Two hundred and thirty-eight. That is how many industrial sites worldwide the World Economic Forum had recognised in its Global Lighthouse Network as of June 2026, out of a global manufacturing base counted in the millions. Whatever “smart factory” means, it is worth starting from the fact that the strictest available list of them is astonishingly short.
Key takeaways
- 238 sites were in the WEF Global Lighthouse Network as of June 2026, across 30-plus countries.
- McKinsey found fewer than 30 percent of industrial digital pilots scale past the pilot stage.
- ISA-95, published as IEC 62264, is the reference model for shop floor to enterprise integration.
- IEC 62443 is the standard series that governs whether that connectivity is defensible.
The word has a traceable origin, and it was a policy programme
The vocabulary everyone now uses came out of a specific room. The term Industrie 4.0 was coined by Henning Kagermann, Wolfgang Wahlster and Wolf-Dieter Lukas, and introduced publicly at the Hannover Messe in April 2011, as part of the German federal High-Tech Strategy. The Industrie 4.0 working group chaired by Kagermann published its implementation recommendations in 2013.
That origin explains a lot about how the term behaves. It began as an industrial policy label describing a direction of travel, not as a technical specification with pass and fail criteria. Fifteen years later, vendors apply it to individual products and factories apply it to themselves, and neither usage carries any obligation to demonstrate anything. This is why the question in the title is worth asking seriously rather than rhetorically.
What separates an instrumented plant from a connected one
Most factories that describe themselves as smart are, precisely, instrumented. They have sensors on machines, screens on walls and historical data in a server. The gap between that and a genuinely connected operation is architectural, and it has a reference model.
ISA-95, published internationally as IEC 62264, defines the layers between the physical process and the enterprise systems, and the information that has to pass between them. It is unglamorous and it is the actual dividing line: a plant where production scheduling in the business system reflects real machine state, and where a quality event on the line changes what the business system does next, has crossed it. A plant where someone exports a spreadsheet every morning has not, no matter how many sensors fed that spreadsheet.
| Commonly cited as proof of “smart” | What it actually establishes |
|---|---|
| Robots on the line | Automation, which predates the term by decades |
| Sensors on every machine | Instrumentation, a prerequisite rather than a result |
| Dashboards in the control room | Visibility, which still requires a person to act |
| A digital twin of the line | A model, useful only if it is fed live and trusted |
| Scheduling that reacts to machine state | Integration, which is the thing being claimed |
Why so few plants get there, in the vendors’ own numbers
The most honest figure in this field comes from the consultancy side of it. McKinsey’s work on what it named pilot purgatory found that fewer than 30 percent of digital manufacturing pilots go on to scale beyond the pilot stage, which means industrial companies fail to capture value from roughly seven out of ten things they try. In a later survey the share of manufacturers reporting themselves stuck in pilot mode rose rather than fell.
That is not a technology failure and the report does not present it as one. Pilots stall because the successful proof of concept ran on one line, with a dedicated team, on data that someone cleaned by hand. Scaling it means confronting the plants that were left out, the machines with no network port, the legacy control systems nobody wants to touch during a production run, and the fact that the maintenance team was never consulted. None of that is solved by the next platform purchase.
Five questions that separate a claim from a capability
When we look at a site that describes itself as next generation, these are the questions that produce the most information fastest. They are deliberately answerable by the plant manager rather than by the vendor.
- Which decision is now made without a human? Name one. If every answer is “we have better visibility”, the plant is instrumented, not automated at the decision layer.
- What happens to the data at 3 a.m.? Systems that only function when the analytics team is awake are reporting tools with a real-time interface on top.
- How many machines are outside the system, and why? The exceptions list is where the truth lives. Retrofitting older equipment is the expensive, unphotogenic work that decides whether the architecture is real.
- Who owns security, and against which standard? The answer should reference the IEC 62443 series, the international standard set for industrial automation and control system security. Connectivity added without it is a liability wearing the language of progress.
- What did the operators get? Sites that scale tend to be the ones where the shop floor gained something usable. Sites that stall tend to be the ones where the shop floor gained a screen that reports on them.
Where artificial intelligence sits in this, honestly
Generative AI has moved quickly into this space, and the Forum’s own reporting on the Lighthouse cohort put generative AI use cases at 23 percent of top solutions in 2025. That is a meaningful share, and it is also a share of solutions deployed at the most advanced sites in the world, which is not the same as a share of factories.
The applications that hold up are narrow and unromantic. Language models used to make maintenance histories and machine manuals searchable in a way a technician can actually use at 2 a.m. Vision systems trained on defects that a specific line produces. Scheduling that reacts to a real constraint. What consistently disappoints is anything that requires the underlying data to be better organised than it is, which returns the problem to integration, where it started. Our look at whether AI can replace factory floor workers takes that question from the labour side.
So what does “truly smart” mean
Our working definition is deliberately narrow. A smart factory is one where the state of physical production is available to the business in a form it acts on automatically, where that link is defensible under a recognised security standard, and where the arrangement survives the departure of the person who built it. Everything else on the usual list is either a prerequisite or a photograph.
By that measure the number 238 stops looking odd. It is not that smart manufacturing is rare because the technology is scarce. The technology is ordinary and mostly for sale. What is rare is an organisation willing to rebuild the connections between its systems and its people at the same time, which is slow, expensive and impossible to demonstrate on a stand at a trade fair.
Interested in the one application that consistently pays back?
Predictive maintenance is where instrumentation turns into a decision most reliably, and where the claims are easiest to check.
Sources: World Economic Forum, Global Lighthouse Network, June 2026 announcement of 16 new awards bringing the network to 238 sites across more than 30 countries and 40 industries, an initiative co-founded with McKinsey & Company, and Forum reporting putting generative AI use cases at 23 percent of top solutions in 2025; McKinsey & Company, “How digital manufacturing can escape pilot purgatory”, for the finding that fewer than 30 percent of industrial digital pilots scale beyond the pilot stage; acatech and DFKI, for the coining of the term Industrie 4.0 by Henning Kagermann, Wolfgang Wahlster and Wolf-Dieter Lukas and its public introduction at the Hannover Messe in April 2011 under the German High-Tech Strategy, with the working group’s implementation recommendations published in 2013; ISA-95, published internationally as IEC 62264, for the enterprise-control system integration reference model, and the IEC 62443 series for industrial automation and control system security. Lighthouse recognition is an assessment against the Forum’s own criteria and is not an exhaustive census of advanced manufacturing sites. Updated August 2026.
