中文
English
Español
Pусский
한국어
日本語

Industrial Router Procurement in 2026: Four Specs That Memory Prices Just Reordered

2026-11-03

For a decade, industrial router selection ran on one piece of folklore: the higher the clock speed, the better the box; the more RAM, the more stable the deployment. That advice cost nothing while memory was cheap. The extra money bought a nicer-looking BOM and not much else.

2026 is not that year. Public tracking from TrendForce, Sourceability and others shows DRAM contract prices more than doubling since 2025, with NAND flash booking double-digit—sometimes 50%-plus—quarter-on-quarter increases, and NAND contract pricing still expected to rise another 15–20% in Q4 2026. On a unit whose entire BOM is a few dozen dollars, memory and flash have moved from rounding error to dominant line item. And an unpredictable one.

The consequences are plain. The same datasheet may quote 30% higher this year. To hold the price, suppliers quietly swap part numbers. So the real procurement question is no longer "are the specs good enough"—it is "which workload am I actually paying for when I buy the extra RAM?"

Here is how the four specs should now be ranked, with criteria you can put straight into a technical agreement.

1. RAM and flash configuration: size to the workload, not to the brochure

Industrial router memory goes three places: packet buffers during forwarding, the resident OS and processes, and local workloads—protocol parsing, caching, containers. The first two are essentially fixed cost on a mature platform. A 5G Wi-Fi 6 industrial router running full forwarding, VPN and serial pass-through is comfortable with 512MB of DDR.

The third bucket is where budgets diverge, and it depends on your operation, not on the vendor's marketing.

A useful filter: ask whether this device has a second job. If it only ships field data back to a platform, 256–512MB RAM with around 128MB of flash is the sane band. If it must buffer several days of data locally, parse protocols, or host a small HMI page, then step up.

Putting that cache in flash rather than RAM is the cheapest move available in this cycle. NAND costs far less per gigabyte than DRAM, and a store-and-forward cache never needed RAM-grade random access anyway.

Take the ComIn I5800 from Starlinkun: 512MB DDR3L, 32MB SPI NOR plus 128MB NAND, paired with 5G SA/NSA, dual-band Wi-Fi 6, gigabit ports and RS232/RS485 pass-through. That covers the vast majority of connect-and-report field deployments without paying for capacity nobody uses. For lighter telemetry, the ComIn I5000 industrial gateway runs 256MB DDR3L with 32MB SPI flash—half the RAM, a visibly lower price, and no problem handling the same MQTT or HTTP uplink load.

One line worth adding to your spec: require the supplier to state memory and storage type (DDR3L, DDR4, LPDDR4X; NOR, NAND, eMMC) and approved alternate part numbers, not just capacity. During an allocation squeeze, two boxes labelled "512MB" can carry very different costs.

2. Container support: it decides the device's second life, not its benchmark score

Containerisation gets marketed as an advanced feature. In practice it changes three things.

  1. Iteration speed. Once a device is in the field, the most expensive action is sending someone to site. Pushing a canary update to a single parsing container is an order of magnitude cheaper and safer than a full firmware upgrade.

  2. Hardware reuse. The same box can run a protocol gateway this year and lightweight vision preprocessing next year—because the application layer is containerised, not because you bought new hardware.

  3. Vendor replaceability. Containers plus a standard image format mean the application can change hands without ripping out the hardware.

Be clear-eyed though: containers are not free. They reliably consume 128–512MB of RAM for images, runtime and page cache, and they need a writable image partition plus rollback space in flash. Container support must therefore be evaluated together with the memory tier. A 256MB device that claims container support will run exactly one minimal image—the single most common misconfiguration in this buying cycle.

Test it instead of trusting it: ask for measured data on target container count, per-container resident memory, and image size. "Supports Docker" is not an answer.

3. Where protocol parsing happens: edge or platform

This is the spec most often settled by default rather than by decision. Plenty of projects parse on the platform out of habit; others push everything to the edge and call it edge computing. Neither is the answer.

Edge parsing—the device converts Modbus, PLC or meter protocols into structured payloads itself—fits high point counts, fast sampling, cellular data plans billed by volume, and sites that must keep working when the uplink drops. It can cut uplink volume by an order of magnitude and removes parsing load from the cloud.

Platform parsing fits projects where protocol variants change often, business users need to adjust rules without a release, and the installed base is a mixed bag of models. Change the rule on the platform, never touch firmware.

For most deployments the pragmatic middle wins: the edge normalises (protocol to a uniform MQTT/JSON structure) and the platform does semantic mapping (which field means which business metric). Edge firmware stays stable and rarely needs upgrading; the platform keeps its flexibility.

4. Supply stability and part-number lifecycle: the cost everyone forgets to model

The biggest hidden cost this cycle is not unit price. It is part-number churn. Memory fabs shifted capacity to higher-margin products, DDR3L and several NAND families were cut back fast, module makers swapped dies, and device makers were forced to re-spin BOMs. For buyers that means:

  • A certified device may need fresh EMC, thermal and long-run stability testing because one memory die changed;

  • A model frozen in a tender arrives as a "pin-to-pin compatible substitute" with an identical-looking datasheet and different headroom;

  • Spare-part commitments evaporate because the original part is gone.

So split this parameter into three questions. Is there a pin-to-pin second source for the critical memory parts? Do part changes go through a PCN process, and what is the notification window? At EOL, can the supplier offer an LTB window and a five-year spare commitment?

Suppliers who answer those cleanly tend to be the ones with their own modules and their own production line—because a part swap forces hardware and firmware to move together. Starlinkun runs that chain end to end, from in-house router core modules (ComIoT series) to ComIn industrial routers, V-series vehicle gateways and the ComCloud device management platform. When a part number changes, firmware adaptation, re-certification and volume delivery get planned as one package instead of being handed to the customer. In a price-surge year, that is worth more than an extra gigabyte of RAM.

Which features are worth paying for in RAM

WorkloadReal memory / storage drawWorth paying more?Recommended tier
4G/5G uplink, VPN backhaul, serial pass-throughLow (under 300MB)No256–512MB DDR
Periodic MQTT/HTTP reporting (under 500 tags)LowNo512MB
Edge protocol parsing, cleansing, store-and-forward (3–7 days)MediumYes—but buy flash first512MB + 8–32GB flash
Edge containers (1–3 apps, canary upgrades)High (+128–512MB)Yes1GB and up, writable partition
On-device vision inference (defect, ANPR)Very highYes, but requires an NPU2GB and up + accelerator
Local HMI or dashboard hostingMediumDepends on concurrent users1GB
Local historian (30+ days)Storage-boundAdd storage, not RAMeMMC or SSD
Dual-SIM redundancy, link backupMinimalNot at all256MB and up

A procurement sequence that holds up

  1. Write down the device's second job—and how many of them there are.

  2. Size RAM to that job, and size the writable flash partition to whether containers are in scope.

  3. Decide which side parses protocols, then validate the decision against your data bill and your offline requirements.

  4. Close with supply terms: second source, PCN notice period, LTB window, spare-part years.

  5. Look at clock speed last. In industrial routing, forwarding throughput is driven mainly by hardware acceleration and packet buffering; a higher clock usually means more power and more heat, which eats into your wide-temperature margin.

Memory prices will eventually cool. The habit of sizing hardware to the workload will not go out of date. Starlinkun's industrial routers and edge gateways cover everything from 4G serial data logging and 5G Wi-Fi 6 backhaul to vehicle gateways; selection guides and solution details are on the official site: www.movingcomm.com.