White Paper

Data Sovereignty by Design: Architecture for Air-Gapped AI

Building AI systems that operate with zero network dependencies

October 2025
5 min read

Executive Summary

Data sovereignty — the principle that data is subject to the laws where it is stored — has moved from a compliance checkbox to a core architectural requirement. As AI processes more sensitive data, where it runs has become inseparable from how it is built. This paper examines the architecture of AI systems designed for air-gapped and sovereignty-compliant deployment.


1. Why Architecture Beats Policy

Data sovereignty was once addressed through legal agreements. The 2020 Schrems II ruling proved that legal frameworks alone are insufficient — architecture must enforce what policy cannot guarantee.

The Sovereignty Enforcement Hierarchy
Strongest Weakest
1

Air-Gapped / Local Deployment

Data physically cannot leave premises

2

Sovereign Cloud Region

Data stays in jurisdiction, on third-party infra

3

Contractual Guarantees

Legal agreement, enforcement varies

4

Terms of Service / Privacy Policy

Self-declared, limited recourse

Air-gapped deployment is the strongest form because it operates on a simple principle: data that never traverses a network cannot be intercepted, redirected, or subject to foreign jurisdiction.


2. Monolithic vs. Modular AI

Rather than deploying massive cloud-scale models, air-gapped environments are adopting modular expert AI — purpose-built small models for specific tasks.

Monolithic LLM vs. Modular Expert AI
Monolithic LLM (70B+)
  • × Massive compute needs
  • × Black-box reasoning
  • × Single point of failure
Hardware48+ GB VRAM
Latency10–30s / response
Modular Expert AI
  • Runs on workstation
  • Auditable decision paths
  • Composable, replaceable
Hardware8–16 GB VRAM
Latency1–5s / response
Source: Adapted from US.inc, "Deploying Expert AI in Air-Gapped Environments," 2025

3. Hardware Requirements Have Dropped

The hardware needed for local AI has decreased dramatically:

ScenarioRAMGPUModelsSpeed
Tactical Edge16 GBNone (CPU)Phi-4 Mini 3.8B15–25 tok/s
Analyst Workstation32 GB12 GB VRAMQwen 3 4B–7B40–60 tok/s
Departmental Server64 GB24 GB VRAMMultiple 7B models80–120 tok/s
Enterprise Cluster128 GB48+ GB VRAM13B+ ensemble200+ tok/s

A standard analyst workstation with a $1,600 GPU can now run competitive AI models at 40–60 tokens per second — fast enough for interactive use with no network dependency.


The Path Forward

Most regulated organizations should target local deployment for sensitive operations and air-gapped deployment for classified data. The technology is available today at costs declining 30% annually.

Organizations that architect for sovereignty now will be compliant by default when the next wave of regulation arrives. Those that retrofit later will pay the premium.


References

  1. DLA Piper. "GDPR Fines and Data Breach Survey: January 2026."
  2. Katonic AI. "Air-Gapped AI: Deploying Agents Without Internet Access." 2025.
  3. US.inc. "Deploying Expert AI in Air-Gapped Environments." 2025.
  4. NIST. "SP 1800-35: Implementing a Zero Trust Architecture." June 2025.
  5. IBM Security / Ponemon Institute. "Cost of a Data Breach Report 2025."
  6. Local AI Master. "Small Language Models 2026."

Independent industry analysis. All data cited from publicly available sources. Published October 2025.