Tech Walks Before Dawn

IoT Hasn’t Tried Its Hardest Yet

The numbers keep growing, yet nothing changes on the shop floor. We go and look at what is missing

Walk 01 First draft September 2026

It is more than fifteen years since the phrase “Internet of Things” (IoT) caught on. The picture of “everything connected to the net” has been painted again and again, and as far as the market statistics go, it really does keep growing. But what you hear from people close to the shop floor sounds quite different. They get as far as fitting sensors and collecting data, and then nothing moves. Today we go and look at what is happening in that “and then”.

1 Starting point ― on paper, it is doing fine

First, the numbers. One research firm estimates the global IoT market at about US$700 billion in 2024, growing past US$4 trillion by 2032 — roughly 24 percent a year. Estimates of the number of IoT devices in operation vary by firm, from a cautious 19 billion or so in 2025 to nearly 27 billion. Either way, several studies point to around 40 billion by 2030 and more than 50 billion by 2035.

Follow the numbers alone and IoT looks like a technology steadily spreading through society. But that is not what people on the shop floor say.

2 First corner ― did the forecasts of ten years ago come true?

Before doubting the growth, it is worth looking back. In the early 2010s, when the word IoT was first catching on, the industry was waving far bolder numbers than today. The famous one is Cisco’s forecast, around 2011, of 50 billion IoT devices in operation by 2020.

What happened? In 2016 the estimates were all over the place: Gartner said 6.4 billion (excluding smartphones and the like), IDC 9 billion, IHS 17.6 billion counting every kind of device. The forecasters themselves had to back down. A former Cisco staffer cut his own figure to 3 billion, and Ericsson arrived at 2.8 billion for 2021 — an order of magnitude below the original 50 billion. The reasons given: the definition of an “IoT device” differed from company to company; the share of things like cars that would actually be online was estimated too generously; guessing demand for products that do not yet exist is inherently hard; and the industry has always had reasons to publish cheerful numbers.

Meanwhile, some things named at the time have still not been solved. The IoT Security Guidelines published in 2016 by Japan’s Ministry of Internal Affairs and Communications and Ministry of Economy, Trade and Industry singled out security and interoperability (devices from different makers working together) as the most important issues. Ten years on, as the state of the smart-home standards we look at later shows, interoperability has not been fundamentally solved.

So the decade sums up like this: the flashy forecasts missed by a wide margin, while the unglamorous problems barely moved at all. Device counts and market size did grow steadily as parts got cheaper, if not as far as first predicted. But the unglamorous, essential problems — connecting things across makers and standards, and linking collected data to actual decisions — have been named for ten years and postponed for ten years. The “ends at the trial” problem we look at next is the continuation of exactly that.

3 The place we most want to see ― why projects end at the trial and never reach production

This is the place we most wanted to see on today’s walk.

Surveys find that 60 to 80 percent of corporate IoT projects ultimately do not succeed. The most telling figure is that more than 70 percent never make it from trial to production. The industry calls this “pilot purgatory”. Most projects get as far as installing sensors and collecting data, then stop at the point of running it as everyday business.

Further, it has been pointed out that less than half of the data collected is actually used in any decision. In other words, the flow that is IoT’s whole point — collect data, decide, change what you do — stops halfway on many shop floors.

Many articles explain this as a communications or technology problem. But read the industry analyses and the reality seems to lie somewhere else. The main cause named for pilot purgatory is “no backing from the top”. Not whether the technology works, but who takes on the authority and responsibility to actually change how work is done on the basis of the data. Add to that the finding that nearly half of companies have no security, data-analysis or connectivity engineers of their own. Collecting data and having people in the organisation who can translate it into decisions are two entirely different things.

One more thing. Looking at the few projects that do deliver, the single factor successful teams share most is “deciding at the outset what would count as making money”. They do not build the technology first and then ask whether it pays; they draw the plan that pays first, and only then touch the technology.

To sum up, pilot purgatory is not one wall but three stacked layers. Most commentary says IoT stumbles at the bottom, technical layer; in practice most projects stop at the two layers above it. There are estimates that adopting standard protocols could cut integration cost by up to 30 percent, so “connecting” is a real burden. But it is one of three.

4 A detour ― why connectivity is still a craft

From here we walk a little through the bottom layer, technology. We come back to the two upper layers in section 5.

IoT’s “three treasures” are often said to be cheap microcontrollers, cheap sensors and cloud data platforms. All three have become astonishingly cheap and ordinary in the past decade or so. Microcontrollers cost tens of yen, sensors have got cheap through volume, and the analysis side in the cloud has matured.

The one part that has not become ordinary in the same way is the communication that connects them. There are three reasons.

First, there are too many standards to choose from. It is said that a company adopting IoT faces more than 30 connectivity options: Wi-Fi, Bluetooth, Zigbee, LoRaWAN, NB-IoT, LTE-M, Sigfox … each with its own strengths and weaknesses in range, battery life, speed and price, and “the designer decides according to site conditions” is still what is demanded. We call this “craft” in this article, and not to make it sound grand: it means the designer has to pile up individual judgements, site by site, about frequency, distance, batteries and regulation.

Second, long range and low cost are hard to have at once. Low-power, long-range radio (called LPWA) was expected to give long battery life, long reach and low price all together, but some analyses find its usable coverage reaches only part of the world’s population. Cellular IoT, on the phone networks, reaches everywhere but the modules cost several dollars or more each and draw more power. “Reaches far but expensive and hungry” or “cheap and frugal but short-ranged”: that choice has not fundamentally changed.

Third, in city-scale deployments there is the problem of carrier fragmentation. Smart meters and traffic systems run for 10 to 15 years, so depending on a single carrier is a long-term risk. Contract with several and you inherit separate procedures, separate bills and separate help desks, and the operator’s burden jumps.

4.1 Does a unified standard solve it? ― the state of the smart home

One answer the industry gave to fragmented communication is Matter, the unified smart-home standard, and Thread, which runs beneath it. Several years after the 2022 release, the picture in 2026 shows what has moved and what has not.

For basic functions — lights on and off, dimming, basic smart locks — devices from different makers are now judged to connect reliably. That is real progress.

But deep problems remain. Version support varies by maker: one is on the latest version while another is still on an old one, or supports only part of it. Many makers ship “minimum” support, and extras such as colour temperature get left behind. One industry commentary sums up Matter in 2026 as “not that it doesn’t work, but that it is only incompletely implemented”.

Unifying the standard was the right idea. But between a standard existing and it being consistently implemented on the ground there is still a wide gap. It is a technology problem and at the same time a problem of each company’s willingness and testing, and no single fix closes it.

4.2 The other constraint ― power

Easily overlooked beside communication is power. For watching a city’s infrastructure, the “meaningful place” to put a sensor does not necessarily have mains or a network. So you end up providing solar-and-battery power one site at a time, which pushes up installation cost and maintenance. And because a lamp post belongs to the power company and a traffic light to the municipality, permits add more delay.

So what the IoT shop floor faces is not a single bottleneck but a tangle: too many standards, the trade-off between range and price, incompletely implemented standards, and constraints on power and sites.

5 The high point ― what has quietly changed, and an assumption worth doubting

This is the spot with the best view on today’s walk. We go back to the two upper layers and look at two things that are quietly changing, and one assumption worth doubting.

5.1 What has changed ― AI has begun to help with “translation”

The first is in the people and organisation layer: generative and agentic AI are entering the workplace. Of 29 industrial AI-agent products identified at Hannover Messe 2026 (a German industrial trade fair), a little over 70 percent had already left the trial stage and were in commercial use, according to one count. The main uses are machine maintenance, fault diagnosis and automatic detection of safety incidents — exactly the territory of “decide from the collected data and act”. The “translation” that only scarce specialists could do is beginning to be taken on by AI.

But a distinction is needed. What AI can take over is the analysis and translation: reading the data and judging what to do next. It cannot take over the authority and responsibility to actually move the organisation on that judgement. The former is a shortage of hands, which AI can fill. The latter is backing from the top, a layer still only humans can occupy. AI lightens the “translation”; the question “who is responsible?” does not go away.

5.2 What has changed ― failure has got cheap

The second is in the money layer: prototyping has become dramatically cheaper. In the early 2000s a prototype circuit board cost tens of thousands of yen for one or two pieces. The US service OSH Park, launched in 2009, brought three boards down to around 1,200 yen, and since the mid-2010s Chinese services such as JLCPCB have pushed that to five boards for 320 yen, delivered in days — a hundredfold drop. Design tools such as KiCad are free and connect design to ordering in one flow.

Here too a distinction is needed. What got cheap is the cost of “trying and checking”, not the cost of “proving that mass production pays”. A 320-yen board says nothing about whether the sums work once production-scale communication fees, maintenance and labour are included. Even so, the change matters. A test that once demanded a single bet of hundreds of thousands of yen can now be run over and over for a few thousand. There is no guarantee of profit anywhere, but when the price of finding out has fallen this far, there is no reason not to find out. Being able to try small and gauge the odds has itself become the surest way to avoid a big loss.

5.3 An assumption worth doubting ― must it be “always connected”?

And this is what the walk most wants to say. Most IoT designs rest on an unspoken assumption: that the device must always be connected to the net. Nobody declared it officially, but it is soaked into the designs. Cellular IoT tariffs are built on constant connection, and monitoring screens treat real-time updates as an obvious requirement.

Yet much of the data from watching a city’s infrastructure does not need that immediacy. For many uses, an update every few hours or every few days is enough. Drop the assumption and, instead of fixed base stations, a realistic design appears in which vehicles already circulating through the city — refuse trucks, buses, delivery vans — collect the data as they pass. This is called a “delay-tolerant network” (DTN). From a design premised on constant connection to one premised on passing contact: by changing the idea, the structural constraint of fixed-infrastructure cost may be sidestepped altogether.

One more: research is advancing on “backscatter” communication, in which a sensor with no battery communicates by reflecting the radio waves around it. Because the sensor emits nothing itself, it is freed from the burden of battery replacement. Both ideas were born by dropping an assumption — “it must always be connected”, “a sensor must transmit its own signal”.

6 End of the walk ― not just talking, but building

To look back: blaming IoT’s slow progress on “immature communications” alone is not accurate. The observation that connectivity is still a craft is not wrong, but trace the real cause of pilot purgatory and, before you reach it, you find the people and organisation layer (who translates data into decisions?) and the money layer (does the money come back?). Working technology is a necessary condition, not a sufficient one.

So what do we do? This problem will not be solved by piling up abstract argument. On the technology layer: which frequency, which chip, how much power, how far, within what regulations — these are things you understand with any conviction only after you have designed a real circuit and antenna and actually sent a signal. And on the money layer, now that prototyping is this cheap, there are more and more things worth testing small, even with no guarantee of profit.

Why does building it with your own hands matter? Because of a set of bodily understandings you never get from reading. The calculated fall-off of a radio signal entering your body as the drop in received strength when you carry the board into the next room. The part that looked perfect on the datasheet turning into a chain of compromises within the real limits of board space and availability. Learning to isolate, with your own hands, whether a board that will not work is failing on the antenna, on power-supply noise, on a setting or on the program. Moving from “knowing” the limit for licence-exempt low-power radio to measuring your own design’s output and judging that it will exceed it. And above all, finding and defining for yourself the problems the textbook does not mention.

None of this is a prescription that guarantees profit. Rather, it is the nearest and surest road to becoming, yourself, one of the people we said in section 3 were missing — the ones who can translate collected data into decisions. A different order of understanding comes only from building. And the cost of trying has never been lower.

Appendix: the limits of this walk

• What we did not look at. Security is named as a ten-year-old problem but not examined. Consumer IoT appears only as far as Matter/Thread
• Conditions on the numbers. Because the definition of an IoT device differs between research firms, device counts and market size are shown as ranges. The failure rates behind “pilot purgatory” also move with each survey’s scope and definitions
• The date. This is the view in September 2026. Matter support and the commercialisation of AI agents can change within months