Edge AI means running inference where the machine is. It is not an architectural preference: latency and bandwidth impose it. This entry explains where that obligation comes from, which computation stays at the edge, what goes up to the platform and what limits the decision has.
What edge inference is
Edge computing places compute inside the plant, next to the equipment. Edge AI is the next step: an already trained model runs on that same device, so data becomes a decision without leaving the perimeter. The distinction that matters is between training and inference. Training needs long, comparable history and substantial compute, and lives where that history is. Inference needs to be close to the phenomenon, which is why it moves down to the plant. The same model is trained above and executed below, versioned, with the ability to withdraw it quickly.
Why latency imposes it
Industrial loops work with common cycle times of 10 ms, 1 ms and 0.5 ms in periodic communication, according to the 5G-ACIA and ZVEI paper on integrating industrial Ethernet with 5G networks published in November 2019. The same document notes that working clock domains, where robots and motion control live, require synchronisation of 1 microsecond or better. With those figures, closing a loop across a non-deterministic network does not fail on bandwidth: it fails on delay variance, because a loop with jitter cannot be tuned. The industrial network profile that addresses this, IEC/IEEE 60802, was published on 29 June 2026 according to the IEEE 802.1 TSN task group, so it is now a standard and not a draft. The same logic appears in the state of the art of robotics: in Helix, presented by Figure AI on 20 February 2025, the reactive policy runs at 200 Hz and semantic reasoning at 7-9 Hz. Two speeds, and the fast one always next to the actuator.
Why bandwidth imposes it
The second reason is volume. Diagnosing a bearing requires observing a vibration band of several kilohertz, and the Nyquist criterion forces sampling above twice the highest frequency of interest, with extra margin if the shape of the signal is also to be reconstructed. Transmitting that waveform continuously, per asset and for years, is neither reasonable nor useful. The sensible approach is to compute the spectrum and the indicators at the edge and send the result with its time window. The same applies to images: an inspection camera produces a stream there is no point in uploading whole so that ninety-nine per cent of the frames turn out to be a correct part.
What stays and what goes up
The rule that orders the split: compute lives at the highest level that meets the latency requirement and still lets the level below degrade safely if the level above disappears.
- Stays at the edge. Acquisition and the filtering that precedes sampling, source timestamping, indicator computation, anomaly detection, in-line visual inspection, the local buffer that survives a link outage and any decision with bounded latency.
- Goes up to the platform. Aggregates with their context, events with their timestamp, episodes clipped around a failure and the quality metrics of the data itself, which are what tell you whether the history is defensible.
- Comes down from the platform. Versioned models, thresholds, recipes and setpoints, always with a tested withdrawal path.
Limits of the decision
The edge carries an operating cost: keeping dozens of devices with versioned models, watching the drift of each one and being able to roll back without travelling to the plant demands a discipline that does not appear on its own. A device that also writes into the process widens the exposed surface, and there the reference framework is the ISA/IEC 62443 series on cybersecurity for industrial automation and control systems. There is also a conceptual limit: moving inference closer to the machine does not solve the problem of knowing what state the plant is in. If data arrives with no unit, no asset and no reliable timestamp, running the model one metre closer does not fix it.
Related terms
Edge AI is how physical AI reaches sensors, cameras and controllers. It rests on the industrial gateway and on data acquisition, it feeds anomaly detection and industrial computer vision, and it is the layer where much of the closed loop runs. On the Physical AI Ladder it already appears at level 1 and becomes decisive at level 3. The full architecture is described in the industrial data platform.