The view I stopped apologizing for
Off-the-shelf single-board computers are a great prototyping tool and a lousy production strategy. I'll go even further: in cloud and edge computing hardware for industrial automation and IoT, the "cheap" SBC is usually the most expensive part of the design.
I've been building industrial edge devices for eight years. I started in 2017 and spent the early years treating SBCs like the obvious choice. They're cheap, well-documented, and easy to order. Then I had a $12,000 eMMC disaster in March 2023. That was the trigger, not the only one. The honest lesson: a board's unit cost tells you almost nothing about its lifecycle cost.
For production devices, I now start with an ARM system-on-module (SoM) and design a custom carrier circuit board around the exact I/O. It means more work in engineering and a higher headline BOM. It also means fewer surprises, fewer truck rolls, and a board I can support for a decade.
The $12,000 mistake that changed my sourcing rules
In 2021, I chose a popular SBC for an edge controller. It was about $85 per board. It ran Linux, had Ethernet, and there were plenty of forum answers. All good.
Nobody stopped me to ask the obvious next question: will this exact SBC still be available in 2025? My senior engineer did warn me to check the lifecycle policy. I didn't listen. I thought "we'll handle that later." Later came in March 2023, when the SBC manufacturer changed the on-board eMMC part because of a supply shortage. The board looked the same. Basically, the storage changed, and nobody tested our write-heavy image against it. After about three weeks of continuous logging, a bunch of devices started corrupting filesystem blocks. Fourteen field units across three customer sites had to be replaced and reconfigured. The replacement units were the new revision, which meant the image we'd built was no longer good enough. Total cost, including travel, remote access, and overtime: close to $12,000.
(Should mention: the cost didn't include the damage to our reputation. One customer wanted to know why we couldn't "just swap the memory card." We're still recovering from that.)
That is the exact failure that turns SBCs from a bargain into a liability. An ARM SoM with a published longevity commitment — like the NXP product longevity program, which can cover 15 years for certain i.MX processors — is easier to defend in a design review. According to NXP, many i.MX products are committed for 15 years. You can build a carrier circuit board around that and not panic when a consumer SBC gets a silent revision.
The hidden I/O cost nobody puts in the quote
Here's the part that used to embarrass me. I would pick a $90 SBC, then add $50 of USB adapters, a DIN-rail mount, and an industrial power supply with weird connectors. By the time it was field-ready, the "cheap" SBC was a $200 assembly held together by adapter cables.
Industrial automation wants Modbus RTU, CANopen, analog inputs, isolated 24V digital I/O. A typical consumer SBC gives you USB, HDMI, and a header with I2C or SPI. So I used dongles. On paper, every dongle works. In the field, every dongle is a point of failure.
I remember one 60-unit order where we specified a USB-to-CAN adapter. It worked in our lab. In the field, three units came back in two weeks. The adapter was seated on a vertical USB port, and vibration was enough to make it unreliable. Shipping two technicians to a plant to rework three units cost more than the adapters did. I should add that this was entirely avoidable.
A custom carrier board solves this at the signal level. You route the SoM's UART, CAN, or SPI directly to industrial transceivers, isolated power, and terminal blocks. It costs more to design, but the per-unit BOM can be lower than the SBC plus dongles, if the volume is anywhere beyond a handful.
For example, we designed a carrier furnace board for a heat-treatment furnace controller. It took the ARM SoM and broke it out to twelve thermocouple inputs, six analog outputs, and two relay outputs. The customer didn't need a general-purpose computer. They needed a predictable controller. No off-the-shelf SBC had that I/O without a rack of external converters. The custom carrier furnace board was not the cheapest first-article cost, but it was super straightforward to scale.
Cloud and edge computing: support matters more than raw speed
Once the device is in the field, it's a cloud and edge computing node. It needs to send data, receive updates, and behave predictably when the network drops. The SBC's benchmark scores don't matter if the vendor's BSP doesn't support a kernel security fix.
I want to be fair here. Not all SBC manufacturers are the same. Some publish solid Yocto layers and answer support emails. But many are still marketing to hobbyists and prototypes. Their release cadence moves fast, and fast is not always good.
With ARM SoM platforms, the support model is different. According to Arm's SystemReady documentation (arm.com, SystemReady Program), SystemReady IR is designed to provide a standardized firmware and boot architecture on SoM-based systems. In plain English, when I move a carrier board from one qualified SoM to another, I have a much better chance that the same OS image boots without a painful bring-up. That's the kind of transparency I can put in a project plan.
Never expected the "boring" SoM vendor to be the one that answered questions like a teammate. Turns out their reference carrier board design was cleaner than anything the SBC community had. They sent the Altium files, the BOM, and a design review checklist. That was the moment I started thinking about total cost, not unit cost.
What about the "but I need it tomorrow" excuse?
I hear it from teams that have never designed a carrier board. They're worried about weeks of schematic work and layout. I understand. But honestly, the SBC path is only faster if you ignore the predictable rework. The client's deadline doesn't care how many dongles you planned to zip-tie to the enclosure lid.
What's actually slow in these projects:
- Getting the connector and pinout wrong on the first layout
- Handling a silent SBC hardware revision mid-project
- Waiting for a support answer from a forum thread
- Realizing the enclosure doesn't fit the cable bundle
A carrier board design has its own traps. Ask me about the 0402 resistor thing sometime. But at least it's your trap, with schematics you control. If I have a clear I/O list, a reference design from the SoM vendor, and a transparent supply chain, the first carrier board can be ready in about three to six weeks. I want to say it took me four weeks on our last project, but don't quote me on that; every design is different. That's still less time than I spent debugging one silly USB-to-serial issue after a field failure.
Transparency is a spec, not a slogan
The vendor who lists all fees upfront — even if the total looks higher — usually costs less in the end.
That quote is my entire pricing philosophy now. I've learned to ask "what's NOT included" before "what's the price."
- Does the SoM price include a reference carrier board schematic?
- Is the BSP maintained for the kernel version we plan to use?
- Are the pinout files in KiCad, Altium, or PDF?
- What happens when a component goes end-of-life?
Some vendors answer. Some don't. I want to be clear: I'd rather pay a slightly higher price to a vendor that answers these questions honestly than save 15% on a board that turns into an archaeology project.
The bottom line
I don't hate SBCs. I use them for prototypes, for internal tools, and for testing. But I don't trust them in production without a serious conversation about lifecycle, I/O, and support.
If you're building industrial automation and IoT gear that will touch a real production line, consider whether your hardware decision is transparent about the total cost. Start with the ARM system-on-module and the carrier board. It feels like more work at the beginning. In my experience, it's way less work than explaining a $12,000 mistake to your customer.
There's something satisfying about seeing a carrier circuit board pass thermal cycling on the first try. After the SBC failures, that feeling is underrated.
