Physical sovereignty

One rack.
One tenant.
Zero shared cloud.

For institutions that refuse to put inference on someone else’s metal, a Corvux appliance you own, in your data centre, with a defined air gap.

Managed sovereign AI is the right wedge for many teams. Tier-1 banks, defence-adjacent agencies, and some SAGAs will still ask: “Whose rack is it, and can anything leave the building?” The air-gapped appliance is Corvux’s answer, a dedicated SKU, outside a repackaged cloud tenant.

1:1
Tenant isolation
Offline
Update path
HSM
Key option
Nairobi
Install team

Inference, storage, control, bundled.

Compute

GPU nodes sized to your concurrent user and batch OCR targets, quoted after workload profiling.

Vector + object

Embeddings and document stores on local NVMe; backup hooks to your DR site only.

Control plane

Admin, routing, audit export, no dependency on Corvux SaaS for runtime.

Evidence kit

Acceptance tests and residency narrative for ODPC and internal audit, see evidence pack.

┌ air-gapped zone ────────────────────────┐

│ apps ──▶ api.corvux.local ──▶ GPUs │

│ ▲ │

│ └── admin (bastion only) │

│ ✗ no outbound LLM API │

│ ✗ no shared Corvux multi-tenant cloud │

└─────────────────────────────────────────┘

Compare with on-premise infrastructure DIY builds and managed sovereign AI if CapEx timing differs. Appliance buyers usually already failed a “use Azure OpenAI” conversation with risk.

Facilities, security, procurement.

Is this a GPU server?

It is a full stack: models, vector store, routing, admin UI, update bundles, delivered as a SKU with acceptance tests.

How do updates work air-gapped?

Signed offline bundles via your change window. No phone-home telemetry to foreign clouds.

Power and cooling in Nairobi DCs?

We size for typical Kenyan colo constraints and provide heat/load sheets for facilities sign-off.

Can it still talk to internal apps?

Yes, over your VLAN. The air gap is outbound to the public internet. Core banking adapters stay on the inside.

Scope an appliance SKU

Share concurrent users, doc volumes, and your data-centre constraints. We return a bill of materials, free of a generic GPU quote.

Request appliance sizing