Defense and military networks
Classified and restricted networks operate disconnected from public internet by design. Operations tools must run inside the perimeter or they are out before the procurement conversation starts.
Air-Gapped Autonomous AIOps
Deployment
True air-gapped AIOps. The platform and the model both run inside your perimeter. No telemetry leaves the environment. Built for defense, government, regulated banking, and sovereign cloud.
What true air-gapped means
The term "air-gapped" has been stretched to the point of being meaningless in vendor marketing. Some vendors call a single-tenant cloud deployment "air-gapped." Others call an on-premises platform "air-gapped" while the model still runs in their cloud. Both are misleading.
True air-gapped means the entire platform runs inside the customer perimeter with no network egress to vendor infrastructure. The AI model itself runs on customer hardware. Telemetry never leaves the environment. Inference happens locally. The platform can operate indefinitely without an internet connection. Model updates arrive as verified offline artifact packages that the customer's security team inspects and stages through their normal change management process.
That is the bar required by defense, government, and regulated banking customers. Anything less is single-tenant SaaS with a marketing label.
Why most AIOps platforms cannot offer this
The economic and architectural reality of cloud-native AI platforms is that the model is the expensive part and the vendor owns it. Selling a model that the customer runs themselves means the vendor cannot continuously update it, cannot observe how it behaves in customer environments, and cannot benefit from cross-customer pattern learning. Most vendors have decided those tradeoffs are not worth the additional engineering investment, so their air-gapped story stops at "the platform runs on-prem but it still phones home to our model API."
That is fine for SaaS-comfortable customers. It is disqualifying for customers whose security model treats vendor cloud as an unacceptable data path.
Who needs true air-gapped AIOps
Sovereign cloud and government
Public sector deployments under sovereign cloud requirements (EU sovereign cloud, GCC sovereign cloud, country-specific data residency rules). Telemetry leaving the sovereign perimeter is itself a compliance violation.
Regulated banking under data residency rules
Indian banks operating under RBI Tier-2 and Tier-3 controls. European banks under specific country-level residency. Banks that have classified their operational telemetry as customer-sensitive and refuse to expose it to vendor cloud.
Healthcare networks under strict residency
Hospital networks and national health systems that treat operational telemetry as patient-data adjacent. Air-gapped deployment removes the data-residency review from the procurement path.
Critical infrastructure operators
Energy, telecom, water, transport. Operators whose infrastructure is itself a national security concern. Exposing operational telemetry to a third-party cloud is treated as a meaningful attack surface, not just a compliance question.
Customers with a zero-egress security policy
Any enterprise whose security architecture treats outbound connectivity from production systems as an unacceptable risk. Air-gapped deployment is the only way the operational AI platform can land in those environments.
Air-gapped vs SaaS vs single-tenant cloud
| Capability | Vendor SaaS | Single-tenant cloud | True air-gapped |
|---|---|---|---|
| Telemetry leaves customer perimeter | Yes | Yes | No |
| Model runs in customer perimeter | No | No | Yes |
| Internet connectivity required | Yes, continuously | Yes, continuously | No |
| Model update cadence | Continuous | Continuous | Batched, offline |
| Cross-customer pattern learning | Yes | Partial | No |
| Acceptable for defense / sovereign | No | No | Yes |
| Time to production | Fastest | Fast | Slowest (security review) |
Each model is correct for its target customer. The honest framing is not "air-gapped is better than SaaS." It is "the deployment model has to match the customer security architecture, and air-gapped is the only valid choice for some customers."
How Opstral implements air-gapped
The air-gapped deployment of Opstral ships as a containerized package that runs entirely on customer infrastructure. Five components run inside the perimeter.
- Sentinel AI (the reasoning agent), including the local inference runtime.
- ProcBot (the execution agent), with the connector library bundled.
- Sherlock (the validation agent), including the MOP scoring engine.
- The MOP library, including pre-built procedures and the customer's custom MOPs.
- The audit store, which holds Action Tickets locally.
No component requires outbound connectivity to operate. The platform is installable from offline artifact packages, runs from offline artifact packages, and is updated via offline artifact packages.
Updates flow through the customer's existing secure transfer process: the offline artifact package is signed by Opstral, the customer's security team validates the signature, the package is staged in a non-prod environment, then promoted to production under the customer's change management. Update cadence is typically monthly for security and capability updates, quarterly for major releases.
What you give up, honestly
Three real tradeoffs that should be visible in the procurement conversation.
- Continuous model updates are not possible. You receive batched updates on your update cadence, not the latest model version the day it ships. For most operational workloads this does not matter. For workloads where the model is heavily tuned to a specific incident pattern, the delay can be felt.
- Cross-customer pattern learning is not possible. Your environment learns from your own incidents, not from the aggregate across all customers. Over time your MOP library compensates, but the first six months of an air-gapped deployment lean more heavily on your own historical data than a SaaS deployment would.
- Vendor support has less visibility into platform behavior. When you raise a support ticket, the relevant logs and traces have to be exported by your team, sanitized, and shared. Some issues that take an hour to resolve in SaaS take a day in air-gapped because of the diagnostic round trip.
None of these is a blocker for the customers who actually need air-gapped. They are the cost of the deployment model, and they are worth paying for the right risk profile.
Related deployment options
Air-gapped is the strictest of a family of deployment models. The full menu, for reference.
| Option | What it means |
|---|---|
| SaaS (managed) | Opstral hosts and operates the platform. Fastest path to value. |
| Private cloud (customer-owned) | Customer owns the cloud account; we deploy and manage. Good middle ground for customers who want isolation without on-prem operational burden. |
| On-premises | Runs entirely in customer data center. Telemetry stays in the perimeter. Model updates fetched over internet from a controlled egress point. |
| Air-gapped | On-prem with offline model update workflow. Zero outbound connectivity required. Model inside perimeter. |
| Regional residency | SaaS deployment regions in US, EU, APAC for data sovereignty without going all the way to on-prem. Available across the above models. |
Frequently asked questions
Yes. For defense and regulated banking customers we share platform source review under NDA as part of the procurement process. Air-gapped deployment includes signed container images that customers can inspect before staging.
We would love to understand your problem first. Tell us about your environment, the operational constraints you are working within, and the outcomes you need. Once we have a solution that fits, the licensing conversation follows. The fastest path is a working session with our team. Get in touch.
Sentinel AI inference is the largest component. A typical mid-size enterprise deployment runs comfortably on a Kubernetes cluster with GPU nodes available for inference workloads. Exact sizing depends on signal volume and MOP execution rate. We size during scoping.
Yes, with the appropriate customer-side approvals. We have walked customers through deployment in classified and restricted networks. The technical deployment is straightforward; the long pole is the customer's own security review and approval cycle.
Discovery and scoping in a non-production environment first. We deploy a representative subset of the platform in the customer's POC environment (typically a smaller cluster), validate the autonomous capabilities against agreed success criteria, then plan the production deployment under the customer's change management process. Realistic timeline from contract to production-grade autonomous operation is 12 to 24 weeks for air-gapped, driven by the security review on the customer side.
For the broader category context, see AIOps platform comparison. For a deeper read on autonomous operations in restricted environments, see Autonomous Operations in Air-Gapped Environments.