You do not have to upload your video to get the security camera cloud features.

In the consumer market, 'security camera cloud' means a subscription (Ring Protect, Nest Aware, Arlo Secure, Eufy Cloud, Reolink Cloud) where your cameras continuously upload video to a vendor bucket and you get a retention window, clip sharing, and an app feed. That model assumes a handful of cameras in a house with plenty of upstream bandwidth. The reframe on this page is: when you have 16 to 25 cameras on a commercial DVR, continuous upload is not bandwidth-feasible, but you can still get the cloud features (AI search, remote alerts, portfolio dashboard, ops reporting) by running edge AI on the DVR's HDMI output and syncing only classified events plus their thumbnails upstream.

On a representative 25-tile multiview, the detector sees about 412,803 candidate detections per 24 hours. The four-stage filter pipeline (overlay_mask, tile-grid zones, multi-frame persistence, tile.label dedup) drops all but about 241 of those. What leaves the property is 241 event JSON blobs (nine fields each) and 241 thumbnails cropped to 480x270 pixels of just the triggering tile. Mean upstream load is under 2 Mbps. Compare that to a true cloud camera setup pushing 16 1080p feeds, which needs 50 to 80 Mbps of sustained upload that most commercial buildings do not have.

The recorder keeps all video forever (or for its native retention window, usually 14 to 90 days depending on disk size). Cyrano does not touch that. The cloud holds the structured event stream: iso8601_ts, property, tile.label, tile.index, tile.coords, layout_id, overlay_mask, event_class, latency_ms, and one 480x270 thumbnail. That is enough to search events in natural language, chart incident trends across a portfolio, export ownership reports, and jump to the exact timestamp on the recorder when you need the full clip. The clip itself never leaves the building.

State tenant privacy laws and property insurance riders increasingly require that continuous camera footage not leave the property without cause. A pure cloud camera setup violates that by default because every minute of video is on a vendor server 24/7. An HDMI-tap event-only sync inverts the default: the raw footage stays on the recorder in the office closet, and only discrete events (with their own retention clock and audit trail) leave. That is the posture most multifamily owners and insurance carriers are already assuming when they sign off on a camera deployment.

No. Cyrano passes the HDMI signal through to the wall monitor untouched and exposes a remote live view to the dashboard, but live view is pulled on demand, not pushed continuously. When you open live view, the unit starts a short-duration WebRTC or similar pull for as long as you watch, then stops. That is a fundamentally different bandwidth profile from 24/7 upload. The rest of the time, the property's upstream link is idle except for the event stream.

Detection keeps running. Events keep getting generated, filtered, and written to a local queue on the Cyrano device. Thumbnails are generated locally. When the link comes back up, the queue drains to the cloud in timestamp order and the portfolio dashboard backfills. A pure cloud camera setup gives you zero detections during an outage because the inference is running on someone else's server that cannot see your video. The HDMI edge-tap path is offline-durable by default.

Search runs against the structured event stream in the cloud, not against the video itself. When you type 'masked man near gate last Tuesday night,' the backend resolves that to event_class in {person_in_zone, loiter}, a tile.label matching a gate camera, an iso8601_ts range, and a thumbnail appearance check. The matching events come back as a list with thumbnails, each with a deep link to the exact timestamp on the recorder. You click the link, the on-site recorder cues up that clip, and it streams to your browser only at the moment you press play. This is why search works without continuous upload.

A cloud camera rollout replaces every camera in the property, runs new cable if the existing camera count or placement is wrong, installs a new NVR, and starts a subscription per camera. That is $50,000 to $100,000+ per property and months of work. The HDMI edge-tap rollout plugs one Cyrano device into the HDMI output of the DVR you already have, passes the signal through to the guard monitor, and connects the unit to the property's network. Install is under 2 minutes per property. Hardware is $450 one-time per property. Software is $200 per month. Nothing on the camera side changes.

The overlay_mask template library covers the recorder brands commonly found in Class B/C multifamily DVR closets: Hikvision DS-7xxx, Dahua XVR and NVR, Lorex, Amcrest, Reolink NVR, Uniview, Swann, Night Owl, Q-See, ANNKE, EZVIZ, Bosch DIVAR, Honeywell Performance, and Panasonic WJ-NX. Each template knows where that recorder stamps the clock, the camera-name strip, and the channel bug so inference does not fire on overlay pixels. A custom layout outside the library gets a one-time mask calibration and then behaves the same as any supported recorder.

Cloud-relay VMS platforms still push continuous video from the recorder to the cloud, just with deduplication and adaptive bitrate smoothing in between. The property still pays the upstream bandwidth cost, and the raw footage still lives on a vendor server. The HDMI edge-tap is strictly event-only upstream. There is no relay tier running continuous video transport. The only thing the cloud sees is the filtered event stream and its thumbnails. You lose the ability to scrub video through the VMS web app, but you gain bandwidth sanity, lower monthly cost, and a tenant-privacy posture most owners already want.

Back-of-envelope: a typical multi-camera cloud plan with 60-day retention runs $10 to $25 per camera per month, so a 16-camera property is $160 to $400 per month in subscription alone, plus the one-time cost of replacing all the cameras. The HDMI edge-tap is $450 one-time for the device plus $200 per month total for the property, regardless of camera count. On a 50-property portfolio that is $10,000 per month vs anywhere from $8,000 to $20,000 per month on cloud cameras, with zero capex to rip-and-replace. The ROI line usually crosses before month 3.

Nine fields plus a thumbnail: tile.label (camera name the DVR stamps on the strip, e.g. 'Loading Dock NE'), tile.index (row-major position in the grid), tile.coords (x,y,w,h in composite-frame pixels), property (site ID), layout_id (recorder layout, e.g. '4x4-std'), overlay_mask (array of overlay classes blanked, e.g. ['clock','cam_name_strip','channel_bug']), event_class (person_in_zone / vehicle_dwell / loiter / tamper), iso8601_ts (recorder clock timestamp), latency_ms (capture-to-delivery time). The thumbnail is a 480x270 JPEG crop of just the triggering tile, stored at an S3 key the event references.