The Decision That Sets Budget, Schedule, and Risk

Somewhere in every connected hardware project, someone asks whether to build the whole board or drop in a system-on-module and design only a carrier around it. The answer is not a matter of engineering pride. It is arithmetic across three variables: how much non-recurring engineering you can fund, how many units per year you will ship, and how much certification you want to inherit rather than earn.

What You Are Actually Buying With a Module

A SoM is a small board carrying the processor, memory, storage, and power management, exposing everything else on connectors or castellations. The physical parts are the least of what you are paying for.

  • The hard routing. DDR memory routing with length matching, impedance control, and an eight- or ten-layer stackup is genuinely difficult and unforgiving. On a module that work is done and proven.
  • Power sequencing. Modern SoCs need rails brought up in a specific order with specific timing. Getting this wrong bricks silicon, and the module vendor has already solved it.
  • A working board support package. A kernel that boots, drivers for the on-module peripherals, and an update path. This alone is often three to six engineer-months on a bare SoC.
  • Radio certification. If the module includes Wi-Fi or Bluetooth, it usually carries a modular grant you can reference.

What you do not buy is a product. The carrier board still needs power input protection, connectors, sensing, I/O conditioning, and an enclosure, and those are where your own PCB design mistakes will still find you.

NRE Versus Unit Cost: The Number That Decides

Run this before anything else. A Linux-class SoM sells for roughly $30 to $95 depending on memory and radios. The same silicon on your own board, at volume, lands somewhere around $18 to $45 in components. Call the delta $25 per unit as a working figure.

Against that, designing a custom board around an application processor costs $90,000 to $250,000 in engineering, plus two to three board spins at $8,000 to $20,000 each, plus BSP work, plus a longer schedule. Take $150,000 as a midpoint and the crossover sits near 6,000 units. If you will ship 1,500 units a year, the module wins by a wide margin and keeps winning. If you will ship 50,000, the custom board pays for itself in the first quarter.

Two adjustments matter. Schedule has cash value: a module typically pulls first revenue in six to nine months against twelve to eighteen for a custom design, and that gap often decides whether a company survives to the volume where the custom board would have paid off. And if the compute requirement is modest enough for a microcontroller rather than embedded Linux, the comparison changes entirely; see microcontroller vs embedded Linux and IoT product development cost.

Certification: The Module's Least-Discussed Advantage

A pre-certified radio module carries a modular approval that transfers to your product if you meet the grant conditions: the specified antenna or one on the permitted list, the specified trace layout at the antenna feed, no modification to the module, and the module's identifier displayed on your label. Honor those and you skip intentional radiator testing entirely, which is often $12,000 to $25,000 and six weeks.

Two cautions. The grant does not cover your finished product's unintentional emissions or immunity, so you still book a lab, and some grants are limited modular approvals with conditions that are easy to break, most commonly by changing the antenna. Read the grant, not the marketing page. The wider process is in FCC certification, and antenna placement stays your responsibility either way.

Supply Risk and Lifecycle

Industrial SoM vendors publish longevity commitments, commonly ten to fifteen years, and hold inventory across shortages because their whole business depends on it. A custom board exposes you directly to every allocation event on every part, including the DDR and PMIC that a module vendor buys in far larger quantity than you do.

The counterweight is concentration. With a module, one supplier failure stops your product completely, and there is no drop-in second source because pinouts are proprietary. Ask for the vendor's change notification policy, their end-of-life notice period, and whether they will sell a last-time-buy. Then check whether any competitor uses a compatible form factor: some ecosystems standardize connector and pinout definitions across vendors, which turns a single point of failure into a two-week port. The general playbook is in a component went end-of-life.

Where Modules Cost You

Height is the first constraint. A module plus its board-to-board connector adds 4 to 8 mm of stack, which a thin consumer product cannot spare. Board-to-board connectors also have limited mating cycles and can fret under vibration, so anything in a vehicle or a handheld tool wants a castellated solder-down module rather than a socketed one.

Thermal behavior is the second. The module's heat spreader was designed for the vendor's evaluation kit sitting in open air, not for your sealed enclosure, and thermal throttling under load is a common late surprise. Budget time for real thermal measurement in the actual housing.

The third is software lock-in. A vendor BSP built on an old kernel with out-of-tree patches can become a security liability you cannot fix. Check whether the vendor upstreams their support and how recently they shipped a kernel update before committing. The same class of risk applies to using consumer boards in products, discussed in using a Raspberry Pi in a commercial product.

Designing the Carrier So the Move Stays Cheap

Start with a module, plan for the possibility of a custom board, and the transition costs a fraction of a fresh start. Keep every carrier-side interface on standard buses rather than module-specific pins. Partition the schematic so power, I/O conditioning, and connectors survive a compute change. Write firmware against a hardware abstraction layer. Keep an accurate power and thermal budget throughout, because that is the data you need to specify the custom design.

Move when two conditions hold together: the volume is real and forecast for at least eighteen months, and the requirements have stopped changing. Migrating while the spec still moves means paying custom-board NRE twice. That checkpoint usually lands around the DVT and PVT builds of a second-generation product, not the first.

Run the Numbers on Your Own Volumes

Projects House takes this decision apart with real quotes rather than rules of thumb: module shortlist, carrier design, custom-board cost model, and a migration plan if the volume justifies it later. Send your compute requirements and annual forecast through our contact form.