At 25 cameras and 1080p, you are pushing 24 terabytes a month upstream just to do detection in the cloud. Edge inference collapses that to single digit gigabytes.

A 1080p H.264 stream at 15 fps with reasonable scene motion lands in the 2 to 4 Mbps range, peaking higher when there is a lot of motion or detail. H.265 cuts that roughly in half (1 to 2 Mbps for the same scene) but only some recorders and cameras decode it cleanly without dropped frames. A 4K stream at the same fps is in the 8 to 12 Mbps range. Most multifamily fleets run 1080p H.264 because it is the path of least resistance with five year old recorders.

Twenty five cameras at 3 Mbps average is 75 Mbps continuous outbound from the property. Continuous outbound at that rate is 24 TB per month per property. At a typical 0.05 dollars per GB cloud egress, that is 1,200 dollars per month per property in transit alone, before storage. Storage on the cloud side at 30 day retention is another 24 TB at the cloud's data rate. Most multifamily portfolios cannot absorb that, which is why most cloud-AI pitches quietly assume sub-720p streams or motion-only upload, both of which damage the detection workload.

Almost all of it. The 25 streams stay inside the property network, where they are essentially free (the LAN is already paid for). The edge box reads the recorder's HDMI output (the same image the wall monitor shows), runs detection locally, and sends only the small structured event payload (a few hundred bytes per event plus a thumbnail image) to the operator. The transit drops from 24 TB per month to single digit GB per month. The cloud storage drops from 24 TB to roughly 50 to 200 MB per month of thumbnails. The forensic record (full pixel history) lives on the existing on-prem recorder, where it always was.

Yes, and that is the point. Most multifamily properties already have a recorder with a four to twelve TB hard drive that is doing 24x7 H.264 capture and rotating at 15 to 45 days. That investment is sunk. The edge AI overlay does not replace the recorder; it sits next to it, reads its HDMI output, and adds the intelligence layer. The recorder is still where forensic clips come from. The edge box is the new component, and it is small (a fanless mini PC), cheap (low four figures upfront), and offline-tolerant.

Below 16, the numbers are small enough that a cloud variant can be afforded by the operator if they really want one (you are talking about a few hundred dollars a month in transit, not thousands). Above 25, multiple recorders and multiple edge boxes are usually in play, and the edge architecture is the only one that scales because it parallelizes per recorder. The 16 to 25 band is the sweet spot where one recorder and one edge box cover the whole property and the cloud math is genuinely punitive.

An edge box keeps detecting during the outage because the model is on the box and the camera feed never crossed the WAN to begin with. Events get appended to a local file with a monotonic counter (one line of JSON per event). When the link returns, the drain worker walks the file forward and posts the events in strict order. The dashboard backfills, the WhatsApp messages catch up, and nothing is dropped. Cloud-AI architectures cannot do this because the inference itself is on the other side of the dead link.