Every video analytics programme eventually faces the same question: should inference run next to the camera, in a data centre, or in the cloud? The honest answer is usually a mix, and the right mix depends on four forces.
The four forces
- Latency: safety and security use cases need sub-second response; weekly reporting does not.
- Bandwidth: a single 1080p stream can consume several megabits per second, around the clock.
- Privacy: raw video is the most sensitive artefact in the system; metadata is far less so.
- Cost: GPUs are expensive everywhere, but egress and storage fees compound in the cloud.
A simple rule
Run inference as close to the camera as the hardware allows, and send upstream only what the next decision needs. Alerts, counts, embeddings, and short evidence clips travel well. Continuous raw video does not.
Move the intelligence to the pixels, not the pixels to the intelligence.
Where the cloud still wins
Cross-site analytics, model training, long-term trend analysis, and governance benefit from aggregation. A hybrid design uses the edge for perception and the cloud for understanding at scale.
All insights3Cubicles Team