When I first moved our MES to the cloud, I made every mistake in the book. The biggest one was assuming the network was ready. It wasn't. We'd tested bandwidth from the office — 200 Mbps, plenty. But the factory floor was a different story. The VPN tunnel kept dropping during shift changes when everyone clocked in and the guest WiFi spiked. Took us three outages before we traced it.
That taught me something obvious in hindsight: test from where the users actually are, not where the IT team sits. Now I make junior engineers spend a full shift on the factory floor before they touch the architecture diagrams. You see things from a desk that simply aren't true on the ground.
Edge computing in industrial IoT is basically that lesson applied to data. The cloud is great for long-term storage, trend analysis, and dashboards that managers check on Monday mornings. But when a vibration sensor on a CNC spindle detects a pattern that means the bearing will fail in about forty minutes, you don't have time to send that data to Virginia, run an inference model, and wait for a response. The spindle doesn't care about your network topology.
I worked with a food packaging plant last year where they were running vision inspection on a bottling line. Twelve hundred bottles per minute. The cloud-based system had a latency of about 380 milliseconds round trip. That sounds fast until you realize a bad bottle moves past the reject gate in 250 milliseconds. They were catching maybe sixty percent of defects. We moved the inference to a local edge box running a trimmed-down model and got that up to ninety-seven percent. Same model architecture, same training data. The only difference was where the math happened.
The mistake people make is thinking edge means dumb. It doesn't. Modern edge devices are running containerized workloads, doing real inference, making autonomous decisions. The cloud still matters — it's where you retrain models, push updates, aggregate data across sites. But the split matters. I usually tell clients: if the decision affects safety or product quality in under a second, it lives at the edge. If it affects next quarter's planning, it lives in the cloud. Everything else is a judgment call.
Security is where this gets uncomfortable. Every edge device is another attack surface. I've seen plants where the edge boxes were on the same flat network as the PLCs because someone didn't want to deal with VLAN segmentation. That's how you get a ransomware story. The edge layer needs its own security posture — device attestation, encrypted storage, signed firmware updates, network microsegmentation. Treat every edge node like it's in a DMZ, because functionally, it is.
The other thing nobody talks about is maintenance. Cloud infrastructure is someone else's problem. Edge infrastructure is yours. Those boxes sit in cabinets next to motors that vibrate, in rooms that hit forty degrees Celsius, with dust that gets into everything. I've had edge devices fail not because the hardware was bad but because the SD card couldn't handle the write cycles. Now we spec industrial-grade storage and build health monitoring into the deployment pipeline. If an edge node goes dark, the operations team needs to know before the shift supervisor calls.
What I find interesting is how this changes the skill set on the plant floor. You suddenly need people who understand both OT and IT. The old divide — where controls engineers owned the machines and IT owned the network — doesn't work anymore. The best edge deployments I've seen had a hybrid team from day one. Not a handoff. A team.
The factories that get this right won't just be more efficient. They'll be the ones that can actually adapt when the next disruption hits, because their systems can make decisions locally without waiting for permission from a data center that might be on the other side of a fiber cut.