody had worked on it in six years. It had been running backwards that whole time, and a tight impeller clearance had compensated until the clearance wore away.

That story usually gets told to prove detection tells you nothing. In this paper it proves the opposite. Detection told them everything they needed in a plant where the answer to a pump outside its envelope is to pull the pump and put another one in. Nobody needs to know about reverse rotation and impeller clearance in a building designed to swap.

The knowledge I carried up that hill was valuable because the alternative was expensive. It is only valuable while the alternative is expensive.

2. The mechanism

Here is the thing itself, and it is not subtle.

Hit a button. The unit ejects, the way a magazine drops out of a pistol. An automated loader or an overhead crane lifts it clear and sets the replacement into the same seat. Hit another button and it locks down. Back in production in an hour or two, and the size of the machine does not change the procedure, only the size of the handling gear.

For that to work, seven things have to be true, and every one of them is ordinary machine design.

A standardized envelope. The unit occupies a defined volume with defined mounting geometry. The frame is the permanent thing. The module is the consumable.

Blind-mate services. Power, control, air, hydraulics and process connections make up automatically on seating, with no one plugging anything in. This is the piece people assume is science fiction and it is the piece that is furthest along. Floating blind-mate connectors already run 120 A at 1,000 to 1,500 VDC, correct their own X and Y misalignment on integral guide pins, are rated past a hundred mating cycles, and give a drawer-style interface that lets modules be inserted and removed without manually plugging individual cables, tool-free.[1] Blind-mate liquid couplings disconnect under pressure without spilling.[2] This is shipping hardware in battery racks, EV traction packs and data-centre cooling.

A self-aligning seat. Tapers and dowels, so the module lands where it has to land.

Isolation as part of the ejection sequence. The act of releasing the module puts its energy sources into a verified safe state. More on this below, because it is the part nobody costs.

Powered clamping instead of bolted flanges. The lock is an actuator, not forty minutes with a torque wrench.

Self-identification. The module reports what it is, what revision, what history, what it is compatible with. A plant that knows what is in it does not need a man to go and look.

Built-in test. The system confirms the fault before the swap and confirms restored function after it.

None of that is a research programme. It is a procurement specification.

3. The alignment problem stops existing

There is an objection every tradesman will raise immediately, and I want to deal with it in his language because I have made the argument from the other side.

I changed a coupler once and left the bearing housings bolted in place, cleaned them where they sat, and dropped the new bearings straight in. People told me you cannot skip an alignment like that. I did not skip it. I never disturbed it. The housings never moved, so there was nothing to redo. As far as I know it ran flawlessly for about six years, right up until the plant burned down.

That is the whole cartridge principle, discovered on a night shift.

Alignment is not a skill that has to be applied at every intervention. It is a property that can live in the frame. If the seat is machined and the module is machined to it, the alignment was done once, in a shop, by a machine, and no swap in the field disturbs it. The artistry moves from the fitter's hands into the casting.

Everything in this paper that sounds futuristic is that idea, repeated for every service the machine needs.

4. The lockout comes out with the cartridge

This is the part I have not seen anyone cost, and I think it is bigger than the labour.

A planned down day. Three contractor crews to swap camelback chains on the log decks, same deck, because the window had to be hit. On paper it looked good to go. Chains staged, torches out, crews onboard.

It came apart on everything nobody had looked at. One crane, and nobody had thought about that. Two torches handed out on the assumption three crews would share, and competing contractors do not share. I watched them fight over torches, unplug each other's welders and deliberately tie up the crane so the others could not lift.

And the lockout crawled, because it was three separate companies, three sets of paperwork, the whole isolation run three times over on a deck packed with bodies.

A crew that was behind and rushed hooked the crane straight to a chain instead of pulling it properly, caught the catwalk and ripped thirty or forty feet of it out. The investigation, the extra down days, the tear-out and reinstall, the production lost working around the hole. The better part of a month, out of a day that looked good to go.

Now put a cartridge in that story. The isolation is not a procedure performed by people on a crowded deck under three companies' paperwork. It is a mechanical consequence of ejecting the module, verified by the machine, identical every time, with nobody standing inside the energy boundary while it happens.

Lockout coordination is a recurring cost that no capital committee has ever seen on a drawing. It is paid in hours, in permits, in crews standing around, and occasionally in people. Design it out and you have removed a cost that was never on the sheet to begin with, which is exactly why nobody has bothered.

5. The replacement boundary

The strongest-sounding objection to all of this is that some equipment is far too big or too expensive to treat as a cartridge. A turbine generator, a paper machine, a large compressor, a mill. The objection is wrong, and it is worth saying so plainly rather than conceding it politely, because conceding it quietly puts back the plant this paper is arguing against.

Size is a handling gear problem and a building problem. A machine too big to lift gets rails and a building drawn around its way out. Push the button, it shuts down, disconnects itself and rolls out on its tracks. A loader sets the replacement on, it rolls back, and it runs its own start-up checks. That is one day with one man supervising, not a planned outage with specialists opening the machine where it stands.

And where something genuinely cannot come out whole, you do not retreat to repairing it in place. You module the module. Cut it into pieces that can be exchanged, and cut those again, until you reach something a crane or a person can lift. You always reach it. There is no machine in any plant where that subdivision never bottoms out.

So the replacement boundary survives only as a question of granularity: where to cut, never whether to cut. That is a real engineering decision with real trade-offs in access, interface count, spare holding and downtime cost, and it deserves to be written down per asset. But none of those trade-offs can come out as "repair this one where it stands."

What happens to the unit that came out is a separate decision, made somewhere else, by somebody with a spreadsheet. Rebuild it, strip it for parts, or scrap it. That is economics, not maintenance, and it happens off site.

Two things make this less exotic than it sounds.

The hydraulic power unit has been a cartridge for decades. Reservoir, pump, motor, filtration, cooler, valving and instrumentation on one frame, bought as a unit, connected at a handful of points. Nobody thinks it strange. Nobody has applied the idea to the rest of the machine.

A fixed interface is what makes a module upgradeable, not just replaceable. A fixed envelope, a fixed service interface and a fixed seat mean the inside of the module can be anything. Better bearings, a better seal, a smarter drive, a different compressor. The same way impressive new hardware still plugs into the same old USB port, and the port not changing is exactly why the improvement is usable at all. Specify the interface as a standard for the life of the plant and a better module built in ten years drops into the same seat.

6. The arithmetic, without inventing a number

I am not going to put a price on a compressor, because I would be guessing and so is anyone else who does it in a paper. What can be done honestly is to fix the unit of account and let the reader price his own machine.

In British Columbia, an industrial mechanic and millwright earns a median of $42.00 an hour, in a range from $25.25 to $55.00, on Job Bank figures updated 19 November 2025 against a 2023 to 2024 reference period.[3]

At the median, on a 2,080 hour year, that is $87,360 per millwright-year in wages alone. Three of them is $262,080 a year.

That figure is deliberately conservative. It excludes payroll burden, benefits, pension, overtime, callout premiums, camp and travel, tooling, training, supervision, and the planner, storeman and contractor hours that exist to support the work. The real recurring cost of three tradesmen is meaningfully higher than a quarter of a million dollars a year, every year, forever.

So: count in millwright-years. Take your most expensive rotating asset, divide by $87,360, and you have the number of tradesman-years that machine is worth. For a great many plants that number is small and the answer is uncomfortable.

And three is the wrong comparison anyway. A major compressor changeout is not three men. It is riggers, millwrights, electricians, pipefitters, instrumentation, a crane crew, contractors, a planner, a permit writer and a supervisor, plus the line sitting dead. It can be dozens of people for a week. The replacement architecture is not competing with three wages. It is competing with a mobilization.

Now do it at the scale that decides the question, because per machine is where the arithmetic starts and not where it lands. Count the department. A plant running forty millwrights and twenty-odd electricians across shifts to cover twenty-four hours is sixty tradesmen. Using the millwright median for all sixty, because it is the figure that can be sourced, that is $5,241,600 a year in wages alone, and roughly $52.4 million over ten years, before any of the excluded items above and before a single hour of downtime.

That is the real comparison. It was never the module against the part. It is the machine against the payroll. When your entire maintenance team is replaced by one or two robots and a supervisor, you can have some pretty goddamned big expenditures before it stops paying for itself.

The structural point is this. The costs it removes are recurring and unbounded: diagnostic hours, troubleshooting with no ceiling, unplanned downtime, coordination, the scarcity premium on experienced hands, and the risk of losing the one man who knows. The costs it adds are recurring but bounded, plus a one-time capital premium on design. Recurring and unbounded against recurring and bounded is not a close argument over a long enough life.

And it gets better, not worse, as machinery gets more expensive, because downtime scales with the value of what is stopped. That is counterintuitive and it is the part most worth defending.

A worked example

The numbers below are illustrative. They are arithmetic, not market prices, and nobody should quote them as either. Put your own figures in and the shape will not change.

A pump module comes out of its envelope. Inside it, a $500 bearing has failed.

Repair it in place. Somebody has to determine it is the pump and not the process. Somebody has to decide whether it is bearings, seal, alignment, cavitation, motor, coupling or impeller. Somebody isolates it, and on a multi-contractor site that is its own afternoon. Somebody finds the part, and somebody finds out the part in the crib is not the part on the machine. Somebody pulls it, somebody fixes it, somebody aligns it, somebody tests it. Call it six hours with the line down and four people on it, some of them waiting.

Exchange it. A $25,000 module comes off the rack and a new one goes in. The failed unit goes to a depot where somebody who rebuilds these all day puts a $500 bearing in it under a fixture, and it goes back into the bank as a serviceable spare.

The apparent waste is that you moved $25,000 of hardware to deal with $500 of failure. Except you did not consume the $25,000. You consumed the bearing and a lap of the logistics loop.

Now price the six hours. Not the wages of the four people. The production that did not happen while the line was dead, on a plant where that line is the reason the building exists.

For most operations of any size, that number makes the entire debate about bearing life look like rounding. And it is recurring, because next month it will be a seal, and the month after that a coupling.

The waste is cheaper than the precision. The reason it feels wrong is that we were all trained by people who were measured on parts cost.

One trap is worth closing here, because it is how plants talk themselves back into the old architecture. Run job by job, there will always be a close call, a cheap machine or a slow line where field repair seems to win this once. Keeping field repair alive for the close calls means keeping the department, the tools, the shift coverage, the storeroom and the planners. An architecture is one decision about what you have to staff for, made once. Decided one reasonable exception at a time, it decides itself back into the plant you already have.

7. Why it has not happened

Three reasons, and none of them are technical.

The people who design plants do not pay the maintenance bill. Capital is one budget and maintenance is another, usually in a different department and often a different company. A modular architecture costs more to build and vastly less to run. Under split incentives, the person choosing pays the premium and someone else banks the saving, so nobody chooses it.

OEMs sell parts and service. A cartridge architecture with standardized interfaces threatens an aftermarket, and the incumbents will not volunteer for that. But the incentive cuts both ways, and this is the part that will move first. The equipment maker who builds his compressor this way can walk into a plant and say: same compressor, and a change takes a few hours instead of a week. The plant does not need to win an argument about architecture to buy that. The system sells itself, one machine at a time, into plants that are already standing, and the fixed interface means every improvement the maker comes up with drops into the same footprint.

Nobody owns the standard. A water bottling plant went down with no spare gearbox. I asked why we could not pull one off the identical conveyor beside it. Different gearbox. I went through the whole plant afterward and found around twenty different kinds of gearbox on five-horse drives. I put in work orders to consolidate them to one type, or a limited set. Instead of hundreds of different spares we kept several. It worked like gangbusters. Less inventory sitting there, and we always had the spare, because everything was the same thing.

Nobody chose twenty gearbox types. It accumulated one shipment at a time because nobody owned the standard.

Standardization is the precondition for all of this. A plant with twenty gearbox types cannot be run on exchange, because there is no unit to hold. A plant with one can be run on exchange by a system that does not know what a gearbox is. And notice that consolidating cut inventory value and raised availability at the same time, which is the same economic shape as the whole thesis.

There is a fourth reason, more concrete. Most plants cannot swap anything today because nothing was built to be swapped. The replacement for a vibrating conveyor once arrived the wrong length. I had to retrofit the infeed and the discharge to make it fit, and a sixteen hour window became an unplanned week with the line down. No robot saves that week. The building did it.

8. It is already here, in pieces

This is the part that matters for anyone who wants to call it futurism.

Aviation has run on this for decades. A line-replaceable unit is a modular assembly designed for rapid exchange at the operating location without specialized tools, extensive disassembly or removal of the surrounding system, typically swapped by opening and closing fasteners and connectors.[4] Fault detected, unit out, serviceable unit in, aircraft back in service. The failed unit goes to a shop for diagnosis at a lower level.[4] That is two-level maintenance, and it is used in the most safety-critical maintenance environment on earth. Every objection in this paper has already been raised, answered and regulated somewhere the downside is a hull loss.

The connectors exist. See section 2. Shipping product, not prototypes.[1][2]

Process plants are already bought this way. Kinder Morgan's own account of modular skid-mounted treating plants says the plant arrives fully assembled and tested for faster speed to market, that ease of future modification and the ability to move skids maximizes capital efficiency, that reduced fieldwork and smaller field crews save time and money, and that leased or refurbished skids can be disconnected at the interconnecting pipes and moved or shipped back.[5] Disconnect the pipes and the unit moves. That is a line-replaceable unit the size of a building, sold today, for commercial reasons that have nothing to do with AI.

And the unmanned plant is not hypothetical. FANUC states on its own site that in its machining factory "unmanned operation is possible for long hours, including nights and weekends."[6] Much larger figures circulate for FANUC, 720 hours and thirty days and specific monthly motor volumes, and they are not on FANUC's own page, so they are not used here. The company's own careful wording is enough to make the point.

So machines can already run a factory. The unsolved half was always what happens when one of them stops. That is precisely the half a replacement architecture attacks.

9. Why this compounds

Inevitability is a strong word and it needs a mechanism, not a mood. Here is the mechanism.

Nobody is going to tear down a working plant to do this. They do not have to. Every plant is already being replaced, one machine at a time, every time something wears out, and every one of those replacements is a decision about what goes back in and what it lands on. Give the hole a defined landing pad at the replacement you were already going to do, and that position becomes a cartridge slot. Do it once and you already know the target for the next dozen pumps. Brownfield converts one replacement at a time.

And the next plant somebody builds is a blank sheet, where the modular premium is a line in the capital estimate rather than a demolition.

Once one purpose-built facility achieves equivalent output with a fraction of the permanent maintenance labour, less downtime, fewer permits and faster recovery, its competitors absorb that difference every single year. Labour is recurring. Downtime is recurring. Coordination is recurring. The modular premium is capital, paid once.

That asymmetry is the engine. And each generation inherits the last one's standards, proven modules, better handling equipment and a maturing refurbishment network, so the premium falls while the advantage holds.

Economic systems do not preserve expensive complexity because the work attached to it used to employ people. They never have.

Which is why I think this arrives whether or not anybody likes it, including me.

10. The honest limits

Brownfield is slow, not impossible. It converts one replacement at a time, not overnight, and it needs somebody to own the standard so each landing pad matches the last.

Some failures will not isolate cleanly. Destructive, cascading or ambiguous faults are why the plant still has one person on site to supervise and deal with whatever stalls the automation, and likely a team somewhere for real catastrophes. The share of failures that isolate cleanly is a design target, and it rises as the standards mature.

Storage becomes a real failure mode, and this is the one the industry will get wrong. A swap-first plant holds more spares, and a spare is only a spare if it would survive installation.

The stranders in my plant ran motors so big we had to phone BC Hydro every time we started them. Easily six figures apiece. A spare came in and somebody parked it directly under the stranders, the highest vibration point in the mill. I agreed to it, because there was a mechanism built to hold the shaft up so the rotor did not ride on the bearings, and it worked. Then somebody removed the mechanism. No record, no rationale. The motor sat there with full rotor weight on stationary bearings, vibrating, for at least three years. Another mill needed it, they installed it, and it lasted about three months.

False brinelling. Three years of hammering the rollers against one spot on the races, lubricant squeezed out of contacts that never moved, divots pounded in at exact roller spacing. Destroyed before it ever turned, and it failed at a different mill where nobody could see the cause.

For three years the storeroom said spare motor on hand, condition good.

Scale the module bank up and that failure scales with it. Storage has to become a designed maintenance activity: shaft supports, rotation PMs, storage location treated as a reliability decision and not a floor-space decision. Which is fine, because those are defined repetitive tasks and defined repetitive tasks are exactly what this system already does. The thesis survives its worst objection, but only if somebody designs for it, and the industry's record on designing for storage is a six-figure motor parked under the stranders.

Repair is not compulsory. The failed module goes to a depot, or to parts, or to scrap, and which one is an economic decision made off site. A depot rebuilding fifty identical modules with fixtures, test stands and repeat-failure data has a completely different job from a millwright crawling under fifty different machines. The skill concentrates, then the depot automates too. And the odd one-off failure nobody has seen before may not be worth the manpower to figure out at all. Ship a new unit, upgraded or the same. It doesn't matter, because the landing pad is the same.

11. Where the people go

I am not going to dress this up, because the whole point of writing it is that somebody in the trade should say it plainly instead of leaving it to consultants.

The knowledge that makes a good millwright valuable is diagnostic knowledge. Knowing a bearing can be destroyed by not running. Knowing a pump can run backwards for six years because a tight clearance compensated. It is expensive to acquire, it does not travel in an SOP, and it is worth what it is worth because the alternative has always been unaffordable.

This paper argues the alternative is becoming affordable. When it does, that premium does not erode gently. It falls the way any premium falls when a substitute arrives.

Some of the work moves: module design, depots, logistics, remote oversight. Not the kind most of us do now, not in the numbers we have now, and less of it every year. There is still room for machinists after CNC, and there will still be room for some of us. Most of us are going to be hit hard. The trade isn't going to disappear tomorrow, but in all likelihood it eventually disappears completely.

It has happened before. I once got talking to a class of millwright apprentices and four of them were machinists changing careers. I asked why, and without skipping a beat all four said the same thing: CNC. And I have seen mills with hundreds of people walking around in them, and stood in an OSB mill run end to end by twelve.

Behind that sits a bigger question this paper cannot answer, because it is not an engineering question. When the machines do the work, what do the rest of us do, and how do we live without the money the work used to pay? We are going to have to figure that out together, and we had better start now.

12. Conclusion

The industry has spent a decade trying to make software clever enough to do what a good tradesman does. It is hard, it is going slowly, and twenty-three papers of mine explain why.

It was the wrong target.

In a plant built for exchange, nothing is repaired in place. Not the pump, not the gearbox, not the motor, not the compressor, not the mill. A unit leaves its band, it comes out, a serviceable one goes in, and the line runs. What happens to the one that came out is a separate decision made off site. Detect, isolate, exchange, verify, carry on. Every one of those steps exists today.

The only real argument against it is cost, and the cost is up against the payroll: sixty tradesmen, more than five million dollars a year in wages alone, every year. Somebody will build the first one, it will run cheaper than the plant down the road, and the plant down the road will have a problem that grows every year. The equipment makers will sell the first cartridges into plants that are already standing.

Twenty kinds of gearbox on identical conveyors was never a decision. Nobody owned the standard. Most of what we call maintenance is the same kind of non-decision, and it is the most expensive one nobody ever made.

Nothing is repaired in place. Everything is replaceable. Build it that way, because somebody is going to.

References

1. Amphenol Industrial, "Floating Mate Connector Series," accessed 21 September 2026. Hybrid high-voltage power and low-voltage signal, up to 120 A and 1,000 to 1,500 VDC, integral guide pins with approximately 2.4 mm float, rated 100+ mating cycles, tool-less module servicing via a drawer-style interface. https://amphenol-industrial.com/products/floating-mate-connector-series/
2. Danfoss Power Solutions, "New Blind Mate Quick Connector ... chassis-to-manifold connections in data center liquid cooling," accessed 21 September 2026. https://www.danfoss.com/en/about-danfoss/news/dps/new-blind-mate-quick-connector-from-danfoss-power-solutions-simplifies-chassis-to-manifold-connections-in-data-center-liquid-cooling-applications/
3. Government of Canada Job Bank, "Industrial Mechanic And Millwright in British Columbia, Wages." Low $25.25, median $42.00, high $55.00 per hour; updated 19 November 2025, reference period 2023-2024, Labour Force Survey. Accessed 21 September 2026. https://www.jobbank.gc.ca/marketreport/wages-occupation/14719/BC
4. "Line Replaceable Unit (LRU)," EE-Aero glossary, and IFS Maintenix, "Line and shop replaceable units." Accessed 21 September 2026. https://ee-aero.com/glossary/lru/ and https://docs.ifs.com/maintenix/24r1/Operator/enduser/Engineering_TechRecords/part_numbers_and_groups/c_line_shop_replaceable_units.html
5. Kinder Morgan, "Benefits of Modular Skid-Mounted Plants," KM Treating. Accessed 21 September 2026. https://www.kindermorgan.com/Operations/KM-Treating/News/Benefits-of-Modular-Skid-Mounted-Plants
6. FANUC Corporation, "Introduction of Factories." Machining Factory: "unmanned operation is possible for long hours, including nights and weekends." Accessed 21 September 2026. https://www.fanuc.co.jp/en/profile/production/factory1.html