Edge AI on cameras: the bandwidth math when you have 16 to 25 feeds per property

1. The cloud AI bandwidth bill on a 16 to 25 camera property

A single 1080p H.264 camera at a moderate bitrate (4 Mbps continuous) consumes around 1.3 TB per month. A property with 16 cameras streaming continuously to a cloud inference server is looking at 21 TB a month uplink. At 25 cameras it is 33 TB.

Most US commercial broadband packages are not engineered for sustained 80 to 130 Mbps uplink. Even on a 1 Gbps fiber drop, the upload pipe is often capped at 100 to 200 Mbps and is shared with everything else the property does. Symmetric fiber that can sustain this kind of upload runs $400 to $1,200 a month commercial.

Add the cloud egress and inference compute on the receiving side, and the cost per camera per month for cloud-AI surveillance lands at $80 to $200. Across 20 cameras that is $20,000 to $48,000 a year for one property.

2. The edge AI architecture, and why bandwidth drops 95 percent

Edge AI flips the pipeline. Inference runs on the camera, on a small box at the property, or on a NVR with a built-in inference accelerator. Frames never leave the property unless an event fires.

An event is small: a 10 to 30 second clip plus metadata. Even at 100 events per day, the upload is in single-digit gigabytes per month per property. Compared to 21 TB on cloud-AI, the bandwidth requirement drops by 95+ percent.

This is why the architecture matters. The bandwidth difference is not a 2x or 5x improvement. It is two orders of magnitude. The economics that did not work for cloud-AI at scale work for edge-AI at the same scale.

3. Latency and outage resilience

Cloud-AI has a network round trip baked into every detection. Frame leaves the property, hits the inference server, returns a result. End to end latency is 200 to 800 ms in the best case. During a network outage the entire pipeline is dead.

Edge-AI runs inference at the camera or in the property's local box. End to end latency is 30 to 100 ms. During a network outage the inference keeps running, alerts queue locally on disk, and drain when the link returns.

For real time intercept (catching a trespass while it is happening, not after), the latency difference matters. For outage resilience (the property's internet drops at 3 AM during a storm), the difference is the entire system staying alive vs going dark.

4. Hardware shapes for edge AI on existing camera systems

The hardware comes in three shapes. First, AI cameras with on-device inference (Verkada, Rhombus, AXIS analytics) which require replacing the existing cameras. Second, AI NVRs (Hikvision DeepinMind, UniFi G5+) which require replacing the recorder. Third, edge boxes that tap the existing DVR's HDMI output (Cyrano, a few others) which require neither.

The third shape is the only one that preserves the existing camera + DVR investment. It taps the multiview composite the DVR already produces for the office monitor, runs object detection on the composite, de-tiles each detection back to a per-camera ID using a fixed grid template, and produces alerts.

On a 16 to 25 camera property, the third shape pencils out at $450 in hardware and $200 a month in software. Compared to a $40,000 camera replacement and $1,500 a month in cloud fees, the gap is large.

5. Has on-device AI caught up to cloud AI in 2026?

Two years ago edge AI inference quality was meaningfully behind cloud AI. The compute was tighter, the models were smaller, and accuracy on hard cases (low light, partial occlusion, distant subjects) was visibly worse.

In 2026 the gap has narrowed substantially. Quantized YOLO variants, distilled detection models, and dedicated inference accelerators (Hailo, Coral, ONNX runtime on Apple Silicon) hit 92 to 96 percent of cloud-AI accuracy on standard surveillance benchmarks. For most security use cases, the residual quality gap is smaller than the variance from camera placement and lighting.

The remaining cloud-AI advantage is on long tail tasks: face recognition at scale across thousands of identities, license plate recognition at acute angles, fine-grained behavior classification. For ordinary trespass / loitering / vehicle alerting on multifamily and small commercial sites, edge AI is at parity.

6. Operational tradeoffs nobody mentions in the brochure

Edge AI means model updates ship as OTA patches to the device. There is no centralized model rollout. The vendor needs a working OTA pipeline; if they do not have one, the model is whatever it shipped with.

Edge AI means thumbnails and clips live on the device. Disk space matters. A device with a 256 GB SSD can hold 60 to 90 days of thumbnails for a 25 camera site. Beyond that the device rotates the oldest data out.

Edge AI means the inference compute is paid for once at hardware time, not monthly. The economics scale better as the camera count goes up, not worse.

7. When cloud AI is still the right call

Cloud AI wins for sites with 1 to 4 cameras where the bandwidth cost is small and the cloud convenience is real. Cloud AI also wins where centralized cross-site search matters: a multi-property operator who needs to query 'every clip across 50 sites where a red sedan appeared this week' is well served by cloud aggregation.

For everything between 4 and 25 cameras per site, edge AI dominates on cost, on outage resilience, and on bandwidth. Above 25 cameras per site, hybrid setups (edge AI for first-pass detection, cloud for retention and search) usually win.