Leo · 2026-10
Somewhere between the camera and the cloud sits a decision European buyers now make deliberately: where inference data lives. Edge AI storage is the on-premises answer, storage hardware that keeps the data an AI model consumes and produces, camera streams, event metadata, embeddings, on a device inside your building, with the inference running on that same device. A third party that never receives the data cannot be compelled to hand it over.
The term gets used loosely in datasheets, so start with the mechanics.
On-device inference means the AI model runs on the appliance’s own processor, next to the storage that holds its input. A camera stream enters the unit, the neural unit classifies it locally, and only the results, an event label and a clip, are written to disk. No stage of that loop requires the raw video to leave the building. In the woCyber line, the published A2 Pro and S2 specifications list an RK3568 system-on-chip with a 1 TOPS NPU, and the S1 performs person, vehicle and pet classification on the appliance with false alarms published at under 3 percent. For heavier workloads, the published X4 specification lists an Intel N100 with support for containerised applications and hardware 4K transcoding. In practice the NPU handles the light models and containers take the heavy ones, and that division of labour is what the storage has to be sized around.
If you want the full data-path picture, our ai-nas-local-processing-where-your-ai-data-actually-runs guide follows one frame from camera to disk.
For buyers who want the camera head, the classifier and the storage in one unit, the products/s1-security-kit combines PoE IP camera input over ONVIF with on-device classification and 7 / 14 / 30-day retention tiers, all administered locally.
Data-sovereignty debates drift into abstraction. The US CLOUD Act makes the argument concrete.
Under the US CLOUD Act, American authorities can compel certain providers to disclose data in their possession, custody or control, wherever the servers physically stand. The obligation attaches to the third party, not to the geography. Run inference on a cloud platform and the provider holds your video, your embeddings and your access logs, which is precisely the possession the Act addresses. Run the same workload on an appliance in your own server room, administered by your own staff, and no provider holds a copy, because a supplier cannot surrender data it never receives. This structural difference is where edge AI storage on your own hardware earns its place, and the tender-document version reads: data processed on hardware we own, in a room we control, under access rules we audit. The EU data-protection framework, Regulation (EU) 2016/679, then governs what you do with that data on top; the European Data Protection Board publishes its guidance at edpb.europa.eu. None of this is a certification claim, and none is made here.
The practical consequence sits one line below the legal text: the custody of your inference data is determined by who holds a copy, not by where the server stands.
Inference does not just consume data. It manufactures three kinds of it, and each needs its own capacity line in the plan.
The vector maths is smaller than most buyers expect. One 1,024-dimension vector stored as 32-bit floats occupies about 4 KB. A million vectors therefore need roughly 4 GB, and a real index with overhead lands nearer 8 to 12 GB, an illustrative planning range rather than a figure from any datasheet. Event clips usually dominate instead: 10 cameras producing 50 detections per day at 3 MB per clip write 1.5 GB daily, or about 45 GB across a 30-day window. Both figures are illustrative planning numbers; size the vector layer against your own model dimensions with our ai-nas-vector-database-sizing-embeddings-and-index-storage breakdown.
Audit logs deserve a sentence despite their size. The logs themselves are kilobytes, but they are the evidence layer for everything else in this article, so ask whether they can be exported intact and timestamped. An unreadable log supports nothing in an audit.
Capacity plans fail quietly when the network is an afterthought.
Keep the same 12-camera example on the LAN: 12 streams at 2 Mbps produce 24 Mbps of continuous traffic, a rounding error for a gigabit switch. Move inference to a cloud API and the same streams become 24 Mbps of permanent upload, plus the retrieval traffic coming back whenever staff search footage. The internal network absorbs the first case without ceremony; the second case turns your internet line into a production dependency, with a provider’s uptime attached. Custody and bandwidth settle these projects long before anyone compares model benchmarks.
Three architectures cover nearly every edge AI project. A simple mapping table does the routing.
| If your deployment looks like this | The architecture that fits |
|---|---|
| Cameras at one site, classification at the source, no server room | On-device NPU |
| Containerised models, frequent retrieval, staff on the same LAN | On-premises server |
| Burst analysis of non-sensitive feeds, elastic capacity | Cloud API |
One exception deserves saying plainly. A cloud API can be the right tool for burst analysis of non-sensitive feeds. Mixing it with sensitive streams, though, quietly changes the custody story of the whole deployment, and that is the sort of change auditors notice before IT does.
Marketing pages list TOPS. Quotations should answer five sharper questions.
For deployments that outgrow an all-in-one unit, a four-bay platform with container support and an NVMe cache option takes over the model-serving role; verify the network ports before the processor, for the bandwidth reasons in the previous section.
Sample units exist for a reason. Connect your real cameras, run your real detection rules across a full retention cycle, and let the retention tier and deletion path run unattended; the production page describes 72-hour burn-in and 100 percent outgoing inspection before shipment, and your workload profile is the part only you can test.
A blunt engineering note while you have the quotation open: skip the NVMe cache for write-mostly event recording and spend that line item on the second network port instead. Cache earns its slot on retrieval-heavy deployments, where staff search footage constantly; continuous ingest barely touches it. Ordering those two options backwards is a specification mistake a one-week sample test catches cheaply, and a tender review catches even earlier.
Picture the purchase order rather than the datasheet. A six-site dental clinic group is replacing a cloud subscription after an insurer’s questionnaire asked for evidence of data custody: fourteen cameras across reception areas, sterilisation rooms and car parks, one small server room per site, and no group IT function to run it.
The trade-off the group faced was consistency against site-by-site pragmatism. Two sites had fibre between buildings; four did not, so a centralised recorder was never viable. The eventual brief specified one appliance per site, on-device person detection so motion alerts stay useful, a 30-day retention tier with the deletion behaviour to be confirmed in the quotation, and separate per-account storage spaces so each practice manager would see only their own site’s footage while the group auditor gets an export path, a feature to confirm per SKU in the quotation.
The wiring bill nearly sank the project at site three, where the records cupboard had no spare network drops. That is the part no datasheet advertises: the last metre of cable has a habit of costing more than the first terabyte. Client records never touch the surveillance appliance; the point of the exercise is that the next insurer questionnaire now gets a one-page answer, per site, from hardware the group owns.
An appliance that never changes is an appliance that quietly ages. Updates are part of the custody picture, not an IT footnote.
The published OEM programme lists a brand OTA channel among its firmware options, which raises three questions worth asking before you sign: which components ride the update channel, who approves a model update before it reaches your fleet, and whether updates can be staged so a pilot unit leads the production batch. Firmware you cannot inspect and updates you cannot stage both weaken the custody argument this article has been building, so settle the update policy with the same rigour as the storage sizing.
Distributors and brand owners ask a different question: not whether the data stays local, but whose brand sits on the box that keeps it local.
The published OEM programme covers a white-label app, branded UI and boot animation, logo application by silkscreen or laser engraving, localised packaging and documentation, and regional power adapter variants. Firmware-side, the programme includes a branded NVR UI, custom AI event rules, retention-strategy defaults and brand-controlled OTA updates. None of those options move the data path; they change the name on it. For brand owners, the name is the product.
Interface localisation deserves a line of its own for EU deployments: the published OEM programme includes user-interface localisation across English, German, French, Italian, Spanish, Japanese, Korean and Finnish, so a German or French end customer can receive a unit whose menus already speak their language. Documentation, packaging and quick-start guides localise alongside the interface.
Commercial terms to anchor the conversation: quantities start at 100 units per SKU, variants needing new tooling or a fresh PCB respin are quoted from 1,000 units, and the woCyber product lines are certified CE, UKCA, FCC, RoHS, REACH, and WEEE. Related reading for the sovereignty angle: our eu-data-sovereignty-storage-appliance overview.
Bring three things to that first meeting: the camera list with bitrates attached, the retention tier your data-protection owner will accept, and a floor plan marking where the server rooms actually sit. Add the spare network drops while you are at it, because they cost almost nothing to pull now and a great deal to add after the walls close.
woCyber AI NAS runs detection, analysis and local storage on-device — your data stays on your own hardware, with no monthly cloud fees.
Key takeaways An FCC-recognized lab must hold the right scope for your device's rule part, not just any accreditation. A 2-bay NAS…
November 3, 2026Shared storage for deskside AI clusters refers to a storage appliance on the office LAN that holds datasets, checkpoints and embeddings for…
November 3, 2026A cloud storage subscription cost is the recurring fee a business pays to keep its files on a provider's servers, billed per…
Ask our export team for the OEM/ODM cooperation guide and a written quotation.
