The right AI video analytics architecture is the one that can sustain your security operation under the conditions your organization actually faces. Connectivity, response requirements, control over sensitive information, and the people available to maintain the system should determine where processing runs.

I recommend starting with a practical question: if the connection to an external service disappears, what must this site still be able to do? The answer often clarifies the deployment decision more effectively than a comparison of software features.

A remote industrial facility, a retail chain, and a restricted government campus may need similar detection capabilities. Their infrastructure, operating responsibilities, and acceptable dependencies can be very different.

Understand what each deployment model means

Edge AI processes video close to the camera, either inside the camera or on a nearby appliance. On-premises AI typically uses servers within an organization’s facility or private data center. These categories overlap: an edge appliance installed at a site is also on-premises infrastructure.

Cloud AI video analytics runs processing on remotely hosted infrastructure. A cloud-managed system, however, may still analyze and record video locally. Buyers should establish separately where inference, recording, search, administration, and alert delivery happen.

Hybrid video surveillance distributes those functions. A site might detect intrusions and retain recordings locally while sending approved event information to a central service. That service could run in a public cloud or an organization’s own data center.

The useful comparison is therefore about individual workloads and dependencies. A label such as “cloud-enabled” does not describe the complete operating model.

Use operating requirements to identify candidate architectures, then validate the exact configuration. Different workloads may require different deployment locations.

Start with connectivity and continuity

Continuous video creates sustained network demand. As an illustrative calculation, 100 cameras averaging 4 Mbps generate approximately 400 Mbps of video traffic, or 4.32 TB per day before network overhead. Actual requirements depend on encoding, scene activity, and which streams leave the site.

Local inference can reduce external traffic when only alerts, metadata, or selected clips are transmitted. But an architecture must distinguish detecting an event from delivering it. A disconnected appliance may continue detecting and recording while a remote operator remains unaware of the incident.

For every critical workflow, specify what happens during an outage: who receives the alert, where evidence is retained, how long local buffers last, and how synchronization resumes. Queued alerts should retain their original timestamps and be identifiable as delayed events.

Resilience also includes power, switches, storage, and compute. Local processing needs backup power and a recovery plan. Cloud processing needs a viable network path and an appropriate service design. Neither location makes the complete detection-to-response chain inherently resilient.

Decide who will operate the system

An on-premises deployment gives an organization direct control, together with responsibility for hardware replacement, operating-system updates, model deployment, monitoring, and recovery. A capable integrator can perform much of this work, but the contract must make ownership explicit.

Cloud services can reduce local infrastructure work, especially across many small sites. Responsibilities still depend on the service. AWS’s shared responsibility model illustrates why consuming cloud infrastructure does not transfer every security obligation to the provider. Running your own analytics on rented cloud servers is operationally different from buying a managed application.

Ask who investigates a failed stream at 2 a.m., approves updates, restores configuration, and tests the system after maintenance. If those responsibilities are unclear, the architecture is incomplete.

Evaluate scale through the full cost of operation

Camera count alone is a poor sizing metric. Capacity depends on resolution, decoding, analyzed frame rate, concurrent models, scene complexity, and the resources reserved for failover. Adding video-language search or report generation can change the workload substantially.

Edge deployments expand site by site, but introduce a fleet of devices to maintain. Central on-premises servers can pool resources, although their failure may affect more cameras. Cloud infrastructure can accommodate changing demand, subject to available capacity, service limits, connectivity, and cost.

Compare options over the same planning period and service level. Include hardware refreshes, licenses, integration, electricity, cooling, connectivity, storage, cloud compute, data retrieval or transfer charges, support, and staff time. Include the cost of resilience and eventual migration.

A low subscription price can look different after continuous processing and retention are included. Equally, owned hardware can be expensive when it is underused or difficult to support. Require cost scenarios for normal operation, expansion, and recovery before approving the investment.

Sovereign AI requires control across the system

Sovereign AI brings questions of authority and operational independence into the architecture discussion. NVIDIA describes national AI capabilities in terms of domestic infrastructure, data, and expertise in its discussion of sovereign AI.

For a security buyer, this becomes a set of concrete questions. Where are video, derived metadata, embeddings, logs, and backups stored? Who can administer the environment? Who controls encryption keys and model updates? Does any component send telemetry or call an external inference service?

Data residency answers where information is located. Sovereignty also concerns who can access it, govern its use, and keep the service operating. A domestic hosting location alone does not settle those questions. Assess provider access, contractual controls, applicable jurisdiction, and technical dependencies together.

Sovereign requirements do not automatically mandate physical isolation. Some organizations can use an appropriately governed cloud. Others require infrastructure under their direct control, with no external connection.

Design air-gapped AI as a complete operating environment

An air gap is stronger than a firewall rule or a private cloud connection. The NIST glossary describes physical separation between systems, with any transfer across the separation performed manually under human control. An isolated security network can still contain many interconnected cameras, servers, and workstations.

For organizations requiring everything to remain air-gapped, the entire operational stack must function inside that boundary. In these environments, sovereign video intelligence means retaining control over video, AI-generated insights, and the infrastructure and services needed to produce and use them, without reliance on external connectivity.

This includes model files, video management, storage, operator interfaces, identity services, logging, and required infrastructure services. Licensing and activation need supported offline procedures. A local analytics engine dependent on cloud authentication or an external language model would leave a critical gap.

I would require four elements in the deployment framework:

● A dependency inventory identifying every service, software package, model, and attempted external connection. ● A controlled update process using authenticated packages, approved transfer media, malware checks, staging tests, and a documented rollback path. ● A local support and recovery plan covering diagnostics, access control, backups, spare capacity, and replacement hardware. ● An acceptance test performed without external connectivity, including cold starts, alert delivery, maintenance, and recovery from failure.

Isolation reduces exposure through external networks, but it does not remove risks from removable media, privileged users, or unpatched software. Maintenance must remain workable throughout the deployment’s life.

Scylla’s February 2026 partnership announcement with Carahsoft describes on-premises video analytics without cloud dependencies for government environments. The procurement lesson is to verify the exact proposed configuration, including integrations, rather than assuming every feature of a broader product portfolio operates offline.

Let the site determine the architecture

Local detection alone does not guarantee response. A resilient hybrid design provides a local response path and makes delayed event delivery visible to central operators.

Consider four illustrative scenarios. These are decision examples, not descriptions of specific customer deployments.

A temporary construction site has a small camera network and intermittent cellular connectivity. Edge processing is a strong candidate because detection can run locally. The design still needs a practical response path during an outage. Scylla Asteria illustrates this pattern: video analytics runs on the device, events can be retained locally, and alerts can be forwarded when the configured connection is available.

A small professional-services office has reliable connectivity, modest analytics requirements, and no server administrator. A managed cloud service may be appropriate if its data policies, connectivity arrangements, and outage behavior meet the business’s needs. The organization should establish which functions stop when connectivity fails and whether that interruption is acceptable.

A retailer with dozens of branches needs consistent policies and centralized oversight. A hybrid design can retain time-sensitive processing and recordings at each branch while centralizing approved events and reporting. The value depends on disciplined device management and a clear rule for what information leaves each location.

A restricted government campus prohibits external connectivity. An air-gapped on-premises deployment is the starting requirement. Capacity, internal redundancy, offline updates, and local operational expertise must be funded accordingly. Cloud-dependent features are outside the acceptable design, regardless of their convenience.

Validate the business decision under operating conditions

Before selecting an architecture, agree on measurable requirements for alert latency, outage autonomy, data movement, recovery time, and ongoing ownership. Compare only designs that meet those requirements, then test the preferred configuration at representative sites.

Operational testing has particular value in physical security. The U.S. Army’s account of Scylla testing at Blue Grass Army Depot describes an eight-month evaluation. That example supports sustained evaluation in context; it does not establish that one deployment architecture suits every facility.

My recommendation is to make architecture an explicit part of the security acceptance process. Demonstrate what the system does when links fail, workloads increase, or equipment needs replacement. The right deployment is the one the organization can govern, afford, and keep working when its security team needs it.