2 Scality ARTESCA+ with the NVIDIA VSS Blueprint puts video understanding and immutable storage in the same building as the cameras. A camera that records all day produces a great deal of video and very few questions worth asking of it. The valuable part is the asking. Which aisle had a spill and when. How long that pallet sat in the loading bay before somebody moved it. Whether a restricted door was propped open during a shift. The reason to ask is that the answer changes what happens next. Now, the spill can be addressed before it becomes a problem. The delivery dispute is settled in minutes instead of an afternoon spent scrubbing footage. The safety review is answered with a fact rather than a recollection. Video nobody can interrogate is a cost center; video you can ask questions of is an operating tool. Getting those answers at any useful scale means running AI over the video, and the first real decision is where that happens. One common architecture sends the video to a centralized cloud service to be analyzed. It remains a useful pattern, and for some workloads it is the right one. But it increases how much video leaves the site, and it puts the running cost on a meter that never stops. A fully local deployment is another option, and it has become a practical one: ARTESCA+ with the NVIDIA Blueprint for Video Search and Summarization (VSS) runs both the understanding and the storage inside the customer’s own environment. Talking about Scality products, Scality ARTESCA is the object storage itself, while ARTESCA+ is the platform that combines ARTESCA with the Kubernetes workload it carries. What follows is the NVIDIA VSS workload validated on that baseline. Keeping it local has five consequences worth being explicit about: Privacy and residency become architectural questions rather than contractual ones. The bandwidth you would have spent shipping video stays yours. The operational cost changes shape, from a meter to an asset. Sizing becomes answerable up front, from two variables. Retention stops being an inference problem and becomes a storage one. 1. Privacy and residency become architectural questions When VSS is configured with local models and Scality ARTESCA storage inside the customer’s environment, recorded video can remain within the customer-controlled location. That reduces external data movement, and where a jurisdiction expects data to stay within it, a system standing inside that jurisdiction is a straightforward way to meet the expectation: no region to select and no residency addendum to negotiate. That is a starting position, not a compliance argument. The outcome still depends on who has access, how keys are managed, how support connects, where backups go, and how the system is operated. What the architecture does is put every one of those questions inside the customer’s control. That control has a second effect, and this one shows up on an invoice long before it shows up in a compliance review. 2. The bandwidth you keep Continuous video from one camera is a continuous outbound stream. Multiply it by the cameras on a site and the sites in an estate, and the network you are paying for is sized to carry raw material rather than answers. Analyzing on-site inverts that. What leaves the building is the result: an alert, a summary, a search index entry. The result is smaller than the footage by orders of magnitude. The video stays where it was recorded, which is also where it usually has to be retained. Bandwidth is the cost you can see on a bill. The one that compounds is what you pay to keep asking. 3. The operational cost changes shape Cloud pricing rewards workloads that are bursty: pay for the peak, pay little at the trough. Video understanding is the opposite shape. Cameras do not stop, so the inference does not stop, and neither does the meter. Storage of continuous video accumulates, and retrieving it later is charged by the byte. The honest comparison is multi-year total cost rather than a single line item: cloud compute, storage, retrieval and network charges on one side; hardware, software, support, power and refresh on the other. What the local option changes is the shape of the spend: predictable and largely fixed, which is the favorable side of the trade when utilization stays consistently high. For a workload as sustained as video analytics, it usually does. A fixed cost is only reassuring if you know what you are fixing it at, which is the next question. 4. Sizing becomes answerable up front Two variables drive the system on site: how many camera streams need to be understood at once, and how long the recorded video must be kept. The first sizes the accelerator, the GPU doing the inference. The second sizes the storage. Both are answerable before anything is bought, and a later post in this series covers how that sizing is done. Sizing tells you what to install. It does not tell you what happens to the recording afterward, and for a good number of these deployments that is the part the auditor actually asks about. 5. Retention becomes a storage problem Video that may later be needed as a record has to be kept for a defined period, unaltered and undeleted. No inference engine can guarantee that. Storage can. Scality ARTESCA adds storage-enforced retention through S3 Object Lock: in compliance mode, a protected object version cannot be overwritten or deleted before its retain-until date, including by an administrator. Governance mode is available where an authorized bypass is wanted instead. Local inference and immutable storage answer two different requirements. One is understanding the video; the other is being able to stand behind the recording afterward. The second is a storage property, and it is the one this architecture adds. Putting a GPU next to the camera does not address it. From consequences to a system somebody has to build Those five make the case for keeping video AI on site. They are also, taken together, the reason it has been awkward to do: somebody has to make an inference stack and a storage platform work as one system, in a single building, at a size that one location can justify paying for. Sizing it, keeping the recording immutable and surviving an audit are not features of either half on its own; they appear only where the two are properly joined. That joining is the work Scality has done with NVIDIA, and tested. What Scality validated Scality has validated two deployment topologies with the NVIDIA VSS Blueprint: A converged host: inference and storage on one machine. The natural fit for a single site with a bounded number of cameras and a self-contained retention requirement. A separate VSS host against an external Scality ARTESCA cluster: the inference machine sits at the edge while recorded video lands on a storage cluster that can be shared, grown and retained independently. The right shape when the retention requirement outlives, or outgrows, any one appliance. Here inference and storage are deliberately separate systems, within the same customer-controlled architecture. We validated both because sites differ in one way that changes the answer: whether the retention requirement is bounded by the site or outlives it. A single store keeping thirty days of footage and a regional operation keeping years of it are not the same system, and forcing one shape onto both means either over-building the small sites or capping the large ones. Validating both means the choice is a sizing decision rather than an architectural compromise. In both topologies the video is understood and retained inside the customer-controlled site. What differs is whether storage is part of the same machine or a cluster of its own – and recorded video cross the same S3 boundary either way. SCALITY ARTESCA+ WITH THE NVIDIA VSS BLUEPRINT – VALIDATED TOPOLOGIES This is an integration validation, not a performance benchmark. In our lab, a single NVIDIA RTX PRO™ 6000 Blackwell GPU ran five camera streams continuously with VSS recording to Scality ARTESCA S3 throughout, without interruption. That proves the two systems work as one. Build on it If you build video analytics, computer vision or physical-security software, the useful property of this arrangement is how ordinary the storage boundary is. Recorded video is exposed through an S3 interface, so you start from something you already know and can validate your application against a defined configuration of this stack. We are looking for software vendors to do exactly that. Here is what that means in practice, and what happens after you get in touch: You tell us what you have built: your product, the use case, and roughly how many camera streams you are aiming at. We give you a place to test. S3 endpoint credentials on the Scality lab. Our engineers work the integration with you, directly. If it validates, we certify it as Validated Design. A documented integration on our Certification list, which is where customers and our own field teams look to see what works. If your application would be better with video understanding underneath it, or you already have one and want it running on a system a customer can put in a store, a warehouse, a hospital or any other edge location, we would like to hear from you. Request a validation slot.
