Industrial PC ODM, Smart Home PCB and PCB Fabrication: What Through-Hole Work Returns

The DIP assembly line is the part of an electronics build that refuses to disappear. Boards loaded with connectors, transformers, relays and high-power parts still go through a through-hole line long after the surface-mount side of the factory has been automated twice over, and the reason is physical rather than nostalgic. What surprises most purchasing teams is not that the process survives but how badly its cost is modelled. Quotations arrive as a per-joint figure, which then disappears into a build where insertion, wave soldering, touch-up and inspection each contribute hours. I have audited enough of these builds to know that the number people argue about in a meeting is rarely the number that decides whether the program is profitable. The three services that matter most in that audit are PCB Fabrication at the front end, Custom Industrial PC PCBA ODM in the middle, and Custom Smart Home Products PCB at the volume end, because each one shifts the arithmetic in a different place. The framework below is the one I use to work out what the line actually returns, which is a different question from what it costs per joint.

Counting What the Line Actually Costs

The first mistake is accepting a per-joint price as the cost of the process. A through-hole line consumes four distinct resources, and only one appears on the quotation. Insertion labour is the obvious item, but board handling, wave-solder consumables and joint inspection all sit behind it. When a program runs a few hundred boards those items are small enough to ignore. At tens of thousands they become the dominant term, and a quotation that looked competitive at prototype stage turns into a forecast nobody can defend. I start by writing the four categories down and forcing a number against each, even a rough one. A visible estimate is more useful than a precise figure that hides two categories inside a single line.

The cost stack above is the first thing to establish, because every later decision depends on it. Once handling, consumables and inspection are separated from insertion labour, the argument stops being about the quoted price and starts being about the build.

Labour Hours Are Not the Whole Story

The second correction concerns labour, which is usually the only number anyone tracks and almost never the deciding one. Insertion time is heavily influenced by board layout, and a layout that spreads through-hole components across the panel can add handling time that no amount of operator skill removes. This is where the front end of the process quietly sets the ceiling on the back end. A fabrication step that groups connectors and keeps through-hole parts to one region gives the assembly line something it can optimise. A design that treats placement as free will be paid for twice: once in fabrication and again in every board that follows. The analysis question is not how fast an operator can insert, but how much of the insertion is determined before the board ever reaches the line.

Rework and Scrap as Hidden Line Items

The third correction is the one that changes conclusions most often, because scrap and rework sit outside the standard cost model by default. A wave-soldered board can emerge with bridges, insufficient fill or thermal damage, and each outcome has a recovery cost that depends on the board's value at that point. On a simple panel rework is cheap; on an industrial control board a single unrecoverable defect can erase the margin on several good units. The practical consequence is that yield deserves to be treated as a cost driver rather than a quality metric. A slower line with a stable first-pass yield will out-return a faster one that generates rework, and the difference appears only when recovery cost is quantified.

Where Volume Changes the Math

The fourth factor is volume, and it does not behave the way most forecasts assume. Through-hole assembly has a fixed component in setup, stencil preparation and line configuration, and a variable component that scales with joints and boards. Below a certain batch size the fixed element dominates, which is why prototypes look expensive per unit no matter who builds them. Above that size the variable element takes over and the curve flattens, but it flattens around a level that depends entirely on the first three factors. Custom Industrial PC PCBA ODM programs tend to sit near the boundary, running in quantities large enough to amortise setup but varied enough that the line is reconfigured often. The analysis has to reflect that reality rather than a clean single-volume assumption, because the mix is what the factory actually experiences.

The industrial build above is the case that makes volume modelling difficult, since the line changes over more often than a pure consumer program would. Treating it as a single steady volume is the fastest way to produce a forecast that will not survive contact with the schedule.

Tooling and Fixture Amortisation

The fifth item is tooling, which is the easiest cost to forget because it is paid once and remembered never. Wave solder pallets, fixtures and any custom handling equipment are bought for a program and amortised across its volume, which means the effective per-unit figure depends on a production run that has not happened yet. If the program grows, the tooling becomes trivial. If it is cancelled at a third of the forecast, the tooling is a loss that was never in the piece price. I prefer to show amortisation on a separate line rather than folding it into a unit cost, because the line makes the risk visible. Teams that see the assumption plan reconfiguration better than teams that inherit a blended number.

A Payback Model for Mixed Builds

The sixth consideration separates useful models from tidy ones: real factories run mixed builds, and the mix usually determines payback. A line alternating between industrial control boards and consumer products carries setup costs across both, and the cheaper program can subsidise line time for the more demanding one. This is why a single-product calculation overstates the case. The model I trust allocates setup, rework and tooling across the actual sequence rather than across an average, and then checks whether the answer still holds when the sequence changes. Programs built on Custom Smart Home Products PCB work illustrate the point well, because household volumes are seasonal in a way that industrial schedules are not. A payback figure that ignores the seasonality is a figure that will be wrong for half the year. The EMS Electronic Manufacturing view of the same factory is a useful cross-check, since it prices the whole build rather than one process step.

The mixed build above is where allocation choices become visible, because seasonal demand and industrial schedules share the same line. Modelling them separately produces two comfortable numbers and one uncomfortable factory.

What the Numbers Look Like Side by Side

The table is ordered from the most visible cost to the least, which is also roughly the order in which the items get forgotten. Reading down the right-hand column gives the failure modes that show up most often in post-mortems, and every one is a modelling gap rather than a manufacturing problem. A calculation covering all six rows can be defended in a review, because the assumptions are inspectable.

Deciding with Imperfect Data

The last point is uncomfortable but necessary: the data needed for a perfect return model does not exist before the program runs. Yield estimates come from comparable builds, tooling costs are quoted rather than incurred, and volume forecasts are commitments that may not hold. That does not make the exercise worthless; it changes what the exercise is for. The value of the model is not the number it produces but the sensitivity it exposes: a short list of assumptions that would have to be wrong before the decision changes. A related analysis of through-hole demand reaches a similar conclusion from the process side. What I would avoid is deferring the decision until the data improves, since on a mixed line the data improves only by running the program. The deciding question is not whether the model is exact. It is whether the assumptions are written down where someone can challenge them.