Architecture

Three Questions That Decide How Your Devices Talk

The Meeting Where This Goes Wrong

Somebody arrives with a technology already chosen. It came from a conference talk, a competitor's press release, or a supplier who was helpful about something else. The rest of the meeting is then spent working out how to make the site fit the radio, which is exactly backwards, and the cost of getting it backwards is not visible until the second hundred devices go in.

The choice is not really a radio choice. Three questions decide it, none of them about modulation, and once they are answered honestly the shortlist is usually one option long. Everything below is organised around those three, because range figures are the easiest specification to publish and the least useful thing to plan with. The only range number worth anything is the one you measure on the site.

Question One: Where Does the Power Come From

This splits the field in half and it is not negotiable.

A device with mains power can use anything. Ethernet, Power over Ethernet, Wi-Fi, a cellular router, a wired fieldbus: all of them are live options, and the decision collapses into what infrastructure the building already has. If a cable tray runs past the machine, wired wins, and the fact that wired is unfashionable does not make it wrong.

A device on a battery rules most of that out immediately. Associating with a Wi-Fi access point costs on the order of hundreds of millijoules per exchange once you count the scan, the association and the DHCP round trip, which is why a Wi-Fi sensor on two AA cells lasts weeks rather than years unless it reports about once a day. Bluetooth Low Energy is genuinely frugal but reaches tens of metres, so every device needs a mains-powered listener within earshot. What is left is the low-power wide-area family, which exists for precisely this case: a few hundred bytes a day, kilometres of reach, and years on one cell.

Solar looks like it promotes a device from the second group to the first. It does not, quite. Panels get sized against the worst month at your latitude rather than the annual average, and a panel large enough to run a Wi-Fi radio through a Northern European December, or a monsoon season, is a piece of infrastructure with its own mounting and its own failure modes. Treat solar as a generous battery budget, not as mains.

Question Two: Who Owns the Ground

This one is commercial rather than technical, and it decides the five-year bill.

If the devices sit on land or in buildings you control, you can own the network outright. Buy gateways, run a server, and the infrastructure is yours: no monthly fee, no carrier contract, no dependency on somebody else's product roadmap, and no sensor data crossing a third party on its way to you. A campus, a farm, a factory, a water utility's district, a portfolio of buildings: all of these justify infrastructure, and once the gateway is up, the marginal cost of the two hundredth device is the price of the device.

If the devices are scattered one to a location across a region, or they move, that inverts completely. No single site justifies a gateway, and a carrier has already built the coverage. You pay per device per month and the coverage problem becomes somebody else's, which is a real service you are buying rather than a tax. What you give up is control of the price and of the technology's lifetime, and several carriers have already quietly deprioritised NB-IoT in favour of LTE-M.

The third case is a vendor who sells connectivity on infrastructure they operate. It is the easiest option to start with and the hardest to leave, because you can neither extend the coverage yourself nor take the network with you when the commercial terms change.

Coverage you own has a large first bill and a flat one after that. Coverage you rent has a small first bill and a line item that never ends. The crossover is arithmetic, not ideology, and it moves with how many devices sit at each site.

Question Three: What Does Year Five Cost

Run both numbers before committing, because intuition is reliably wrong here.

Infrastructure you own is front-loaded and then flat: roughly 300 to 600 EUR for an outdoor gateway, 20 to 100 EUR per device, a server that runs on a small virtual machine, and after that nothing per device per month. The risks are your own: coverage gaps you have to fix, a gateway that needs a truck roll, and the expertise to keep a server alive.

Cellular inverts the shape. Modules are often cheaper than the equivalent gateway-plus-device stack at low volume, and then each unit carries somewhere between 0.50 and 5 EUR a month. A thousand devices at 2 EUR over five years is 120,000 EUR in connectivity alone, before a single sensor is bought, and that line does not stop when the deployment does.

Twenty devices across twenty sites will nearly always be cheaper on cellular. Two thousand devices on one campus will nearly always be cheaper on infrastructure you own. Between those, do the sum. The number that usually decides it is devices per site, not devices in total, and it is worth checking whether the deployment can be reshaped to raise it.

What Each Option Is Actually For

Wired still wins more often than the industry admits. Inside a plant or a machine room, an RS485 run daisy-chained between meters costs a few euros a metre, never drops a reading, and has no battery to replace. What kills wired is the cost of installing the cable, not the cable: trenching a yard, drilling a listed façade, or working in an occupied tenancy is where the quotation goes wrong. Where the route already exists, use it, and see the retrofitting article for how existing fieldbus equipment gets bridged onto a modern pipeline.

Wi-Fi is the most over-specified choice in IoT, and its problems are organisational rather than technical. The IT department owns the network and will reconfigure it without telling you. The access points do not reach the basement plant room or the far end of the yard, which is exactly where the interesting measurements are. And the power cost rules out batteries for anything that reports more than a few times a day. Where it works well is mains-powered equipment inside a building whose network you actually control.

Bluetooth Low Energy is a layer, not a system. A beacon broadcasting an identifier runs for years on a coin cell and costs a few euros, which is what makes tagging thousands of pallets, tools or bins economical when satellite trackers would only ever be justified for the highest-value items. It needs mains-powered scanners spread through the space, so it usually sits underneath a wider network rather than replacing one.

Cellular splits into two that get conflated. LTE-M carries real throughput, handles handover properly and supports devices that move, which makes it the right answer for anything on a vehicle. NB-IoT is the narrower, cheaper, more power-frugal sibling with good building penetration in the sub-gigahertz bands, but its coverage depends entirely on where your carrier chose to deploy, and rural gaps are common because carriers build where the business case is. Battery figures of five to ten years are quoted for both and are achievable, though the protocol stack is heavier than a low-power radio's and maintaining carrier registration costs energy that a simple wake, transmit, sleep cycle does not spend.

Private low-power radio, LoRaWAN in nearly every case I am asked about, is what the unlicensed sub-gigahertz bands make possible. Three to fifteen kilometres rural, one to five urban, and building penetration that higher frequencies cannot match: one indoor gateway commonly covers several thousand square metres through concrete and multiple floors. Devices reach five to ten years on a cell because they sleep almost continuously and wake to send a payload measured in tens of bytes. The cost of owning it is owning the coverage, and a device that wanders off your gateways simply goes quiet.

The Comparison That Is Worth Tabulating

	Private LPWAN	NB-IoT	LTE-M	Wi-Fi	Wired
Battery life	5-10 years	5-10 years	2-6 years	days to months	not applicable
Typical reach	1-15 km	1-15 km	1-10 km	30-80 m	cable length
Handles movement	no	limited	yes	no	no
You own the network	yes	no	no	yes	yes
Recurring per device	none	monthly	monthly	none	none
First-site cost	gateway	none	none	none	cable install
Fails because	coverage gaps	carrier decisions	carrier decisions	IT changes	install cost

Where This Advice Stops

Two situations break the framework and are worth naming.

High-rate data does not belong on any of these. Vibration or acoustic monitoring sampled in kilohertz, or anything with a camera in it, is a different engineering problem: the sampling happens at the edge, the reduction happens at the edge, and only a summary travels. A project that plans to stream raw waveforms over a low-power link has not been costed properly.

The other is the deployment that needs two answers at once, which is more common than any single-technology supplier will tell you. A private network on the sites dense enough to justify a gateway, cellular on the scattered outliers, and one ingest pipeline behind both is frequently the cheapest total answer. It is slightly more work to build and considerably cheaper to run, and nobody whose business is selling one of the two will propose it.

What I Provide

Most of my involvement in this decision is a coverage survey and a spreadsheet, in that order. The survey is measurement rather than modelling: real devices at the awkward positions, over the working day, with the doors shut and the machinery running, because a propagation model has never once told me what a steel switchboard door does to a signal. The spreadsheet is the five-year comparison with your device counts and your site geography in it rather than a vendor's.

What comes out is a recommendation with the arithmetic attached, and often a hybrid that nobody proposed. From there I can build the whole thing, gateways through to dashboards, or hand you the plan and stay out of it. Either way the reasoning is written down, because the person who inherits this in three years deserves to know why each decision went the way it did.