In May 2026, industrial production in the eurozone was 1.2 per cent lower than a year earlier. With weak demand and pressure on margins, technological investments are judged on the basis of reductions in downtime, quality issues and operating costs, rather than on the number of platforms implemented. In this context, edge computing serves a specific purpose: it allows data to be analysed where it is generated, without making the production process dependent on an internet connection or a remote data centre.
It is not a question of moving the entire cloud into the factory floor. The edge is a computing layer between the machine and the central IT environment. Controllers, cameras, sensors and quality control systems transmit data to a local industrial computer or server. This is where processing, filtering and preliminary analysis take place. The results, aggregated data and information required for training models and comparing plants are sent to the cloud.
This division eliminates three problems: latency, transmission costs and the uncontrolled dissemination of information about the production process.
Milliseconds matter
Data used to stop a machine, adjust a robot’s parameters or reject a defective component should not have to wait for a response from the cloud. Network latency is not constant, and a lost connection must not deprive the plant of its ability to react.
This also applies to maintenance. At Siemens’ Nuremberg stamping plant, nine presses have been connected to a local analytics platform developed by Schuler and Siemens. The system records a slight increase in pressing force that accompanies tool wear. The change is too subtle for a single sensor to be sufficient for its reliable interpretation. Only by analysing a sequence of measurements can an anomaly be detected and maintenance scheduled before a breakdown occurs.
Edge computing is therefore transforming the significance of machine data. Rather than merely serving as material for a subsequent report, it becomes a signal that influences ongoing production.
Not every gigabyte deserves a trip to the cloud
The most obvious example is industrial image analysis. Cameras can record a product from multiple angles around the clock, although most frames show no anomalies. Transmitting the entire stream means paying for the transfer, storage and processing of data whose informational value is minimal.
Emirates Global Aluminium analyses around 2.9 petabytes of imagery annually at its plants, including for the monitoring of overhead cranes and the inspection of carbon anodes. Following the launch of a hybrid on-premises platform, the response time of AI models has been reduced by a factor of ten to thirteen, whilst the reported cost of image- and video-based applications has fallen by 86 per cent. The model used for anode inspection was trained on 100,000 images.
These figures have been published by the technology provider and therefore do not constitute a universal benchmark. However, they clearly illustrate the economics of large data streams: raw images should be retained locally, whilst detected defects, metadata and selected samples should be transmitted centrally.
3M applied a similar approach. In one of its factories, machine data and product data were generated in different systems, and network bandwidth was limited. A local database and predictive model made it possible to continue the analysis even without access to the cloud. The transfer of a dataset, which previously required manual operations and took weeks, was reduced to a matter of minutes. A single engineer completed the integration of the solution into the existing environment in around six hours.
The product formula is more valuable than telemetry
Detailed recipes, processing parameters, machine settings, quality control images and data revealing the productivity of individual production lines should not leave the factory without a clear need to do so. In many industries, it is not patent documentation but the process configuration itself that constitutes the most valuable part of know-how.
This also has a regulatory dimension. The EU Data Act, which comes into force on 12 September 2025, regulates access to data generated by connected devices, including industrial machinery. A company must know what information is generated, where it is stored and who is authorised to access it. A local architecture facilitates this control, though it does not replace it.
‘Edge’ is not synonymous with security either. Every local server and industrial computer requires inventory management, updates, network segmentation and access control. ENISA recognises ransomware as the most severe cyber threat in the EU, and phishing and the exploitation of vulnerabilities as the main points of entry into systems. A distributed infrastructure without centralised management may increase the attack surface rather than reduce it.
A mature industrial architecture keeps control, image analysis and sensitive process data on-site. It uses the cloud for model training, software management and the analysis of results from multiple locations. This is not a compromise between the edge and the cloud, but a technical division of responsibilities: the factory retains its operational capability, whilst the organisation does not sacrifice scale.
