I spent the better part of 2018 convincing a food processing plant to send all their sensor data to AWS. It was the right call at the time. Cloud was cheap, scalable, and the dashboards looked great in the boardroom. But by 2021, that same plant was running into a problem nobody had budgeted for: latency on the kill switch.
They had a thermal anomaly detection system that was supposed to shut down a conveyor line when bearing temperatures spiked. The round trip — sensor to cloud, cloud to actuator — was averaging 800 milliseconds. That sounds fast until you realize a jam on that line could destroy $40,000 worth of product in under two seconds. The cloud was too slow for the thing that mattered most.
That is the core argument for edge computing in industrial IoT, and it is not a new one. What is new is the economics. Five years ago, putting a real compute node on a factory floor meant buying a ruggedized server that cost as much as a used car. Now you can get an ARM-based edge gateway with enough horsepower to run a TensorFlow Lite model for under $300. The hardware got cheap enough that the conversation shifted from 'can we afford edge?' to 'can we afford not to?'
The other thing that changed is the data volume. A single CNC machine with vibration sensors can generate 20 gigabytes per day. Multiply that across a plant with 200 machines and you are looking at four terabytes daily. Sending all of that to the cloud is not just slow — it is expensive. Egress fees alone can eat a mid-size manufacturer's IT budget. Edge lets you filter, aggregate, and act locally, then push only the summaries upstream. You keep the signal and throw away the noise.
I made a mistake on a steel mill project last year that taught me something about where edge belongs. We put all the inference at the edge — defect detection, thickness monitoring, the works. It worked beautifully for six months. Then the model started drifting because the raw material supplier changed and the edge node had no way to retrain without a manual update. We ended up with a hybrid: inference at the edge for speed, retraining in the cloud for accuracy, and a pipeline that pushes updated models down on a schedule. That is the pattern I see working best now. Not edge versus cloud. Edge and cloud, each doing what it is good at.
The security story is worth mentioning too. Every sensor that talks to the cloud is a potential attack surface. I have seen plants where the OT network was air-gapped by policy but someone plugged a cellular modem into a PLC because the vendor needed remote access. Edge computing lets you keep sensitive process data local while still getting the benefits of centralized monitoring. You reduce the blast radius without reducing visibility.
If you are starting this journey, do not try to edge-enable everything at once. Pick the use case where latency or bandwidth is actually costing you money. Prove it there. Then expand. The plants I see struggling are the ones that treat edge as an architecture decision before it is a business decision. Start with the pain, not the technology.