Updated July 2026 · 7 min read · Adente Vision Engineering Team
"Edge or cloud" is one of the first architecture questions on any AI project, and for most software the answer is genuinely open. For line-speed visual inspection it is not. The decision has to be made in milliseconds, on data the factory will not let leave the building, on a line that cannot stop because the internet did. Those three constraints point the same way: the inference runs on the edge.
This post explains why, and, because honesty is part of the argument, where the cloud still earns a place.
Latency: why milliseconds beat the round-trip
A vision unit has one job per part: capture, decide, signal the line, all before the next part arrives. On a line running 100 parts a minute, the whole cycle has to finish in well under a second, and the decision itself in tens of milliseconds.
Edge inference makes that budget. The image never leaves the compute unit; the model runs locally and returns a pass/fail in a fixed, predictable time, on one delivered cap line, roughly 30 milliseconds per part. A cloud round-trip adds network transit, queueing and the variability of a shared connection to that same budget. Even when the average looks acceptable, it is the worst-case latency that matters on a line, and the worst case over a network is neither small nor predictable.
For a machine that has to keep pace with production, deterministic beats fast-on-average every time.
Data sovereignty: images that never leave the line
The second constraint is not technical, it is a rule most plants already have: raw production imagery does not go off site. Those frames can reveal part geometry, process detail, throughput and yield, exactly the information a manufacturer treats as confidential.
Edge inspection sidesteps the problem entirely. The image is analysed on the device and discarded or logged locally; only a pass/fail result, and, if you choose, an aggregate metric, ever needs to move. Nothing sensitive crosses the factory boundary. This is also why edge deployments clear internal review faster: when no image leaves the line, there is far less for IT and security to sign off before the system can be switched on.
A cloud-first inspection system inverts this. To make its decision it has to send the raw frame out, which is the one thing the plant was trying to avoid.
Uptime and IT approval
Edge inference keeps working when the connection does not. A production line cannot pause inspection because a switch failed upstream or the site link dropped; the check has to run on every part regardless. A model on the device does exactly that, it has no external dependency to lose.
There is an operational bonus. An edge unit that speaks to the PLC directly over an industrial protocol behaves like any other device on the line: one cable, one trigger, a pass/fail per part. There is no cloud account to provision, no data-processing agreement to negotiate before the first inspection, and no ongoing question about what happens to the imagery. For a plant, "runs locally, nothing leaves the line" is a much shorter approval path than "streams every frame to a third-party service".
Edge vs cloud, side by side
| Factor | Edge (on-device) | Cloud |
|---|---|---|
| Decision latency | Milliseconds, deterministic | Network-dependent, variable |
| Data privacy | Images stay in the factory | Raw frames leave the site |
| Uptime | Runs even if the link is down | Needs a live connection |
| Line integration | Direct PLC signalling | Extra hop and buffering |
| IT / security approval | Short, nothing leaves | Longer, data leaves the site |
| Best role | The pass/fail decision | Fleet analytics, dashboards |
When the cloud still helps
Edge for the decision does not mean cloud has no place, and a vendor who pretends otherwise is overselling. The cloud is genuinely useful for everything that happens after the pass/fail, where latency and raw imagery are no longer the issue.
- Fleet analytics. Aggregated results from many units across several sites are easier to trend centrally, defect rates by line, drift over time, comparisons between plants.
- Dashboards and remote monitoring. A web view of status and summary results lets a quality manager watch a line without standing next to it. Note what moves here: metrics and summaries, not raw frames.
- Model management. Distributing an updated model to a fleet is a coordination task the cloud handles well.
The pattern that works is a clean split: the decision on the edge, the overview in the cloud. The frame stays local; only results and aggregates travel. That keeps line-speed latency, data sovereignty and uptime intact while still giving you central visibility. (For how the on-device architecture fits together, see the AI Visual Inspection guide and the system architecture.)