Air-gappedLLM inferenceFine-tuningRAG

Secure Regulated-Industry AI Platform

A platform where compliance and isolation are first-class constraints: no internet egress, signed and scanned supply chain, strong tenant and data isolation, and audit evidence for everything.

Customize in Designer
Business objective

Run AI workloads on sensitive/regulated data with demonstrable compliance and data sovereignty.

Personas
Security engineerCompliance officerPlatform administratorSRE
Key assumptions
  • ·No outbound internet; all artifacts arrive through a controlled inbound path.
  • ·Controls must map to frameworks (NIST, ISO 27001, SOC 2, sector-specific).
  • ·Strong tenant isolation, including GPU memory and data remanence handling.
Logical architecture

Rendering diagram…

Design

Compute

Dedicated node groups or MIG-partitioned isolation per tenant; scrub GPU memory between tenants to address data remanence.

Network

No egress. Segmented internal networks, default-deny network policies, and an inbound mirror/scan path as the only way artifacts enter.

Storage

Encrypted at rest with managed keys; per-tenant volumes; strict lifecycle and secure deletion for sensitive data.

Kubernetes

Hardened Kubernetes with an admission policy engine (Kyverno/Gatekeeper), signed-image enforcement, pod security, and runtime detection.

Security

Supply chain: mirror → scan → sign (Cosign/SLSA) → verify at admission. Model/dataset provenance tracked. Controls mapped to frameworks with collected evidence.

Observability

Full audit logging, SIEM integration, runtime security events, and compliance drift detection against a known-good baseline.

High availability
In-cluster redundancy for control plane and critical services; no dependency on external SaaS that the air gap would break.
Disaster recovery
Encrypted backups restored within the boundary; documented, tested restore procedures that never require internet access.
Capacity
Sized conservatively with isolation overhead (dedicated pools reduce packing). Plan spare capacity since elastic cloud burst isn't available.
Cost profile
CapEx-heavy with a compliance premium: isolation reduces utilization, and audit/evidence tooling is part of the run cost.
Risks
  • highSupply-chain gap — an unscanned artifact entering the boundary.
  • highTenant isolation failure exposing regulated data across tenants.
  • mediumConfig drift from the compliant baseline going undetected.
Implementation phases
  1. 1

    Boundary + supply chain

    Mirror/scan/sign pipeline and internal registry with admission verification.

  2. 2

    Hardened platform

    Policy engine, pod security, secrets/KMS, network segmentation.

  3. 3

    Isolation

    Per-tenant compute/data isolation and GPU memory scrubbing.

  4. 4

    Governance

    Audit evidence, SIEM, drift detection, control-to-framework mapping.

Validation checklist
  • No egress path exists from workload networks.
  • Unsigned/unscanned images are rejected at admission.
  • Cross-tenant data access blocked under test.
  • Audit evidence collected and mapped to required controls.
Recommended GPUs for this pattern
NVIDIA

H100 SXM5

80 GB · 700 W

NVIDIA

A100 80GB SXM

80 GB · 400 W

AMD

Instinct MI300X

192 GB · 750 W

NVIDIA

L40S

48 GB · 350 W