Part I asked what may leave the fence. Part II asks what cannot wait for the trip. Line-speed decisions are made in milliseconds next to the machine, inside a power and thermal budget, and a single network hop can consume the whole window. Where inference runs is part of the control design, not an IT preference.
Three tiers, set by the decision window and what the data reveals: device/edge (vision reject, safety, robot guidance; under ~10–20 ms; data never leaves), plant server room (twin sync, OEE, predictive inference on sensitive OT; tens of ms; data stays inside), and cloud (multi-site training, historical analytics, non-critical copilots; seconds, after classification only). The field has moved: edge systems co-located with equipment are reported at 15–45 ms against 800–2,400 ms for cloud, with line-speed vision targets under 10–20 ms, and low-power accelerators make always-on inference viable at single-digit to low-tens of watts. Edge isn’t only speed: it keeps running when the WAN doesn’t.
Four categories stay on-premise: safety interlocks and zone monitoring, line-speed reject, closed-loop write-backs, and any prompt or image revealing process IP. The test for each use case: does delay or data exit have a safety, quality or IP consequence? Your action this week: ask what the maximum allowable latency is for your next vision project, and whether anyone wrote it down. Full tier model at renegrywnow.com.
Reflection questions
What’s the maximum allowable latency for your next vision or robot-guidance project, and has anyone written it down?
Would your line degrade safely if the cloud link were cut mid-shift?
Are safety interlocks, line-speed reject and write-back loops confirmed as on-premise, with no cloud path?
Series: Energy Dominance · Week 45 · Part II
Previous: Part I — Data Sovereignty in Industrial AI: Cloud vs. Plant Control.
