Most AI procurement decisions in industrial operations get framed as a question about capability: which platform has the better model, the better dashboards, the better API. The more important question is where the model runs.
That question has a specific answer for specific use cases.
When cloud AI fits
Cloud AI suits use cases where the latency of a round-trip (80–200ms on a reliable connection, 500–800ms on satellite) is not a constraint, where the data volumes are manageable to transmit, and where the AI does not need to trigger a control action.
Practical examples: production scheduling, fleet-level anomaly trend analyzis, quarterly maintenance planning, management dashboards. These can wait. Sending data to the cloud, processing it, and returning a recommendation works fine when the recommendation informs a decision a human makes over hours or days.
When edge AI is required
Edge AI is required when the model needs to trigger a control action faster than a cloud round-trip allows, when connectivity is unreliable, or when the data cannot leave the asset.
Control loop timing. A battery storage system adjusting to a grid frequency deviation needs to respond in under 42ms. A CNC machine control loop runs at 1–10ms. A drill floor safety system needs 150ms. None of these can absorb a cloud round-trip and still meet the timing requirement. The model runs at the asset or the control loop does not close.
Connectivity. An offshore rig on a GEO satellite link has 600ms+ latency. A vessel in a port without reliable cellular has intermittent connectivity. A remote pumping station may have no connectivity at all during certain periods. A cloud-dependent AI system stops when the uplink does. An edge-resident model keeps running.
Data sovereignty. Some operators cannot send raw operational data to a third-party cloud. The data is processed locally and what travels upstream is structured output. The model lives at the asset because the data has to.
The architecture Helin uses
On the Helin platform, AI models deploy as containerised applications to the edge device. The model runs locally, processes sensor data continuously, and triggers control actions at the asset with no cloud round-trip. When the decision is made, the result goes to the fleet management layer: an alert, a log entry, a dashboard update.
The control loop closes in under 100ms. At BP's drilling operations, Red Zone Manager works this way: camera feed, vision model, light-fixture activation in under 150ms. The cloud does not participate in that decision.
Sunrock's Smart Grid Manager runs edge AI across 300 solar farms. Each farm executes curtailment and trading decisions locally. The central platform manages the fleet and reports on outcomes.
Two tests for an architecture decision
For each AI use case in scope, ask what happens when the uplink drops. If the use case stops working, it belongs at the edge. If it returns a result a few minutes later with no operational consequence, cloud is fine.
Then ask what the latency requirement is. Under 200ms almost certainly means edge. Over one second and cloud is likely viable. Between 200ms and one second is worth testing against the specific equipment.
Most industrial operators with distributed assets need both, running in the same platform. The question is which platform manages them together.
Back to the main guide: Industrial edge computing, what it is and how it works
By Helin