Edge AI cameras vs server-side processing — which to pick for a Saudi industrial site?
Pick edge AI when bandwidth is constrained or latency must stay under 80 milliseconds — typical for remote Aramco wellpads, NEOM corridor sites, or mobile rigs. Pick server-side when you have over 24 cameras, fibre or stable Ethernet, and need a single model to evolve without touching every device. Hybrid (edge cameras for hot streams, server for the rest) is the most cost-effective architecture for any site over 80 cameras.
The architecture decision matters because it locks in years of cost and operational behaviour. Edge means the model runs on a chip inside the camera or a small box next to it. Server-side means streams are sent to a central GPU server (on-prem or in a Riyadh-region cloud) and processed there.
Decision matrix
| Factor | Edge wins | Server-side wins |
|---|---|---|
| Latency requirement | under 80 ms (e.g. crane interlock, gate access) | under 500 ms is fine (most safety alerts) |
| Bandwidth available | under 50 Mbps backhaul | over 100 Mbps |
| Camera count | 1 to 24 streams | 24+ streams |
| Number of model classes | 1 to 4 simple classes | 5+ classes or complex models |
| Model update frequency | quarterly | weekly retraining |
| Site power budget | tight (PoE only) | servers on UPS |
| Deployment locations | fragmented (many sites) | concentrated (one campus) |
Total cost of ownership — 36-camera site over 3 years
| Line | Edge (SAR) | Server-side (SAR) |
|---|---|---|
| Camera or device hardware | 360 cams × SAR 3,600 = 129,600 | 36 cams × SAR 1,400 = 50,400 |
| GPU server (one-off) | 0 | 48,000 |
| Switch + cabling delta | 4,000 | 9,000 |
| 3-year licence | 320,000 | 360,000 |
| 3-year support | 62,000 | 72,000 |
| 3-year bandwidth (remote site) | 4,000 | 54,000 |
| Total 3 years | ~519,600 | ~593,400 |
Edge wins on remote sites with metered or satellite backhaul. Server-side closes the gap as backhaul gets cheaper and breaks even at full fibre.
Hot patterns we deploy in 2026
Pure edge — small remote site
Aramco wellhead, MEWA pump station, single-tower construction site. 8 to 16 edge AI cameras, no central server. PPE and intrusion alerts go via 4G/5G to the central command. Bandwidth need under 2 Mbps.
Pure server — single campus
NEOM site office, Riyadh logistics warehouse, refinery control building. 60+ ONVIF cameras, single GPU server (NVIDIA L4 or RTX A4000), all classes processed centrally. CCTV integration takes about a week.
Hybrid — large multi-zone project
Megaproject with 200+ cameras spread over 4+ zones. Hot streams (crane zones, perimeter gates) run edge for under-50 ms latency; cold streams (parking, breakroom) feed the central server. Model updates push to edge devices over MQTT every quarter.
What people get wrong
- Believing edge is always cheaper — for under 24 cameras on a fibre site, a cheap GPU server beats edge cameras on TCO.
- Believing server is always smarter — heavy models (DINOv3, Sapiens) cannot fit on most edge devices, but lightweight YOLO-class models hit 92% mAP on edge with no real disadvantage.
- Ignoring update operations — edge fleets need an OTA update mechanism. Without one, 200 edge cameras become 200 maintenance trips.
- Skipping PDPL hosting check — server-side cloud must be in a Saudi region. See PDPL compliance.
