A hardware company can usually make one object. The problem is making the next hundred objects while still not knowing exactly what the object should be.
A prototype proves that a design can exist. It does not prove that strangers will pay for it, operators can use it, suppliers can reproduce it, technicians can repair it, or the product can survive six months outside the laboratory. Those facts arrive only after real production begins.
This creates a missing institution in the industrial economy. Prototype shops are optimized for one impressive object. Contract manufacturers are optimized for a controlled design and a forecast. Dedicated plants are optimized for repetition. None is naturally organized around producing dozens of changing products in small but real batches, placing them into the world, measuring what happens, and rapidly expanding the few that work.
The factory should be the result of successful product search. It should not be the entrance ticket.
The machines required for this system already exist. They are distributed across machine shops, sheet-metal fabricators, industrial printers, electronics assemblers, composite specialists, moldmakers, test laboratories, finishing companies, and small contract manufacturers.
What is missing is not another universal machine. What is missing is a financial and operational abstraction over the machines.
The missing middle
Software companies can search cheaply. A team can deploy a new interface, route a fraction of users toward it, measure behavior, reverse the change, and try again. The company can maintain several versions at once and let real use reveal which one should survive.
A robotics company cannot do that as easily. Changing a robot arm may require a new toolpath, fixture, material order, inspection routine, assembly instruction, packaging insert, safety review, and field test. A product that performs well in simulation may fail because a technician cannot reach a fastener without removing three other components.
Physical reality is an expensive test environment. Hardware companies therefore tend to converge too early. They choose one architecture before they possess enough information, refine it into an attractive prototype, raise money around it, and discover during production that the prototype was only one hypothesis among many.
Production is therefore not merely what happens after research and development. It is part of research and development. Kenneth Arrow’s 1962 paper on learning by doing formalized the broader economic idea that productive knowledge emerges through activity and accumulated experience.1
The most attractive prototype can still be the wrong product. The real product often appears only after the first fifty units have been assembled, deployed, broken, repaired, returned, and redesigned.
The abstraction of production
Modern industrial history is usually narrated through the giant factory: Ford’s moving line, steel mills, shipyards, chemical complexes, and vertically integrated corporations. These institutions transformed established demand into enormous, repeatable flows.
But the assembly line depended on earlier abstractions. Parts had to become interchangeable. Dimensions had to become measurable. Drawings had to become transferable. Machine tools had to become accurate enough that one organization could make a component that would fit a component made somewhere else.
The decisive abstraction was the part. A dimensioned part could travel between organizations. A drawing could tell another shop what to make. A gauge could determine whether it had succeeded. Production became divisible before it became continuous.
-
1800s
Interchangeable parts
Dimensions, gauges, and machine tools let components cross organizational boundaries.
-
1913
Moving assembly line
Ford combined interchangeable parts, subdivided labor, and moving material for repetition at scale.2
-
1940s–70s
Job shops and numerical control
General-purpose shops acquired increasingly precise, programmable production capability.
-
1980s–2000s
Fabless-foundry split
Chip designers could access capital-intensive fabrication without owning a fab.3
-
2010s
Digital manufacturing networks
CAD-to-quote systems made fragmented capacity easier to discover and transact with.
-
Next
Industrial search foundry
Programs, evidence, and financing become composable—not merely individual part orders.
Ford’s moving assembly line arrived in 1913 after interchangeable parts and machine tools had matured enough to support it.2 The line was not a universal theory of production. It was an answer to a narrower question: how can a known object for a large market be produced repeatedly at falling unit cost?
It was not an answer to: which object should we produce?
Continuous selection
SHEIN is usually described as a fast-fashion company. A more useful description is that it operates a continuous product-selection system.
The company says it begins each item with an initial batch of roughly 100 to 200 pieces and uses observed demand to decide what to replenish.4 The important innovation is not the number itself. It is the loop:
A weak design does not receive the same inventory commitment as a strong design merely because both appeared in the original plan. Many designs enter. Most remain small. Some expand. The supply chain reallocates effort toward what the market is selecting.
Industrial products require richer evidence than a product-page click. A drone may need flight hours, repair events, battery-cycle data, payload performance, operator error, environmental exposure, warranty cost, and a paid second order. A robot gripper may need task completion rate, installation time, calibration drift, and changeover time.
The general principle still transfers:
Do not commit the full production system before the market has selected the product.
A search portfolio, not a single forecast
Illustrative allocation across four rounds. The number of programs falls while the batch behind each survivor grows.
The latent foundry
The search foundry should not begin by purchasing every machine. That would recreate the capital problem it is meant to solve.
The initial foundry is already distributed across existing firms. One shop has underused five-axis capacity. Another understands thin-wall aluminum better than anyone nearby. A sheet-metal company can turn revised enclosures in days. An electronics assembler can produce fifty boards without demanding an annual forecast. A composite specialist knows how to make a drone arm survive vibration. A test laboratory has the chamber and instrumentation required to expose a failure quickly.
This base is unusually fragmented. U.S. Census analysis found that 52 percent of employees in machine shops and closely related turned-product and fastener manufacturing worked at establishments with fewer than fifty employees.5 These shops are not obsolete remnants. They are distributed stores of machines, process knowledge, labor, local supplier relationships, and judgment.
The problem is that they remain difficult to compose.
Who can make this part, at this quantity, by this date?
Which product branch, process route, batch, test, and future commitment should be funded next?
Digital manufacturing networks have already made CAD-to-quote and supplier routing more legible. That is an important layer. But a quote marketplace clears an order. It does not know why the batch exists, what evidence it must produce, what should happen if the program succeeds, or which future capacity rights should be negotiated before success becomes obvious.
The next abstraction must be built over the program, not merely the part.
The Build Lot
Hardware is usually financed with objects that are too large or too late.
Equity finances a company. Debt finances an asset, receivable, or predictable cash flow. A purchase order finances a known delivery. A grant finances a declared project. None naturally represents:
Build thirty units of revision D, test four mechanical configurations under real customer use, terminate three branches, preserve all process data, and reserve the right to make 150 units of the winner.
That should become a recognizable economic object. Call it a Build Lot.
A Build Lot is not merely an order for parts. It is a bounded industrial experiment with a controlled revision, participating shops, manufacturing route, acceptance plan, field-test objective, cost ceiling, data rights, and explicit decision gates for the next batch.
The Build Lot as a contract primitive
Each lot converts capital into both physical units and a defined increment of evidence.
Some lots validate manufacturability. Others test customer behavior, reliability, serviceability, certification, channel economics, or supplier repeatability. A batch of twenty can be rational if every unit produces valuable operational evidence. A batch of five hundred can be irrational if no one has defined what the company expects to learn.
The relevant output is not only the inventory. It is the posterior belief about the product.
The financial stack
The institution around the Build Lot can be called an Industrial Search Facility. It commits capital to a portfolio of product programs and releases it in tranches as programs clear technical, production, and market gates.
It does not require one magical security. It requires a small stack of claims that separates risks currently bundled together.
Money as part of the production control system
Capital moves toward shops before the final product is proven. Evidence and future claims return as the program matures.
The batch receivable
The shop should not be forced to become a venture capitalist merely because its customer is early. Once a Build Lot is accepted, the shop holds an ordinary receivable from the facility rather than a speculative promise from a thinly capitalized startup.
The ramp option
The product company buys the right—but not the obligation—to expand into defined quantity bands, lead times, and pricing logic. The shop receives compensation for making future capacity legible without leaving machines permanently idle.
The process-participation claim
A shop that contributes a durable process improvement can receive a cash floor plus bounded upside if its method survives into later batches. This may be a per-unit kicker, ramp bonus, tooling buyout, or transfer payment when the process moves into a dedicated plant.
The tooling loan
Fixtures, molds, test rigs, dies, and custom automation sit between flexible production and the plant. They should be financed only when the expected unit-cost reduction justifies surrendering some flexibility.
The warranty reserve
A unit can pass outgoing inspection and fail months later. A portion of program economics remains in reserve until field reliability and returns mature, separating manufacturing acceptance from product performance.
When should a product leave flexible production?
Adjust an illustrative batch. The model compares flexible production with a dedicated tool. It is a decision aid, not a supplier quote.
A drone program
Consider an agricultural-drone company with eight employees. It has three airframe architectures, two battery enclosures, and several landing systems. Simulation suggests all are viable. The founders have enough capital to build an excellent demonstration aircraft, but not enough to explore the combination space through ordinary contract manufacturing.
Under the conventional sequence, the company selects one architecture, produces a polished prototype, raises capital, and orders one hundred similar units. It learns late.
Under the search-foundry sequence, the first Build Lot contains thirty production-intent drones distributed across three architectures. They are not laboratory toys. They use controlled components, serial numbers, work instructions, acceptance tests, and repair records.
A product earns specialization in stages
An illustrative drone program from fieldable prototype to a dedicated supply chain.
Production-intent variants
Three architectures across five farms. Learn repair access, handling, vibration, heat, dust, and operator behavior.
Selected architecture
Expand the serviceable design. Add fixtures, standardize wiring, redesign the sensor mount, and remove two weak branches.
Commercial pilot
Introduce bridge molding, an end-of-line test rig, standard harnesses, and customer deposits tied to field evidence.
Specialized production
Custom composites, dedicated testing, specialized motors, automated inspection, and a concentrated supplier network.
The first lot may reveal that the theoretically best airframe is difficult to repair in the field. A lower-performance architecture may survive rough handling and be serviced with ordinary tools. The second lot expands that branch. The third lot proves commercial repeatability. Only then does the program finance tooling that would have been reckless at the beginning.
The search foundry has not replaced the factory. It has discovered what factory deserves to be built.
The production graph
The deepest asset of the foundry would not be its list of shops. It would be the graph connecting product decisions to production and field outcomes.
For every serialized unit, the system should know the design revision, bill of materials, material lots, machine and process route, operators, work instructions, inspection results, accepted deviations, firmware, test conditions, deployment environment, failures, repairs, and subsequent engineering changes.
From CAD file to industrial causality
A useful system preserves not only what happened, but the context linking design, process, unit, field behavior, and the next decision.
Two parts with identical geometry can behave differently because of material condition, heat treatment, tool wear, print orientation, finishing, assembly torque, adhesive cure, supplier substitution, or a deviation that never reached the drawing.
Over time, the graph can answer questions ordinary sourcing systems cannot:
- Which enclosure features repeatedly create sealing problems?
- At what volume does this geometry usually move from machining to casting?
- Which shops produce reliable thin-wall parts under repeated revision?
- Which design changes reduce field repair cost without increasing assembly time?
- Which apparent customer problems are actually process-variation problems?
The platform accumulates a model of industrial causality. That model lowers the cost of the next company’s search.
Graduation, not permanent flexibility
Flexible manufacturing carries a premium. CNC machining can cost more per unit than casting. Additive production can cost more than molding. Manual assembly can cost more than a dedicated cell. Generalist suppliers cannot always match a plant designed around one stable product.
That is not a defect. The foundry is buying option value.
It pays more per unit to avoid committing too early to the wrong design, supplier, tool, or plant. The correct comparison is not between a machined part and a molded part after demand is certain. It is between flexible production and the expected cost of tooling the wrong product.
The industrial option ladder
As uncertainty falls, the product gives up flexibility in exchange for lower unit cost, tighter process control, and ownership of strategic bottlenecks.
General-purpose network
Maximum search breadth. High unit cost. Minimal fixed commitment.
Fixtures and recipes
Repeatability begins without fixing the entire product architecture.
Bridge tooling
Selected components trade option value for better economics.
Reserved supply chain
Volume commitments make supplier investment and lending possible.
Vertically integrated plant
Ownership internalizes mature process learning and strategic coordination.
Some products will never graduate. A specialized robot, research instrument, or replacement component may remain inside the flexible network indefinitely because customization and availability matter more than minimum unit cost.
Other products will leave quickly. A design that discovers a large market may move into molding, casting, automated assembly, custom electronics, and global suppliers within a few generations.
The platform should support both outcomes.
Failure modes
The search foundry can become a low-quality broker that routes work to the cheapest shop and loses process context between batches. It can become a consulting company whose economics require more humans for every program. It can transfer startup risk onto shops unable to carry it. It can reward experimental inventory without useful evidence. It can let founders choose vanity metrics that keep failed programs alive.
It can also mistake flexibility for permanent superiority and keep products in general-purpose production long after they should have graduated.
Supplier variation can overwhelm the learning signal.
A search system that cannot kill a branch becomes a subsidy.
Product IP, background process knowledge, and new process learning need separate rules.
Deferred claims require serial-level auditability and bounded waterfalls.
Safety-critical products cannot iterate as casually as consumer interfaces.
Option value becomes waste once product and demand are stable.
Build Lots therefore need independent acceptance criteria. Shops need a cash floor. Field tests need precommitted metrics. Warranty reserves need time to mature. Intellectual-property rules must distinguish background knowledge, product rights, and newly created process knowledge.
The thesis
The industrial economy contains more general-purpose capacity than small firms can effectively access. The machines are divided among thousands of shops. Expertise is divided among thousands of workers. Capital is divided among banks, venture funds, strategic customers, equipment lenders, public institutions, and suppliers.
The divisions are not the problem. The absence of an abstraction across them is the problem.
Raise capital → build plant → choose product → produce inventory → discover demand
Generate hypotheses → finance Build Lots → deploy → measure → expand winners → integrate the plant
Interchangeable parts made industrial work divisible. The assembly line made repetition economical. The semiconductor foundry separated design from ownership of fabrication. Digital manufacturing platforms made quoting and routing more legible.
The next institution should make physical product search financeable.
A small firm should not need to own a factory to discover a serious industrial product. It should need a well-specified experiment, access to qualified shops, and capital whose commitments expand only as the evidence improves.
The factory is what a successful product earns.
Sources and notes
- Kenneth J. Arrow, “The Economic Implications of Learning by Doing,” The Review of Economic Studies, June 1962.Source ↗
- The Henry Ford, “Henry Ford: Assembly Line.” The museum describes Ford’s 1913 combination of interchangeable parts, subdivided labor, and moving material.Source ↗
- Semiconductor Industry Association, “Semiconductor Industry Primer: The Stages of Production and Business Models.”Source ↗
- SHEIN Group, “Process.” SHEIN states that an item begins with a small initial batch of 100–200 pieces and is replenished according to demand.Source ↗
- U.S. Census Bureau, “Small Manufacturers Play a Big Role in the U.S. Economy,” November 16, 2017.Source ↗
All quantities, costs, funnels, and product-program examples in the figures are illustrative unless explicitly attributed. They are intended to explain the proposed institution, not to forecast a specific factory or investment return.