Model choice
Open-source, NVIDIA NeMo and Nemotron-class, or small right-sized models. Swap models without re-architecting.
SOVEREIGN INDUSTRIAL AI · REFERENCE ARCHITECTURE
The reference architecture for sovereign industrial AI.
The full stack that runs inside your plant, air-gapped, from OT data to governed action. Hardware-agnostic, model-agnostic, and entirely under your control.
Five layers, one boundary.
From the silicon to the decision, every layer runs inside your fence line. Nothing depends on an outside service.
05
Decision
MAGSCognitive decision teams that detect, decide, coordinate, and execute the next best action, with humans in governance.
04
Model engine
AI FlowThe industrial-grade agentic harness. Runs and governs the models you choose, with guardrails and provenance.
03
Context
Digital TwinsThe operational model of the asset: physics, constraints, and topology, so agents reason about the real plant.
02
Data
Data StreamsReal-time OT/IT ingest from SCADA, historians, sensors, and control systems, on site.
01
Infrastructure
Your hardwareCustomer-controlled compute: NVIDIA-accelerated, AMD, or your own silicon, from the edge to an industrial data centre.
Runs on the accelerators you already have — NVIDIA-accelerated, AMD, or your own silicon — from the edge to an industrial data centre.
Run and govern any model, on your own hardware.
AI Flow is the industrial-grade harness that puts models to work in operations, without sending anything out.
No data egress
Prompts, context, and operational data never leave the boundary. Inference happens where the data lives.
Governance & guardrails
Stage gates, policies, and the Human Agency Scale keep autonomy inside defined limits.
Provenance & audit
Every recommendation and action is traceable, with an immutable audit trail for regulated operations.
Right-model routing
Route each decision to the model that fits it, balancing capability, latency, and cost on your own hardware.
Runs offline
Full autonomy continues when the site is disconnected or permanently air-gapped.
Sized to the operation, from one asset to the enterprise.
The same architecture scales down to an edge node and up to an air-gapped data centre. Start where the value is.
At the asset
- Footprint
- A single node on the line or at the asset
- Hardware
- Edge accelerators (e.g. AMD Ryzen™ AI) or CPU
- Models
- Small, right-sized models
- Best for
- Low-latency local inference, intermittent connectivity
Across the site
- Footprint
- Site servers running the full loop for a facility
- Hardware
- On-prem servers, optional GPUs
- Models
- Right-sized to open, mid-size models
- Best for
- Site-wide agentic operations, central governance, local execution
Across the operation
- Footprint
- A GPU cluster for larger models and multi-site operations
- Hardware
- NVIDIA-accelerated, AMD, or your own silicon
- Models
- Open and NeMo-class models
- Best for
- Fully air-gapped, multi-site, larger reasoning workloads
Configurations are confirmed with a solution architect against your connectivity, latency, and governance requirements.
Air-gapped by design.
Security isn’t a setting bolted on afterwards. The architecture assumes no outside access from the start.
No external access
Nothing calls out. There is no path from the control network to the public internet, so there is nothing to breach from outside.
Data stays on site
Operational data, IP, and models remain inside the plant boundary. Sovereignty is physical, not policy.
OT-safe integration
Reads and writes to control systems happen through governed, auditable interfaces, not ad-hoc connections.
Governed autonomy
Agents act only within defined boundaries, with human oversight calibrated on the Human Agency Scale.
Full auditability
Every decision carries its provenance. The immutable trail supports authority-to-operate and regulated missions.
From your infrastructure to governed action.
1
Install on your infrastructure
Deploy onto customer-controlled compute, from an edge node to an air-gapped data centre.
2
Connect OT data
Data Streams ingest live signal from SCADA, historians, sensors, and control systems, on site.
3
Model the operation
Digital Twins capture the physics, constraints, and context of the real asset.
4
Bring your models
AI Flow runs and governs the models you choose, entirely on-premise.
5
Deploy decision teams
MAGS agents detect, decide, and coordinate the next best action.
6
Govern every action
Execution stays within boundaries, with provenance and an immutable audit trail.
Design your sovereign deployment.
Walk through the reference architecture with a solution architect and map it to your plant, your hardware, and your models.
