We have asked this question again in every project for five years and the answer has come out the same almost every time: video is processed on site, events and metrics go to the cloud. That is not an ideology; it is the result of four calculations.
The bandwidth arithmetic
A single 1080p camera produces 2–4 Mbps with H.264. A store with 16 cameras means 48–64 Mbps of continuous upstream traffic. Most retail stores in Türkiye have asymmetric lines; download may be 100 Mbps, but upload is 10–20 Mbps. So you cannot even move a single store to the cloud. A factory may have a wider line, but a 60-camera site goes above 200 Mbps and that traffic competes with ERP, MES, telephony and everything else.
Downscaling the stream before sending it is an option; people counting works at 640×360. But PPE detection, plate reading and quality inspection do not work at that resolution. Cloud is possible for part of the project, not all of it, and running two architectures side by side is a burden of its own.
Latency
When the distance between a forklift and a pedestrian drops under 2 metres, the warning has to reach the equipment within 300 milliseconds. You cannot guarantee that on a stream that goes to the cloud and back; the average may be fine, but the worst case reaches seconds, and in safety the worst case is what matters. In retail metrics latency is irrelevant; 5-minute aggregates are enough. Again the decision depends on the application.
Data protection
Raw video is personal data. As long as it does not leave the store or the factory, the processing activity stays inside the premises and retention and access are managed from one point. Once video goes to the cloud, a whole series of topics opens: controller and processor roles, the country of the server, transfer agreements. The first question a customer's legal team asks is "where does the video go"; the answer "nowhere, only numbers come out" halves the meeting.
We do no face recognition and processing is anonymous, but even that does not change the fact that the footage itself is personal data. Processing on site and never storing the video is the cleanest path.
Three-year cost
On the edge side the cost comes up front: one GPU server sized to the camera count, a single machine for 8–16 cameras. On the cloud side the cost comes month by month: GPU instance rental, data egress, upgrading the site's line. A rough comparison for a 16-camera location:
| Item | Edge (3 years) | Cloud (3 years) |
|---|---|---|
| Processing hardware / GPU rental | One-off server | Monthly GPU instance × 36 |
| Bandwidth | Existing line suffices | Symmetric line upgrade × 36 |
| Storage and egress | None, an event is a few KB | Continuous |
| Maintenance | Remote updates + one site visit a year | None |
I have deliberately left out the figures; GPU prices and exchange rates move every quarter. The ratio does not move: edge pays for itself in 12–18 months and over three years lands around a third of the cloud cost. In a single location the difference is small; across a 40-store chain it is large.
Where the cloud is right
- When there are 2–3 cameras and the line is adequate, putting a server on site is overkill. A small café, a single-entrance office: cloud processing makes sense.
- Multi-location comparison and forecasting happen in the cloud. Without collecting the metrics of 40 stores in one place there is no benchmarking, no demand forecasting, no campaign-impact measurement. But what goes to the cloud is the metric, not the video.
- Model training happens in the cloud or at head office. Anonymised, labelled samples are collected from sites, the model is trained centrally and goes back to the sites as an update.
- Temporary projects. For a trade fair or a one-month campaign measurement, renting cloud capacity beats buying hardware.
The line we draw
On site: stream intake, decoding, detection, tracking, rule engine. In the cloud: metric aggregation, multi-location dashboards, forecasting, model training, remote updates. The name for this boundary is hybrid, but the weight is clearly on site. When a customer says "we want everything in the cloud", we do the bandwidth arithmetic together first; usually the arithmetic makes the decision.