AI features depend on a dependable video path. Plan the camera stream, network links, power, recording, processing and recovery together. A system that works on one demonstration channel may fail when every camera is active.
Draw the four data paths
Identify where video is recorded, where AI runs, where searchable metadata is stored and how users authenticate. Draw which links carry live streams, cloud uploads, events and playback. A hybrid system can record locally while sending metadata or selected video elsewhere; the word “hybrid” alone does not specify traffic or failure behavior.
Network and PoE checklist
| Component | What to verify |
|---|---|
| Camera connection | Link quality, cable suitability, port speed, supported power requirements and installation conditions. |
| PoE switch | Per-port capability and total power budget under expected loads, including heaters or accessories where applicable. |
| Uplinks | Combined camera traffic, viewing traffic, additional AI streams and headroom. |
| WAN | Actual upload capacity, other business traffic and the proposed cloud upload pattern. |
| Network separation | Approved access between cameras, recorder, AI processing and user systems; no unnecessary direct internet exposure. |
| Time and naming | Consistent time settings and meaningful camera names for search, event correlation and exports. |
Have the responsible network team validate the design. An unused physical port does not prove sufficient power or uplink capacity. A separate AI appliance may request another stream from each camera; confirm simultaneous connection limits and whether the recorder can supply the supported stream instead.
Estimate recording storage, then validate it
For continuous recording, a rough decimal estimate is: storage in TB = cameras × average Mbps × 86,400 × days ÷ 8,000,000. This estimates video payload only. Add the actual system’s overhead, audio if used, metadata, storage protection, spare space and growth requirements. Variable bitrate changes with scene activity and settings.
Illustrative calculation: 16 cameras averaging 4 Mbps for 30 days produce approximately 20.7 TB of video payload. That is not a recommendation to purchase a 20.7 TB array. Usable capacity differs from drive label totals, and the recorder’s retention behavior must be tested.
Event recording can reduce volume but makes retention depend on trigger behavior and pre/post-event settings. Confirm whether recording continues when the AI component is unavailable. Keep metadata retention separate from footage retention: a search result without its video may have limited value.
Size AI processing for the actual workload
Compute demand depends on the model, resolution, frame rate, codec and number of concurrent analyses. Ask the supplier to state the supported configuration and spare capacity. Do not treat a GPU model or a headline channel count as a universal sizing rule. Decoding, recording and indexing can compete for resources.
For server-based search of historical footage, ask how quickly a backlog can be processed and whether live analysis slows during that work. For on-camera analytics, confirm which applications can run together. For cloud services, ask about upload, indexing delay and limits associated with the subscription.
Plan power and outage behavior
A UPS design should account for the equipment needed to preserve the required function: cameras, switches, recorder, relevant routers and any local AI appliance. Required runtime is a project decision. Have a qualified designer check the electrical and environmental requirements rather than choosing solely by a UPS’s headline rating.
Test an approved outage scenario and recovery. Determine whether recording continues, alerts are delayed or lost, local caches fill, and uploads resume. Document the expected result so an operator can distinguish a normal recovery from a failure.
Handover records
- Camera and firmware inventory, network diagram and meaningful labels.
- Measured stream settings, observed traffic and storage assumptions.
- License, compute and retention requirements.
- Owner-controlled administrator access and supported recovery process.
- Update, maintenance and support responsibilities.
- Failure-test results and remaining capacity.
See the retrofit guide before adding processing to an existing network, and the integration and ownership guide for operational boundaries.