Architecture Overview
The RGS Carbide Enclave is a fully airgapped, government-grade Kubernetes platform built from commodity hardware and open-source components. It is sovereign in the strict sense: once the airgap boundary is crossed, nothing enters the enclave that wasn't explicitly collected, verified, and transported through Hauler. There is no phoning home, no pull-from-internet, no surprise dependency on an upstream registry.
What makes this different from a standard on-premises Kubernetes deployment:
- Zero internet access on the far side of the airgap. DNS resolves only internal names. There is no default route to the internet from any node inside the enclave.
- Every artifact is auditable. Hauler manifests enumerate every image, chart, and binary before it enters. If it isn't in the manifest, it doesn't exist in the enclave.
- AI workloads run on a physically separate arm64 node (NVIDIA DGX Spark) joined to the same management cluster. arm64 parity is a hard requirement for every artifact that touches it.
- Internal PKI everywhere. step-ca is the root. cert-manager issues all certificates. No self-signed certs, no
--tls-skip-verifyin production configs.
Components
| Component | Role | Why This Choice |
|---|---|---|
| Harvester | Bare-metal HCI hypervisor — runs all platform VMs | Purpose-built for edge/airgap; tight Rancher integration |
| RKE2 | Kubernetes management cluster (3-node HA) | FIPS-capable, hardened by default, airgap install supported natively |
| Rancher Manager | Cluster lifecycle, RBAC, and multi-cluster visibility | Central control plane for all downstream clusters |
| Harbor | OCI container registry + Helm chart proxy | Self-hostable, OIDC auth, supports OCI artifacts (models) |
| Keycloak | OIDC identity provider | Single IdP for Rancher, Harbor, and DGX workloads |
| cert-manager + step-ca | Internal PKI / TLS certificate automation | ACME support, cert-manager integration, fully airgap-capable |
| Hauler | Airgap artifact transport and bootstrap registry | RGS-native; images + charts + binaries in one store |
| NVIDIA DGX Spark | arm64 AI inference node | NVIDIA GB10 — purpose-built for edge AI; 128 GB unified memory |
| ExMachina | AI model gateway and serving orchestration | Ties vLLM/Ollama into a single access layer |
Hardware
| Host | Role | Model | CPU | RAM | NICs |
|---|---|---|---|---|---|
| nuc-00 | Bastion / admin | NUC10i7FNK | i7-10710U | 32 GB | 1 |
| nuc-01 | Harvester node 1 | NUC10i7FNH | i7-10710U | 64 GB | 2 |
| nuc-02 | Harvester node 2 | NUC10i7FNH | i7-10710U | 64 GB | 2 |
| nuc-03 | Harvester node 3 | NUC10i7FNH | i7-10710U | 64 GB | 2 |
| spark | AI inference | NVIDIA DGX Spark (GB10) | arm64 | 128 GB | 1 |
| nas | NAS / NFS storage | ASUS X99 | Xeon E5-2630 v3 | 94 GB | 1 |
nuc-01, nuc-02, and nuc-03 each have two NICs: one for Harvester management traffic and one for VM network traffic. The second NIC is required — Harvester uses it for the VM VLAN overlay.
Network
| Item | Value |
|---|---|
| Domain | carbide-enclave.kubernerdes.com |
| Subnet | 10.0.0.0/22 |
| Gateway | 10.0.0.1 |
| Bastion (nuc-00) | 10.0.0.10 |
| NAS | 10.0.0.11 |
| Rancher VIP | 10.0.0.30 |
| Keycloak VIP | 10.0.0.98 |
| Harbor VIP | 10.0.0.99 |
| Harvester VIP | 10.0.0.100 |
| DGX Spark | 10.0.0.251 |
| DHCP pool | 10.0.0.172 – 10.0.0.254 |
Build Sequence
The enclave is built in a strict dependency order. Nothing can be skipped — each layer depends on the one below it.
Day 0 / 1 / 2 Framework
| Phase | What happens | Key output |
|---|---|---|
| Day 0 | Design: hardware, network, airgap strategy, credentials | Everything is planned before a single cable is plugged in |
| Day 1 | Build: install the full stack in dependency order | A running, secured, airgapped Kubernetes platform with AI serving |
| Day 2 | Operate: updates, monitoring, compliance, break-fix | Sustainable operations without internet dependency |
Ready to start? Getting Started walks through what you'll build and how to orient yourself in the docs.