AI NVR Appliance Manufacturer: What Integrators Verify

Leo · 2026-10

An AI NVR appliance manufacturer builds the recorder, the on-device inference stack and the storage layer into one unit, so integrators vet three things before discussing price: how detection is tested, how cameras onboard, and how long footage stays local. A datasheet covers none of the three in useful detail.

What Does an AI NVR Appliance Manufacturer Actually Ship?

An AI NVR appliance is a recorder that classifies events on the device itself instead of shipping frames to a cloud service for analysis. The manufacturer ships three layers in one chassis: camera intake, inference, and retention.

Each layer produces a different question at the counter. Intake settles which cameras connect and how they draw power. The inference layer settles whether an alert deserves a human look. Retention settles how far back an incident can be reconstructed when a client calls six weeks later.

The published S1 specification describes this combination directly: an edge AI NVR paired with private-cloud storage, on-device classification of people, vehicles and pets, a false-alarm rate below 3%, PoE IP camera intake over ONVIF, and retention tiers of 7, 14 or 30 days. Integrators who want the full picture can review the S1 security kit alongside their own camera list.

Two design choices follow from that description. Computation stays on the appliance, so a site keeps recording and classifying when the internet drops. Accounts are physically isolated on the device, which matters when a building manager and a tenant share one recorder but must not see each other’s footage.

Which Claims Need a Bench Test Rather Than a Datasheet?

Three claims need a bench test: the false-alarm rate, the camera onboarding time, and retention behaviour under continuous write. Each one is cheap to test with eight cameras and one afternoon.

The false-alarm figure is the one buyers quote most and understand least. A published rate below 3% describes classification accuracy on the vendor’s test material, and the number means little until the integrator repeats a version of the test in the client’s lighting, with the client’s camera angles. Ask how the rate was measured, on how many events, and across which scenes.

Onboarding time is measurable in minutes. A platform that publishes QR pairing and automatic camera discovery should absorb eight cameras faster than a technician can label them in the client’s documentation. A platform that needs a static IP per camera before it will talk to the recorder will cost a second visit.

Retention behaviour shows up in week two, not hour one. Continuous write plus event-triggered clips is a different disk workload from file archiving, and the only honest test is to run the array for a week at the planned bitrate, then read the oldest footage back.

How Many Terabytes Does Sixteen Cameras Need?

Sixteen cameras recording at 4 Mbps for 30 days need about 20.7 TB before any headroom. The number comes from one formula that works for any site.

Storage in terabytes = cameras x bitrate in Mbps x 86,400 seconds x retention days / 8 / 1,000,000,000,000. The divisor of 8 converts bits to bytes, and the final divisor converts bytes to terabytes.

Worked example, using a hypothetical 16-camera site at 4 Mbps and 30-day retention:

• One camera per day: 4 x 86,400 / 8 = 43,200 megabytes, or 43.2 GB.
• Sixteen cameras per day: 16 x 43.2 = 691.2 GB.
• Thirty days: 691.2 x 30 = 20,736 GB, or 20.7 TB.

Drop retention to 14 days and the same site needs 9.7 TB, because the daily figure does not change. Add a 10 percent allowance for index overhead and clips that run past the retention window and the 30-day figure becomes 22.8 TB. That allowance is a planning assumption, not a vendor specification.

The formula also shows where sites go wrong. Doubling camera count doubles storage; doubling bitrate doubles storage; moving from 14 to 30 days doubles storage again. A client who asks for 30 days at 8 Mbps on 32 cameras has multiplied the original estimate by eight, and no retention tier rescues a recorder that was sized for the wrong number.

Camera Onboarding: Where PoE and ONVIF Decisions Land

PoE and ONVIF decide how much of the installation is cabling and how much is configuration. PoE carries power and data on one cable, which removes a power run per camera; ONVIF gives the recorder a published, profile-based way to talk to cameras from different vendors.

The ONVIF question to ask is which profiles and which versions the appliance implements, and whether discovery works with the specific camera models on the drawing. A generic “ONVIF supported” answer is not enough for a 40-camera retrofit, because discovery failures surface on site, not in the quotation.

Terminology matters here. ONVIF support is a capability that a device implements; it is not a certificate a vendor holds. The ONVIF organisation publishes the profiles and their scope on its own site (ONVIF profiles, retrieved October 19, 2026), and an integrator can check a claimed profile against the standard before accepting the claim.

Power budget belongs in the same conversation. A PoE switch or injector plan that assumes every port draws the same wattage will fail on the first pan-tilt-zoom camera added after handover, and the recorder is usually blamed for a power design that was never written down.

Factory Evidence an Integrator Can Request

Factory evidence is what separates a trading company from an AI NVR appliance manufacturer, and an integrator can request it before the first purchase order. The requests cost one email each.

• Production sequence: SMT lines, assembly, functional test, a 72-hour burn-in room, and 100 percent outgoing inspection with batch reports archived.
• Audit access: third-party factory audits through SGS, Bureau Veritas or TÜV, video audits, and visits by the customer’s own engineers.
• Certification documents: CE, UKCA, FCC, RoHS, REACH and WEEE, with the CE declaration of conformity, FCC test report, and RoHS and REACH declarations travelling with the shipment.
• Track record: 15+ years of OEM and ODM work, 500K+ units shipped, and sales into 30+ export markets.

An integrator buying a private-label programme is buying the factory’s process, and a supplier that cannot state where development ends and production begins is describing a supply chain it does not control. woCyber, the export brand of a Beijing-based storage hardware manufacturer, runs research and development in Beijing and mass production in Shenzhen, and the factory page describes the burn-in and inspection sequence in the same terms listed above. NIST’s guide to general server security treats physical and logical access control as part of the server security baseline (NIST SP 800-123, retrieved October 19, 2026), and the same logic applies to a recorder: the factory that builds it should be able to describe who touches it before it ships.

Channel Terms That Decide the SKU

Channel terms decide the SKU long before the hardware does, and four of them are worth settling in writing at the quotation stage.

Minimum order quantity. The published OEM terms list MOQ from 100 units per SKU, rising to 1,000 units where the project needs a new PCB respin or new tooling. An integrator planning a 60-unit pilot should know that the pilot does not clear the threshold, and should price it as a sample-and-evaluation phase rather than a production run.

Customisation scope. White-label app listings, branded UI and boot animation, logo by silkscreen or laser engraving, localised packaging and documentation, regional power adapters, branded NVR interface, custom AI event rules, retention defaults and branded over-the-air updates all sit inside the OEM programme. Each item changes the project timeline, so the scope belongs in the first quotation rather than the third.

Tooling and NRE. Non-recurring engineering is quoted per project and credited against agreed volume, and customer-specific tooling transfers to the customer once paid. Those two sentences decide whether the integrator owns the mould at the end of the programme.

Documentation for retail. EAN and UPC codes, energy and compliance labelling, WEEE producer responsibility marking and country-of-origin marking are the items a distributor asks for at onboarding, and they are cheaper to settle with the factory than with a labelling contractor.

Two Deployment Paths Compared

Integrators usually choose between a combined recorder-and-storage appliance and a NAS that accepts cameras. The table compares the published specifications without ranking them, because the right path depends on the site.

The combined path suits sites where the camera count is stable and the client wants one box and one support call. The NAS path suits sites that already own cameras and want storage first, with camera recording as an additional duty; the A2 Pro compact NAS is the published two-bay option for that role. Above sixteen cameras at 4 Mbps, price the uplink separately from the drives, because a single gigabit port sets a ceiling that extra capacity cannot lift.

What Belongs in the First Email to a Supplier?

Six questions belong in the first email, and each one should require a written answer rather than a brochure.

1. Which ONVIF profiles does the appliance implement, and which camera models have been tested for discovery?
2. How was the published false-alarm rate measured, and on what sample?
3. What are the retention tiers, and can the default be changed per project?
4. Which certifications ship with the unit, and which documents travel in the box?
5. What is the MOQ per SKU, and what triggers the higher threshold?
6. Which customisation items change the lead time, and which are catalogue options?

Two answers are worth more than the rest. A supplier that describes its burn-in and outgoing inspection routine in specific terms has a factory process it can point to, and a supplier that separates compliance documents from compliance claims understands the difference between shipping a declaration of conformity and holding a certificate.

Integrators building a repeatable programme usually read three further documents before committing: how cctv-storage-appliance-oem-for-integrators“>CCTV storage is specified for integrator projects, which ai-nas-security-features-to-verify-before-deployment“>security features to verify before deployment, and how oem-odm“>the OEM and ODM programme is structured from sample to volume.

Five Answers That Save a Site Visit

Does an AI NVR need a cloud subscription to work?

The published S1 specification describes 100 percent local computation, so recording and classification run on the appliance without a cloud account. An integrator should still ask which optional off-site copy targets the firmware exposes, and treat any such target as an add-on to the local design rather than a dependency.

What does a false-alarm rate below 3% actually measure?

The figure measures how often the on-device classifier raises an alert that a human judges to be wrong, on the vendor’s test material. Integrators should repeat a version of that test with the client’s cameras and lighting before treating the number as a site-level expectation.

Can an existing camera estate be reused with an AI recorder?

Reuse depends on ONVIF support in the cameras and on the profiles the recorder implements, not on the camera brand. The practical test is discovery: connect three cameras of each model, confirm the recorder finds them without static addressing, and record for an hour before committing the full site.

How much storage does a 30-day retention policy need?

Storage in terabytes equals cameras x bitrate in Mbps x 86,400 x days / 8 / 1,000,000,000,000. Sixteen cameras at 4 Mbps for 30 days need 20.7 TB, and the same site at 14 days needs 9.7 TB, before any headroom the integrator chooses to add.

What should an integrator ask about account isolation?

Ask whether accounts are separated by software permission or by physical isolation on the device, because the two answers behave differently when a tenant or a sub-tenant shares one recorder. The published S1 specification from woCyber describes physically isolated accounts, and the integrator should confirm the mechanism on the demonstration unit.