ProjekinAja. Embedded Systems Stage 6 · Bring-Up & Deployment Lesson 6.8
● Stage 6 · Lesson 6.8

Cost, BOM, and Scaling From 25 to 25,000

At 25 units the engineer's time is the cost. At 25,000 units a two-cent resistor is a decision.

⏱ ~26 min read 💰 Cost engineering 🧭 Beacon Rev A

🎯 By the end you'll understand

Beacon's real bill of materials, line by line, and why requirement R5 was both missed by one per cent and measuring the wrong thing. How non-recurring cost — stencils, jigs, tooling, certification, and above all engineering hours — divides by volume, and why at 25 units certification alone costs roughly forty-six times the parts. Where price breaks actually come from and what changes between 25, 2,500, and 25,000. How lead times, allocation, and end-of-life notices turn a cheap part into a stopped production line, and what second sourcing costs when you try to design it in on a 35 × 25 mm board. What a respin costs at each volume, and why over-the-air updates are the highest-return line item in the whole requirements table.

KEY INSIGHTUnit cost is a function of volume, and so is which costs are real

Ask "what does a Beacon cost?" and the honest answer is a curve, not a number. At 25 units it costs about eleven and a half million rupiah. At 25,000 it costs about a hundred thousand. The hardware is identical.

That is not a rhetorical trick; it is the arithmetic. Twenty-five units carry the same certification, the same test jig, the same stencil, and the same four hundred and twenty hours of engineering as twenty-five thousand would. Divide those fixed costs by twenty-five and each unit inherits an enormous share. Divide them by twenty-five thousand and they nearly vanish, leaving parts and assembly — which by then have fallen by half through volume pricing.

The practical consequence is that cost-reduction effort has to be aimed at the volume you are actually at. At pilot scale, agonising over a two-cent resistor is theatre; the same afternoon spent reducing engineering hours or avoiding a respin is worth thousands of times more. At twenty-five thousand units the polarity reverses completely: an engineer's week costs about the same as a two-cent part across the run, so a week spent removing two cents is break-even and a week spent removing ten cents is obviously correct. Teams get this backwards constantly, usually because the BOM is the number that is easy to see and the engineering hours are the number nobody counts.

🔑 Hold onto this

Three cost layers behave differently and must be tracked separately. Recurring per-unit cost — parts, PCB, assembly, enclosure — falls with volume along price breaks and is the only layer that ever appears on a BOM. Non-recurring cost — tooling, stencils, jigs, certification, engineering — is fixed and divides by volume, dominating everything below a few thousand units. Risk cost — lead time, allocation, single sourcing, end-of-life, currency — has no line on any spreadsheet and is the one that actually stops production. A cost model with only the first layer is the model that gets a company into trouble.

CORE CONCEPTBeacon's actual bill of materials

Here is Rev A's bill at the pilot's 25 units, in rupiah at an assumed IDR 15,800 to the US dollar. Prices are indicative distributor and PCB-house quotes at that quantity, not contract prices, and the currency assumption is itself a cost risk discussed below.

LinePart@25 (IDR)@2,500@25,000Share @25
MCU moduleESP32-C3-MINI-136,34029,23024,49024.0 %
Temp / humiditySHT40-AD1B37,92024,49018,96025.0 %
AccelerometerLIS2DH1217,38011,3809,16011.5 %
Buck regulatorTPS6284015,01010,4308,2209.9 %
ChargerMCP738318,6906,0004,4205.7 %
PCB4-layer, 35 × 25 mm, ENIG17,3806,6404,74011.5 %
ProtectionDW01A + FS8205 + PRTR5V0U2X5,3723,3202,3703.5 %
Connectors & I/OUSB-C, JST-PH, button, WS2812B7,5844,6603,3205.0 %
Discretes3 × PMOS, inductor, ≈32 passives5,8463,3202,2903.9 %
BOM totalparts + bare PCB151,52299,47077,970100 %
Assemblyoverseas turnkey, ≈55 placements44,0009,2806,030—
Landed costBOM + assembly, excluding enclosure195,522108,75084,000—

Two facts jump out of that table. The first is arithmetic: IDR 151,522 against a requirement of IDR 150,000 is a miss by one per cent. Not a dramatic failure, and entirely fixable — dropping the accelerometer's grade or moving to a two-layer board would do it, and both cost something the requirements also asked for. R5 is missed, and the post-mortem says so.

The second fact is more interesting and less comfortable. R5 was written as a bill of materials number: parts plus the bare board. It says nothing about assembly, and assembly at 25 units is IDR 44,000 — twenty-nine per cent on top of the number the requirement measures. The landed cost of a Beacon at pilot volume is nearly two hundred thousand rupiah, and everyone involved remembers the figure as "about a hundred and fifty." That gap is not dishonesty; it is a requirement that measured the quantity that was easy to state rather than the quantity that matters. Rev B's requirements change is to define R5 as landed cost at a stated volume, which is a one-line edit and would have changed several decisions in Stage 2.

⚠ Two parts are half your bill, and neither has a second source

The module is 24 % of the bill and the SHT40 is 25 %. Together they are 49 % of the BOM, and both are effectively single-sourced. The SHT40 has functional alternates that meet requirement R3's ±3 %RH — other digital temperature-and-humidity sensors clear it comfortably — but none shares its footprint, so switching means a board change. The module has no drop-in alternative at all: a different module means a different footprint, a different antenna, a different RF layout, and different regulatory evidence. This is the single most important sentence in Beacon's cost model, and it is not a cost sentence. It is a risk sentence, and Lesson 6.9's list of what a real product would still need starts here.

HOW IT WORKSNon-recurring cost: the part that divides

Non-recurring engineering cost — NRE — is everything you pay once regardless of how many units you build. Its per-unit contribution is simply the total divided by the volume, and that division is the most dramatic arithmetic in hardware.

NRE itemIndicative costPer unit at 25Per unit at 2,500Per unit at 25,000
Stencil for assembly400,00016,00016016
Test jig (Lesson 6.3)950,00038,00038038
Injection-mould tool (Lesson 6.5)87,000,0003,480,00034,8003,480
Certification: FCC + CE + SDPPI174,000,0006,960,00069,6006,960
Engineering — 420 h across six stages105,000,0004,200,00042,0004,200
Total NRE, excluding the mould tool280,350,00011,214,000112,14011,214

Read the "per unit at 25" column again. Certification alone contributes almost seven million rupiah to each pilot unit — roughly forty-six times the entire bill of materials. Engineering time contributes another four point two million, about twenty-eight times the BOM. The parts, the thing everyone optimises, are under two per cent of what a pilot unit truly costs. This is why the dek says what it says, and it is why a small hardware team that spends a week saving five thousand rupiah of BOM at pilot scale has spent about a quarter of a million rupiah of engineering to save one hundred and twenty-five thousand across the batch.

Now read the "at 25,000" column. Total NRE falls to just over eleven thousand rupiah a unit against a landed cost of eighty-four thousand — meaningful but no longer dominant. The mould tool that was absurd at 25 is a rounding error, and the certification that was ruinous is under seven per cent of a finished unit. Everything flips, and the flip happens somewhere between one and five thousand units for a product shaped like this one.

💡 Count engineering hours, or the model is fiction

Engineering time is the NRE line nobody puts in the spreadsheet, because it feels like it is already being paid for. It is not free, and at pilot volumes it is usually the largest single item. The number above — 420 hours across six stages, at a loaded rate of IDR 250,000 an hour — is deliberately modest for a project of this scope, and it still produces IDR 105 million. Count the hours. Then, when someone proposes a design change that saves two hundred rupiah of BOM at 25 units and costs a day of layout, the conversation has numbers in it: two hundred rupiah times twenty-five is five thousand rupiah of saving against two million rupiah of engineering. That conversation takes thirty seconds and it is one nobody has, because the second number was never written down.

SEE ITExplore the cost curve — and find where the money and the risk actually live

Set the volume, choose how the boards get assembled, decide whether to design in a second source, pick a certification path, and choose the enclosure route. The chart shows landed cost and total cost per unit across the whole volume range, with your chosen volume marked. Watch which decisions matter at 25 units and which only matter at 25,000 — and watch the lead-time exposure, which does not care about volume at all.

⚙ Interactive · Cost curve explorer

Prices are indicative Indonesian quotes at IDR 15,800 to the dollar. Certification, tooling, and engineering figures are illustrative order-of-magnitude estimates. The selling price is fixed at IDR 450,000 so the margin readout means something.

total incl. NRE landed (parts + assembly + case) production volume (log) — cost per unit in IDR (log)
BOM / unit
—
Landed / unit
—
NRE / unit
—
Margin at IDR 450k
—

Four things to try. At 25 units, switch assembly to in-house hand building: the landed cost rises sharply, because fifty minutes of your time per board is the most expensive assembly method available and the only one that feels free. At 25 units, turn on the full certification path and watch total cost per unit pass eleven million rupiah — the number that tells you a 25-unit pilot must be internal, not sold. At 2,500, switch the enclosure to injection moulding and watch it become the right answer. And at every volume, turn on the dual footprint for the sensor: it costs about two thousand six hundred rupiah a unit for a slightly larger board and it is the only control on the page that changes the lead-time exposure, which is the risk that actually stops a production line.

COMPARISONLead times, allocation, and the parts that disappear

The cost model above assumes you can buy the parts. That assumption failed spectacularly and industry-wide between 2021 and 2023, when lead times on many microcontroller and analogue lines stretched well past fifty-two weeks, and it fails locally and quietly all the time. Four mechanisms are worth understanding because each has a different countermeasure.

MechanismWhat happensCountermeasureWhat it costs you
Lead timeDistributor stock runs out and the factory quotes 20, 40, or 52-plus weeksBuy ahead against a forecast; keep safety stock of the long-pole partsWorking capital tied up in inventory, and obsolescence risk if the design changes
AllocationDemand exceeds supply and the manufacturer allocates against historical purchase volume. A first-time buyer of 2,500 pieces gets nothing at allBuild a purchase history before you need one; work through a distributor who will advocate for youPaying more, earlier, for parts you do not yet need
End of life and product change noticesThe manufacturer announces a change, or an EOL with a last-time-buy window commonly around six to twelve monthsSubscribe to change notifications for every part; check lifecycle status (Active / NRND / Obsolete) before designing anything inAlmost nothing, if done at design time. A respin, if not
CounterfeitsShortage drives buyers to brokers, and the market fills with relabelled, recovered, or fake parts that pass a visual checkBuy from authorised distribution or the manufacturer, full stopPaying list price during a shortage instead of a broker's price
CurrencyA rupiah-priced product with a dollar-priced bill of materials moves with the exchange ratePrice with a currency assumption stated, and review it; hold some inventory as a natural hedgeMargin volatility that has nothing to do with engineering

The currency line deserves emphasis for a product built in Indonesia. Beacon's BOM is quoted in dollars and its requirement is in rupiah. At IDR 15,800 to the dollar the BOM is 151,522 and misses R5 by one per cent. At IDR 17,000 the same parts cost 163,030 and miss by nine per cent, with no engineering change of any kind. If a requirement is written in local currency against a dollar-denominated bill, the exchange rate is a silent requirement and it belongs in the assumptions column.

Second sourcing on a 35 × 25 mm board

The textbook answer to single-source risk is to qualify a second part and design its footprint in alongside the first, populating whichever is available. On Beacon that runs straight into requirement R8, which caps the board at 35 × 25 mm. A dual footprint for the humidity sensor needs roughly a further 12 mm² of board plus its own routing, and the board has to grow to about 38 × 25 mm to absorb it — which breaks R8 as written, or forces something else off the board. The choice is therefore explicit: relax a size requirement, drop a feature, or accept that a quarter of the bill of materials has no alternative. Beacon's Rev A took the third option without ever writing it down, which is the part worth criticising. Taking a risk deliberately and taking it by omission look identical until the day the part goes on allocation.

The module is harder still. There is no pin-compatible, footprint-compatible, similarly-certified alternative to the ESP32-C3-MINI-1 that you can populate on the same pads. A second source there means a second board variant, a second RF layout, and a second set of regulatory evidence — a project, not a purchasing decision. The honest mitigation for a single-sourced module is not a second source; it is inventory, a relationship with a distributor, and a design that keeps the module's interface narrow enough that a future variant is a board respin rather than a firmware rewrite.

IN PRACTICEWhat a respin costs, and why R7 is the best line in the requirements table

Every defect on Rev B's list has a cost that depends entirely on when it is found and how many units exist. The same fix — say, moving the button off GPIO9 — has four different prices.

1

Found before fabrication: hours

A layout change, a re-run of design rule and manufacturability checks, and a new set of Gerbers. Cost: an afternoon. This is what the review in Lesson 5.9 was for, and it caught the GPIO9 issue and accepted it — which was a decision, not a miss, and it is worth noting that the decision was made without knowing the field consequence.

2

Found at 25 units: about IDR 5 million and three weeks

Twenty-five new boards, twenty-five new assemblies, the shipping, and the calendar. The old boards become bench units. This is the cheapest a hardware defect ever gets after fabrication, and it is the entire economic argument for a pilot: you are buying information at the only volume where information is cheap.

3

Found at 2,500 units: inventory, re-test, and possibly re-certification

Two thousand five hundred landed units is roughly IDR 272 million of inventory that is now wrong. Reworking may or may not be possible; scrapping usually is not affordable. If the change touches the RF layout or the antenna environment, the module's integration conditions are back in play and the certification evidence may have to be redone — which is another IDR 174 million and eight weeks. The cost of a defect has jumped by two orders of magnitude, and most of the jump is not the boards.

4

Found at 25,000 units in customers' hands: a different kind of conversation

Now the cost includes retrieving units, or not retrieving them and living with the failure rate, plus the commercial consequences and whatever your contracts say. No engineering number covers this. What is worth internalising is only that the curve is steeply superlinear, which is why the two cheapest activities in this entire course — a design review and a 25-unit pilot — have returns nothing else comes close to.

🔑 The requirement that pays for itself

Requirement R7 asked for over-the-air updates with rollback, and at the time it looked like a convenience. On this cost curve it is the highest-return line in the table, because it moves an entire class of fixes off the hardware curve and onto the firmware curve — where a defect found at 25,000 units costs a build, a staged rollout, and a monitoring window rather than a recall. Of Rev B's changes, look at which ones are firmware: the fresh-press gate on the update path, the motion threshold and duration, the telemetry re-allocation, the low-battery cutoff. Four of the twelve changes can reach deployed units without touching a board. That is what R7 bought, and it is why "we'll add updates later" is one of the more expensive sentences in hardware.

BEACONThe verdict on R5, and what a Beacon really costs

Requirement R5: unit BOM cost at 25 units of no more than IDR 150,000, excluding the enclosure. The measured figure is IDR 151,522. R5 is missed by 1.0 %, and the post-mortem records it as a miss rather than rounding it away, for a reason worth stating: a requirement that is quietly relaxed once has stopped being a requirement.

QuestionAnswer at 25 unitsAt 2,500At 25,000
BOM, parts and bare PCB (what R5 measures)151,52299,47077,970
Landed, plus assembly195,522108,75084,000
Plus the enclosure248,522148,05091,980
Plus amortised NRE, full certification path11,462,522260,190103,194
Gross margin at a selling price of IDR 450,000deeply negative42.2 %77.1 %

The bottom two rows are the honest business answer to "is this a product?" and it is not a question a course on embedded systems can dodge, because the answer changes what you should build. At 25 units, with certification, a Beacon costs roughly eleven and a half million rupiah. No price recovers that; the pilot is a research expense, and pretending otherwise is how a company mistakes a prototype for a business. At 2,500 units the same device is around two hundred and sixty thousand rupiah all-in against a four hundred and fifty thousand rupiah price, which is a real if unexciting margin. At 25,000 it is comfortable. The volume at which this product becomes viable is therefore somewhere between one and two thousand units, and that number should have been derived in Lesson 1.9 alongside the requirements rather than in Lesson 6.8 alongside the post-mortem.

🧪 What would actually get R5 back under target

Four candidates, with their costs stated honestly. Drop to a two-layer board: saves about IDR 7,000 at 25 units and breaks the RF layout that requirement R2 depends on and the module's integration conditions assume. Rejected. Substitute a lower-grade humidity sensor: saves IDR 12,000 to 18,000 and still clears R3's ±3 %RH, at the cost of accuracy margin you may want later — a genuine option, and the one a cost-down would take. Remove the RGB LED and its high-side switch: saves about IDR 2,500 and removes the only local status indication, which the pilot showed technicians using constantly. Rejected. Renegotiate R5 to a landed-cost basis at a stated volume: costs nothing, changes no hardware, and makes the requirement mean what everyone thought it meant. That last one is the recommendation, and it is a requirements change rather than a design change — which is the correct kind of fix for a requirement that was measuring the wrong quantity.

WEAKNESSWhere this bites

✗ Optimising the BOM at pilot volume

At 25 units, parts are under two per cent of the true cost of a unit. A week spent shaving five thousand rupiah from the bill saves one hundred and twenty-five thousand rupiah across the batch and costs about two million rupiah of engineering time. The same week spent on the design review that prevents one respin saves five million. This is not an argument for carelessness about cost; it is an argument for aiming cost-reduction effort at the volume you are actually at, and for revisiting the whole BOM properly when a real volume commitment exists — at which point the price breaks are different anyway and the earlier work would have been redone.

✗ Treating single sourcing as a cost decision

The cheapest part is usually the one with no alternative, and choosing it looks like good cost engineering right up until the line goes to allocation. Beacon's two largest lines are both single-sourced, and that fact appears nowhere in the bill of materials, because a BOM has a column for price and no column for "what happens if this is unavailable for nine months." Add the column. For each line, record the lifecycle status, the current lead time, and whether an alternate exists — and if none does, say so in the design review rather than discovering it during a shortage. The mitigation is often cheap at design time (a dual footprint, a slightly larger board) and impossible afterwards.

✗ Quoting a unit cost without a volume attached

"A Beacon costs a hundred and fifty thousand rupiah" is a sentence with no truth value, and it will be repeated in meetings until somebody builds a business plan on it. Every cost figure needs three qualifiers: at what volume, including what, and under what currency assumption. The version that survives contact with reality reads more like "IDR 195,500 landed at 25 units, excluding the enclosure and all non-recurring cost, at IDR 15,800 to the dollar." It is longer, it is less quotable, and it is the only version that will not embarrass you later.

KEY INSIGHTHow this connects

Cost is the discipline that decides whether any of the preceding six stages produced a product or an artefact. It is also the one where an engineer's instincts are least reliable, because the intuition that a cheap part is a good decision was formed at a volume nobody is at.

GLOSSARYKey terms introduced

Bill of materials (BOM) cost
The cost of parts plus the bare board, per unit, at a stated quantity. It excludes assembly, test, packaging, and everything non-recurring — which is why a BOM figure alone systematically understates what a unit costs.
Landed cost
BOM plus assembly, test, and shipping — the cost of a finished unit arriving at your door, before any fixed costs are amortised. The number a requirement should be written against.
Non-recurring engineering (NRE)
Costs paid once regardless of volume: stencils, jigs, tooling, certification, and engineering hours. Divided by volume, it dominates unit cost below a few thousand units and becomes negligible above tens of thousands.
Allocation
The state in which demand exceeds a manufacturer's supply and parts are distributed according to purchase history. New buyers, however urgent, receive nothing — which makes purchase history itself a supply-chain asset.
Last-time buy
The window, commonly six to twelve months, following an end-of-life notice during which a part may still be ordered. Missing it converts a component change into a board respin at whatever volume you happen to be.
NRND
"Not recommended for new designs" — a lifecycle status meaning the part is still supplied but should not be designed in. Checking it costs a minute and is skipped constantly.
Second sourcing
Qualifying an alternative part and, ideally, laying its footprint in alongside the primary so either can be populated. It costs board area, which on a size-constrained design is a direct conflict with a requirement.
Respin
A new revision of the board and a new build. Its cost is superlinear in volume: hours before fabrication, millions at pilot scale, hundreds of millions plus re-certification at production scale.

ACTIVE RECALLCheck your understanding

Answer in your own words first, then reveal.

1A teammate proposes spending three days replacing three parts to save IDR 6,000 of BOM. The pilot is 25 units. Evaluate the proposal, then evaluate it again at 25,000 units.›

At 25 units: the saving is 6,000 × 25 = IDR 150,000 across the entire batch. Three days of engineering at a loaded IDR 250,000 an hour is roughly IDR 6 million. The proposal destroys about forty rupiah of value for every rupiah it saves, and it also consumes three days of calendar during which the pilot is not deployed and no information is being bought. Reject it, and say why in those terms rather than as a matter of taste. At 25,000 units the same saving is IDR 150 million against the same IDR 6 million of engineering — a twenty-five-fold return, and obviously correct. The parts did not change and neither did the engineer; only the divisor did. The transferable rule: multiply any per-unit saving by the volume you have actually committed to, compare it with the fully-loaded cost of the hours, and be suspicious of any cost-reduction argument that never states the volume.

2R5 asked for a BOM under IDR 150,000 at 25 units, excluding the enclosure. The measured BOM is IDR 151,522 and the landed cost is IDR 195,522. Which number should the post-mortem report, and what should change?›

Both, with the discrepancy named. Against the requirement as written, the result is IDR 151,522 and R5 is missed by one per cent — reported as a miss, because a requirement quietly relaxed once has stopped functioning as a requirement. But the more important finding is that R5 measured the wrong quantity: it excluded assembly, which at pilot volume is IDR 44,000, twenty-nine per cent on top. Everyone involved remembers "about a hundred and fifty thousand" for a unit that actually costs nearly two hundred thousand before the enclosure and before any fixed costs. The change is to redefine R5 as landed cost at a stated volume with a stated currency assumption — a one-line requirements edit that costs nothing and would have altered several Stage 2 part choices, because assembly cost scales with placement count and a requirement that ignores placements does not discourage adding parts.

3Beacon's module and humidity sensor are 49 % of the BOM and both are effectively single-sourced. What does that cost, and what would fixing it cost?›

It costs nothing on the bill of materials and everything on the risk side, which is precisely why it is invisible — a BOM has a price column and no availability column. If either part goes onto allocation or receives an end-of-life notice, production stops, and neither event gives useful warning to a customer with no purchase history. Fixing it differs by part. The sensor has functional alternates that clear R3's ±3 %RH, so a second source is a dual footprint on the board: roughly 12 mm² more area, which pushes the board to about 38 × 25 mm and breaks R8's size limit as written. That is a real trade — relax a requirement, drop something else, or accept the exposure — and the fault in Rev A is not that the exposure was accepted but that it was accepted silently. The module is worse: there is no footprint-compatible, similarly-certified alternative, so a second source means a second board variant, a second RF layout, and a second set of regulatory evidence. For that one the realistic mitigation is not a second source but inventory, a distributor relationship, and keeping the firmware's interface to the module narrow enough that a future change is a respin rather than a rewrite.

4Why does requirement R7 — over-the-air updates with rollback — have the highest return of anything in the requirements table, on cost grounds alone?›

Because it moves an entire class of defects from the hardware cost curve to the firmware cost curve, and those two curves diverge violently with volume. A hardware defect costs hours before fabrication, about five million rupiah at 25 units, hundreds of millions plus possible re-certification at 2,500, and something no engineering number describes once units are with customers. A firmware defect, on a fleet with working updates and rollback, costs a build, a staged rollout, and a monitoring window at every one of those volumes. Look at Rev B's list: the fresh-press gate on the update path, the pin hold before deep sleep, the motion threshold and duration, the telemetry re-allocation, and the low-battery cutoff are all firmware. Five of nine changes reach deployed units without touching a board. Rollback matters as much as the update mechanism, because an update path without it converts one bad build into a fleet-wide brick — which is the failure mode that makes teams afraid to deploy, and an update mechanism nobody dares use has all of the cost and none of the return.

Next: thirty days in a real workshop, reported honestly →