- Edge computing processes data close to the source (at the machine or plant level) rather than sending it to a central cloud first — which matters when latency, bandwidth, or connectivity reliability are constraints.
- Most manufacturing AI applications don't require millisecond latency — predictive maintenance and OEE analytics run comfortably on cloud or on-premise servers. Edge is necessary for real-time control decisions and vision inspection at line speed.
- The right architecture is often hybrid: edge for latency-critical applications, cloud/on-premise for analytics and long-term model training.
Edge vs cloud is an architecture decision, not a preference
“Edge computing” gets positioned as the advanced, next-generation approach — which implies cloud is somehow inferior. That’s a vendor framing, not an engineering one. The right question is: does your application have latency, bandwidth, or connectivity requirements that cloud processing can’t meet?
For most manufacturing AI applications — predictive maintenance, OEE analytics, quality root-cause analysis — the answer is no. These applications run on data that’s seconds or minutes old, not milliseconds. Cloud or on-premise server processing is entirely adequate.
Edge becomes necessary when: the AI must respond in milliseconds (vision inspection at line speed, real-time control decisions), internet connectivity is unreliable at the plant location, or data privacy requirements prohibit sending raw production data off-site.
A predictive maintenance model that predicts a bearing failure 18 days out doesn’t need to run at millisecond latency — running it every 5 minutes on a cloud server is functionally identical. The latency sensitivity is at the control layer, not the prediction layer. Confusing these two leads to expensive edge infrastructure deployed for applications that don’t need it.
When edge computing is genuinely necessary in manufacturing
Vision inspection at line speed: A packaging line running at 200 units/minute needs a reject decision in under 300 milliseconds. That round-trip to a cloud server and back introduces too much latency. Edge inference on an industrial PC at the inspection station is the right answer.
Safety-critical real-time control: Any AI that feeds directly into a control system decision (not just an advisory) needs to run at the control system’s latency requirements — often sub-100ms. That’s always edge.
Air-gapped environments: Some manufacturing environments (defence, nuclear, certain pharmaceutical) prohibit any internet connectivity. All AI must run on-site.
The practical hybrid architecture
FactoLake supports both edge and cloud deployment patterns — edge nodes at the plant level for data collection and latency-critical inference, with synchronisation to a central cloud or on-premise server for analytics, model training, and cross-plant intelligence.
Frequently Asked Questions
Not typically. MaintIQ, ProdIQ, and QualIQ run on data that’s seconds to minutes old — cloud or on-premise server deployment is adequate. FactoLake edge nodes handle the data collection layer at the plant.
Industrial PCs (IPCs) from vendors like Beckhoff, Siemens IPC, or Advantech are the standard. For vision AI inference, NVIDIA Jetson or GPU-enabled IPCs. Ruggedised to handle plant floor temperatures, vibration, and EMI.
FactoLake edge nodes collect and buffer data locally, then synchronise to a central instance via MQTT or HTTPS when connectivity is available. In poor connectivity environments, the local buffer ensures no data is lost during disconnection.
Related: Complete Guide to AI in Legacy Factories · Using OPC-UA for AI Integration · How to Connect PLCs to AI