Cloud analytics runs the video analysis on remote servers — rented GPUs or a managed AI service — instead of on the camera or a local box. Frames or whole streams are uploaded to the cloud, processed there, and the results (and often the video) are stored and served back. It buys effectively unlimited, elastic compute and the heaviest models without buying and maintaining hardware on site.
The appeal is capability and elasticity: complex models, large vision-language reasoning, and bursty workloads that would overwhelm a camera or a single server are easy in the cloud, and you pay for what you use. It suits forensic, after-the-fact analysis and centralised multi-site processing well. But three meters run the whole time — compute, storage, and egress — and they decide the bill.
The pitfalls are cost and latency. A per-minute analytics API sounds cheap until it runs continuously: analysing one camera around the clock can cost on the order of thousands of dollars a month, versus roughly tens of dollars to rent a GPU that does the same work — the "per-minute API trap". And the network round-trip adds latency (often several hundred milliseconds), so cloud analytics is poorly suited to instant, on-scene alerts. Most real systems therefore filter at the edge and send only the interesting fraction up; the full economics belong to the edge-vs-cloud articles.
