Where the inference runs
Two engagements this week converged on the same constraint: the places that most need continuous machine perception are the places with the least reliable connection to a data centre.
Two separate engagements this week converged on one constraint. The environments with the strongest case for continuous machine perception — underground works, heavy industrial floors, infrastructure in service — are also the environments with the least reliable path to a data centre. The perception has to happen where the thing being perceived is.
What the models are asked to do. Detect deterioration in a structure from imagery captured on a repeat pass. Maintain an inventory of assets that were never registered anywhere. Notice that a condition has changed against a learned baseline rather than against a threshold someone set once. These are ordinary computer-vision and sensor-fusion tasks, and their difficulty is not the model. It is the round trip.
Distributed inference is the structural answer. Splitting a model across nodes near the sensors, with the network carrying results rather than raw streams, changes three properties at once: the latency of a decision, the bandwidth a site has to sustain, and the amount of raw material that ever leaves the premises. The third of those is a governance property as much as an engineering one.
The lined and the unlined case. In tunnelling the distinction between lined and unlined sections is a good illustration. A lined section presents a surface whose expected condition is known and whose deviations are measurable. An unlined section presents rock, where the same camera pass is a harder inference problem and the value of a continuous record is higher, because there is no drawing to compare against.
Our read. Network-enabled intelligence close to the physical asset is where this settles, ahead of hauling everything to central compute. The argument is availability first — a site that stops perceiving when the link drops has bought a capability it cannot rely on — and attack surface second. This is our position rather than a published finding, and we state it as such.
This note is drawn from several conversations during the week of 24 August 2026. No organisation, site or individual is identified.
Where this goes next
Want this applied to your organisation?
One call is enough to know if we're a fit.