Edge Compute Topology
The choice of where computation sits along the path from device to region, trading latency and autonomy against manageability and consistency.
"Edge" spans several genuinely different tiers, and conflating them produces designs that do not fit any of them.
On-device compute gives the lowest latency and full offline autonomy, constrained by power, memory and the difficulty of updating. Gateway or on-premises compute aggregates a local population of devices, which is where most industrial architectures place real processing — it can run substantial workloads, survives a site's internet outage, and keeps data local for residency reasons. Edge point of presence compute in a provider's distributed network is milliseconds from users, ideal for request manipulation and cached personalisation, and constrained in runtime and state. Regional cloud is where anything requiring strong consistency, large state or heavy compute belongs.
The forces that push work outward are latency floors set by physics, bandwidth cost, and the requirement to keep operating when the link fails. The forces pulling work inward are manageability, consistency, and the fact that debugging a fault distributed across four thousand sites is a different discipline from debugging one in a region.
The rule that holds up: push outward only what genuinely requires it — the control loop that must close in ten milliseconds, the filtering that reduces bandwidth by two orders of magnitude, the function that must work when the connection is down — and keep the rest central.