Last year, I stood on the plant floor watching bottles race past at 600 a minute. Our new inspection system was supposed to catch label misalignments. Instead, it was kicking out perfectly good product. 120,000 bottles in a shift, and our reject rate had tripled. The operator looked at me like I’d personally broken his line.
We had built a beautiful cloud-based vision AI. Every image from the line camera got uploaded, analyzed, and a reject signal came back. On paper, 200 milliseconds of latency seemed trivial. The thing is, 200 milliseconds at 10 bottles per second means two bottles pass the camera before the decision arrives. The reject piston would fire late, hitting the wrong bottle, and operators started overriding the system because it was more trouble than it was worth.
The fix felt almost too simple. We took the same trained model and deployed it on a Jetson Nano plugged directly into the camera’s Ethernet port. No cloud, no round trip, just local inferencing. Latency dropped to under 10 milliseconds. The reject piston hit the right bottle every time. The line went back to normal within an hour, and I stood there feeling foolish for ever thinking the cloud was the answer to everything.
That incident stamped a rule into my workflow: control loops faster than human reaction time don’t belong anywhere but on the edge. I’d read about fog computing for years, but until you watch thousands of dollars of product hit the reject bin because a packet took a detour through a congested VPN tunnel, it’s all academic.
Then I saw the same play out at a chemical plant. They tried running closed-loop pH dosing via a virtual machine in a data center two states away. A network spike caused a 300ms delay in the PID loop. The acid pump kept pumping, and the batch turned into a useless slurry. The plant manager put it bluntly: he’d rather have a cheap PLC with local logic than a perfect AI model with unpredictable jitter. He was right.
Edge computing in industrial IoT isn’t about being anti-cloud. It’s about recognizing that data has gravity. The decisions that keep machines from tearing themselves apart need to happen where the physics are. Now I design architectures so that the edge handles immediate actions: safety stops, defect rejection, motion control. The cloud handles what it’s good at: training models on historical data, spotting long-term drifts, and comparing performance across sites.
I learned one more lesson the hard way. On a packaging line, I set up an edge gateway to buffer vibration data locally for seven days in case the WAN went down. The WAN went down on a Friday evening. By Sunday midnight, the gateway’s SSD filled up with uncompressed sensor data and the whole thing crashed. The line ran blind for hours before anyone noticed. Now I treat edge storage like a server room: strict retention policies, compression quotas, and disk monitoring that pages someone immediately, not on Monday morning.
When I’m asked about edge computing today, I don’t talk about fancy marketing terms. I talk about bottles and chemical batches and that pit in your stomach when you realize the network isn’t deterministic. Some things can’t wait for a round trip. That’s not a philosophy, it’s just physics.