Last updated: 27 September 2026
Short answer: edge computing means running the computing close to where data is produced, on a machine, a factory-floor PC or a server in a shop, instead of sending everything to a cloud data centre. It pays off when a decision has to be made faster than a round trip to the cloud allows, when the raw data is too large to send, when the data should not leave the site, or when the work has to carry on without an internet connection. The training, fleet management and long-term storage usually still happen in the cloud. Three real deployments below show where the line falls.
What does "edge computing" actually mean?
The Linux Foundation's Open Glossary of Edge Computing defines it as "the delivery of computing capabilities to the logical extremes of a network in order to improve the performance, operating cost and reliability of applications and services." NIST's Fog Computing Conceptual Model (SP 500-325) describes a similar middle layer, which it calls fog, sitting between end devices and the cloud for latency-sensitive applications. In practice the words overlap, and it helps to think in layers:
| Layer | Where it runs | Typical example |
|---|---|---|
| On-device | Inside the sensor, camera or phone | A smart camera that flags a defect itself |
| Device or on-premises edge | An industrial PC or small server on site | A PC beside a production line running an inspection model |
| Infrastructure edge | A nearby data centre, such as a telecom operator's | A city-level cloud zone offering single-digit millisecond latency |
| Cloud region | A large data centre, often hundreds of kilometres away | Model training, dashboards, long-term storage |
Most real systems use more than one layer. If you have read my comparison of local and cloud AI for small businesses, this is the same trade-off applied to machines instead of office work.
Why not just send everything to the cloud?
There are four practical reasons, and a project usually needs at least one of them to justify edge hardware.
1. Latency. Microsoft publishes median round-trip times between its Azure regions: about 8 ms between two neighbouring US East regions, 69 ms from East US to West US, and around 200 ms from Central India to East US. Those are backbone figures; a factory's own internet link adds more. A conveyor moving at one metre per second travels 20 cm in 200 ms, which is too late to reject a part at a fixed station. A PC beside the line answers in a few milliseconds.
2. Bandwidth. Industrial cameras produce a lot of data. A single GigE Vision camera can stream around 125 MB per second, and the arithmetic for vibration monitoring is similar in spirit: one accelerometer sampled 25,600 times a second at 16 bits produces about 51 kB/s, or roughly 4.4 GB a day, per channel. Sending a summary of that data, such as a frequency spectrum or a trend value, is cheaper and more reliable than streaming the raw signal.
3. Keeping data on site. Images of a customer's product, a factory's process or people in a shop may be commercially or legally sensitive. Processing them locally and sending only results reduces what leaves the building, which also simplifies the privacy questions I covered in what engineers should know about data privacy laws.
4. Working offline. A production line or a restaurant kitchen cannot stop because the internet connection drops. AWS describes its IoT Greengrass edge software as letting devices "act locally on the data that they generate" and "react autonomously to local events", which is exactly this requirement.
Example 1: predictive maintenance in a factory
The case: Siemens runs its own condition-monitoring software, Senseye, together with its Industrial Edge platform at its Cham electronics plant. The pilot monitored more than 90 motor-driven assets.
What happens at the edge: edge devices on the shop floor collect torque, current, voltage, temperature, energy use, cycle counts and running time from machines over industrial networks such as PROFINET and Modbus TCP. They clean and condense that data locally, and the analysis looks for changes in behaviour that come before a failure, so maintenance can be planned instead of reacting to a breakdown.
The result claimed: Siemens reports a "significant reduction of unplanned downtime" and a rollout to other plants. It does not publish a percentage, and like all vendor case studies it should be read as a claim, not an independent measurement.
Why edge fits: the machines speak industrial protocols that do not belong on the open internet, the raw signals are high-volume, and the useful output is a small amount of information: this motor is drifting and needs attention within two weeks.
Example 2: AI visual inspection on a welding line
The case: Audi uses Siemens Industrial Edge to detect weld spatter on car bodies. Each body has about 2,000 welds, the line builds around 1,000 cars a day, and the inspection has to fit inside a cycle time of about 60 seconds.
How the work is split: images from eight cameras are used to train the model in the cloud, where heavy computing is easy to rent. The trained model then runs on an industrial PC at the line for near-real-time inference, and lights mark the spatter locations for workers to fix.
Why edge fits: thousands of high-resolution images per hour would be expensive to upload, and the answer is needed within the cycle, not after it. This train-in-the-cloud, run-at-the-edge pattern is now the standard shape of industrial AI. I described how the same inspection records feed quality documentation in how AI is changing engineering documentation and quality work.
A caution about platforms: this market changes fast. Amazon's Lookout for Vision inspection service stopped taking new customers in October 2024 and was discontinued on 31 October 2025, and Microsoft retired its Azure Percept edge kit in 2023. If you build an inspection system, make sure the model and data can be moved to another platform.
Example 3: a Kubernetes cluster in every restaurant
The case: Chick-fil-A runs a small Kubernetes cluster in each of its roughly 2,800 restaurants, using three small computers per restaurant and the lightweight K3s distribution. The clusters run business-critical applications locally, and tens of thousands of connected kitchen devices send data to them.
How it is managed: the configuration lives in Git, and each restaurant's cluster pulls its settings from there, an approach called GitOps. A small central team manages thousands of sites without logging into each one.
Why edge fits: a restaurant's ordering and kitchen systems have to keep working during an internet outage, and three nodes mean one computer can fail without stopping the store.
The contrast: Amazon's camera-based Just Walk Out checkout is the famous counter-example of a retail system that proved harder to run than expected. Amazon began removing it from its US Fresh grocery stores in 2024 in favour of smart carts, and in early 2026 announced it was closing its Go and Fresh stores, although, according to GeekWire, the technology continues in more than 360 third-party locations such as stadiums and airports. The lesson is not that edge computing failed; it is that the business case has to survive the cost of the hardware and the people who maintain it.
What does edge hardware look like in 2026?
- Industrial PCs from automation vendors, built for dust, vibration and wide temperature ranges.
- AI modules such as NVIDIA's Jetson family. The top model, Jetson AGX Thor, went on sale in August 2025 with a developer kit from $3,499, aimed at robotics and multi-camera vision. The chips inside these devices come from the same few foundries I mapped in the semiconductor supply chain overview.
- Cloud providers' on-site hardware, such as AWS Outposts racks and servers, which bring cloud services into your own building.
- Edge software platforms, such as AWS IoT Greengrass version 2, Azure IoT Edge and the newer Azure IoT Operations, or plain Kubernetes. Check support dates: Greengrass version 1 reaches end of support on 7 October 2026, and Azure IoT Edge 1.5 on 11 November 2026, according to Microsoft's lifecycle page.
What are the downsides?
- Fleet management. One cloud region is easier to update than 500 small computers in 500 locations. Patching, monitoring and replacing failed hardware is the real long-term cost.
- Security. Every edge device is a computer on a site network that someone must keep updated and physically secure.
- Platform churn. As the retirements above show, a managed edge service can disappear within a few years.
- Duplicated effort. You still need the cloud for training, dashboards and backups, so edge adds a layer rather than replacing one.
When should you use edge computing?
Use it when at least one of these is true: the decision is needed in milliseconds, the raw data is too big to send economically, the data should not leave the site, or the process must keep running offline. If none applies, a cloud service is usually simpler and cheaper, and you can move specific workloads to the edge later once you know where the bottleneck is.
FAQ
Is edge computing the same as IoT?
No. IoT refers to the connected devices themselves. Edge computing is where their data gets processed. Many IoT systems send everything straight to the cloud and have no edge layer at all.
What is the difference between edge and fog computing?
Fog is an older term, used by NIST among others, for a layer of computing between devices and the cloud. The Linux Foundation glossary calls fog computing "an early edge computing concept". Today most people simply say edge.
Does edge computing replace the cloud?
Rarely. In all three examples above, the cloud still handles model training, central configuration or reporting. Edge takes over the time-critical and data-heavy parts.
Is edge computing only for large companies?
No. A single small PC running a vision model beside one machine is edge computing. The hard part at larger scale is managing many sites, not the concept.
