Private AI cloud: plan the operating model.
A private deployment is not a hosting choice. It is a decision about who holds the keys, who can administer the system, and what happens to your workloads if the relationship ends.
What private can mean
The word covers four materially different arrangements. Being specific about which one you need changes the cost, the timeline and the responsibilities.
Dedicated hosted
A dedicated hosted option. Capacity, operating location, isolation and support responsibilities need to be confirmed in the deployment agreement.
Your cloud tenancy
Deployment in your cloud account. Review outbound integrations, telemetry and support access against your existing agreements and network controls.
On-premises
Your datacenter and your hardware. Used where residency or latency requirements are absolute, or where an existing investment should be used.
Disconnected
A target configuration with local inference and a controlled update process. Verify that every required dependency works within the agreed network boundary.
Controls to specify and verify
- Model independence
Open-weight and private models hosted inside the boundary, with no requirement to call an external provider.
- Workload isolation
Separation between tenants, environments and agents at the compute and network layer.
- Key custody
Agree key ownership, storage, rotation and recovery; verify which components support customer-held keys.
- Operator access control
Define who may administer each component, the approval required and the evidence retained for review.
- Sovereign operations
Agree runbooks and test whether your operating team can upgrade, recover and maintain the selected configuration.
- Portability
Specify which data, policies, evidence and agent definitions must be exportable, and test the required formats.
- Standard Kubernetes
Conformant Kubernetes rather than a proprietary control plane, so the deployment is not a lock-in mechanism.
- Residency by construction
Document permitted regions, model routes and integrations, then verify actual placement and outbound traffic.
- Evidence stays local
Choose the storage, retention and export rules for decision records, logs and backups within the agreed boundary.
The honest trade-off
A disconnected deployment removes an entire class of risk and adds an operational burden: you own patching, capacity and recovery. That is the right trade for some environments and the wrong one for many. It should be a decision, not a default.
Who does what
| Responsibility | Dedicated hosted | Your cloud | On-premises / disconnected |
|---|---|---|---|
| Hardware and capacity | NeuroCluster | Your cloud provider | Customer |
| Platform operation | NeuroCluster | Shared | Customer, with support |
| Upgrades | NeuroCluster | Scheduled with customer | Customer-initiated |
| Key custody | Customer option | Customer | Customer |
| Monitoring | NeuroCluster | Shared | Customer |
| Incident response | NeuroCluster, with customer | Shared | Customer, with support |
Questions
- Can we start shared and move to private later?
- A move can be planned, but the work depends on data, identity, integrations and operating responsibilities. Agree the export and migration requirements before choosing a starting configuration.
- Do models work as well without a frontier API?
- For a great deal of enterprise work, yes — particularly extraction, classification, retrieval and structured reasoning over a well-modelled ontology. For some tasks a frontier model is meaningfully better. The architecture lets you make that call per class of work rather than once for the whole system.
- What is the minimum viable footprint?
- It depends on the workload mix, model sizes and availability requirements. This is a sizing conversation with real numbers rather than a published figure, and we would rather scope it properly than publish one that misleads.
- Available
- Existing implementation. Scope, readiness and operating controls must be validated for each deployment.
- In development
- Being built now. Not yet generally deployable.
- Planned
- On the roadmap. Not yet in development.
- Research
- Under investigation in Labs. May not become a product.
- Strategic direction
- Intent, not a delivery commitment.
Continue
- Infrastructure overviewThe full stack and every deployment posture.
- SovereigntyThe six control dimensions a private deployment addresses.
- ModelsRunning open-weight and private models inside the boundary.
- RuntimeThe execution layer being deployed.
- ArchitectureReference architecture for security review.
- GovernmentWhere disconnected deployment is most often required.
Bring us one operational problem.
You do not need a finished brief. Bring the problem — we will work out the next step together.
Or book a call with the team