The Radio That Lasted Six Years

On 20 November last year AT&T confirmed that it had stopped selling NB-IoT data plans and stopped certifying NB-IoT devices, and that the planning to decommission its NB-IoT network in the United States was underway. The project was to be finished as early as the first quarter of this year. The network opened in 2019, six years end to end for a technology whose entire commercial premise was that a battery-powered device could be installed and forgotten about for ten.

The wording was careful and, as a carrier statement, entirely reasonable: the company was improving its services for business customers by moving them from NB-IoT to LTE-M, which offers more data capacity. No customer was abandoned, existing ones were being worked with, and the destination really is the better technology. All of that is true. None of it helps the person with water meters in the basements of four hundred buildings.

Then, on 17 July, the same carrier announced nationwide 5G RedCap coverage, over two hundred million points of presence, described in the reporting as replacing the decommissioned NB-IoT network. Read those two announcements together and the shape is clear enough. Inside eight months a low-power cellular technology was retired and a replacement for the replacement was launched, and neither event was a generational shutdown. LTE is fine. LTE-M is fine. A thing inside them is gone.

Retirement inside a generation is the mechanism the rest of this is about, and it is not really about NB-IoT. The unit of retirement in a mobile network was never the generation. The unit is any sub-technology inside one, and a sub-technology can be switched off while the generation carrying it runs for another decade, which means the sentence everyone repeats to justify a fifteen-year deployment does not say what they think it says.

"LTE will be around into the 2030s" is probably true and not the point. It is a statement about a generation, and generations do last a long time; the traffic, the handsets, and the spectrum economics all say so. What a device holds is not a generation. It holds a specific radio technology, a specific band, a specific set of features negotiated at attach, and any one of those can be withdrawn on its own. When it is, the generation is still there and your device is still off. The reassurance and the risk are about two different objects.

The narrower the technology, the easier it is to retire. A general-purpose LTE band carries phones, and phones are what stop a carrier turning something off. NB-IoT carried none. Its subscribers were meters, sensors, and trackers, none of which vote or call support, and most of which sit inside a small number of enterprise accounts a carrier can telephone individually. The property that made it cheap to run, a tiny, quiet, purpose-built slice, is what made it cheap to remove.

And the replacement is on its own clock. LTE-M is the recommended destination and a fine technology, and it is also a sub-technology inside LTE carrying exactly the same structural exposure as the one being retired. 5G RedCap, the newer answer, arrived in Release 17 of the standard, with an enhanced variant in Release 18 reaching toward the low-power range NB-IoT and LTE-M cover today. Every migration here is a move from one retirable slice to another.

No US operator has published an end date for LTE at all. That is the fact to sit with, and it is the opposite of reassuring. No countdown to argue with, no date to plan a refresh against, no document anybody signed. The absence gets read as a commitment, as though no published date meant it was not going to happen, when the reverse holds. The carrier owns an option it can exercise whenever the spectrum is worth more doing something else, and the fleet owner is short that option without having been told they wrote it.

Telstra and Optus both closed their 3G networks on 28 October last year, Vodafone's Australian network having gone before them at the end of 2023. Telstra had announced a June 2024 closure years earlier and moved it, which is worth noting for what it says about how firm these dates are in either direction. They slip. A fleet owner who planned to the original date spent early, while one who assumed the slip would repeat did not spend at all.

The extension existed because of what was still connected. Handsets, yes, but also the long tail nobody had an inventory of: personal alarms worn by people living alone, lift emergency phones, payment terminals, farm and vehicle trackers, and medical devices reporting home over a modem chosen a decade earlier. The warnings in the run-up were about equipment silently ceasing to function on a particular morning.

And the failure was total rather than degraded, which is the part people underestimate. A device that cannot register on the remaining network is off. Australian law goes further and prohibits providers from carrying a device that cannot reach the national emergency number, so a handset that could not do voice over LTE lost service entirely instead of losing some of it. The device has no partial mode to limp along in while procurement catches up.

The AT&T decision matters more than its size suggests. Those Australian devices were not stranded by a whole generation ending. A specific radio capability went away underneath equipment whose owners had no mechanism for knowing it was there. Shrink that from a generation to a sub-technology and you get the same failure with a fraction of the warning and none of the newspaper coverage.

The part can reach end of life before the network does. In its own release of 7 April this year, u-blox records having divested its cellular business in March. u-blox is a long-standing supplier of cellular modules to industrial customers, and it is leaving the category by corporate decision, entirely independently of what any carrier does. A module can go end of life, a supplier can exit, and a certification can lapse while the network it talks to runs perfectly well. The device is stranded either way.

Certification is a per-carrier, per-device fact with no floor under it. Note what AT&T stopped first, well before the network: the certification of new devices and the sale of plans. That step ends the product long before it ends the deployment. You can no longer buy the thing you standardized on, so the spare in the van is the last one, and its replacement is a different module with a different footprint, a different firmware image, and a certification of its own.

Not one party in the chain signed anything about service life. The carrier's terms let it modify or discontinue a service. The module vendor warrants a part against defects, not against a network existing to talk to. The standards body specifies a technology and has no authority over anyone who deploys it. Meanwhile the device is sold on a ten-year battery and installed by somebody costing it over fifteen. Every party is behaving reasonably inside its own contract, and the service life the buyer was sold is in none of them.

Spectrum is finite and it is the input everything else is built from. A generation held open forever is capacity denied to everyone currently using the network, and refarming is the mechanism by which the network any of us use gets better. A carrier that retired nothing would be running a museum funded by the people trying to make a call. The complaint is the distance between how long a device is sold to last and how long anybody will say out loud that they will carry it.

The counterexample is a published date, with not one extra year of life in it. On 8 December 2021 the UK government and the mobile operators stated that the operators do not intend to offer 2G or 3G networks past 2033 at the latest, and the reasoning attached to it is the whole argument in one sentence: establishing the date was necessary to provide clarity about the likely commercial longevity of these technologies, so that users could plan with confidence.

Individual operators remain free to switch off earlier, which several said at the time they would, so the date binds nobody the way a warranty would. The date is still worth more than the alternative, because it turns an unbounded risk into a number a buyer can put in a spreadsheet. A date you dislike can be planned against. An absence cannot, and it gets priced at zero by everybody in the purchase, which is how it ends up costing the most.

Price the option, since you are the one short it. The device costs what it costs plus the expected cost of replacing its radio before the equipment wears out, and that second term is not zero merely because nobody quoted it. Put a number on the truck roll. An engineer, a site visit, an access arrangement, a commissioning test, times the count of installed units. In most fleets that figure dwarfs the module, which reorders the procurement decision and is why it usually goes uncalculated.

Buy the radio as a replaceable part where you can. A modem on a socketed module or a separable gateway turns a fleet-wide replacement into a component swap. It is the design decision that most reduces exposure. It costs more per unit and it is the first thing value engineering removes, because the saving is visible now while the cost lands on somebody else in year eight. Splitting the saving from the cost is a governance problem. Worth arguing as one.

Ask for the sunset terms in writing before you buy. You will probably not get a date. What you can get, and few buyers ask for, is notice: a contractual minimum period between a technology being withdrawn from sale and withdrawn from service, plus a commitment on which technologies the account may migrate to. Notice is a negotiable term rather than a law of nature, and asking costs one clause in a document somebody is already drafting.

And keep an inventory of what your estate actually speaks. Every organization stranded so far shared one property: nobody could produce a list of which devices used which radio technology on which band. The Australian long tail was long precisely because it was invisible. A field in an asset register naming the radio module and what it supports is close to free to maintain, and it is the difference between a six-month project and a very bad month.

A published date buys less than it appears to. The UK statement is an intention rather than an obligation, dates move in both directions, and a buyer planning precisely to 2033 will be wrong about something. Nor is this an argument against cellular for machines: the alternatives have their own version of the problem, and a private network you run yourself only moves the retirement decision onto your own desk, where at least you can see it.

The risk is held by people who never agreed to hold it. A carrier makes a rational spectrum decision, a module vendor makes a rational portfolio decision, a standards body specifies and moves on, and somewhere a sensor bolted to a pipe in a plant room stops reporting on a Tuesday because a technology it never heard of was withdrawn from a network it cannot see. Fifteen years was never promised. Everybody sold it that way.