Sovereign Inference: An Architecture for Offline, Owner-Operated Foundation Models in Regulated Enterprises

Capable foundation models are usually delivered as a remote service, so that using them on sensitive material requires that material to leave the organisation's control. For regulated enterprises this is often a barrier to adoption in itself, independent of model quality. This paper describes an architecture for sovereign inference: running a capable, general-purpose model entirely offline, on hardware the organisation owns, with the network egress boundary enforced as an engineering invariant rather than a policy. We set out four requirements, locality, confidentiality, provenance and ownership, and describe the components that meet them: a loopback-only inference engine, an egress boundary tested as a release gate, retrieval-grounded private knowledge bases that keep large corpora usable within a bounded model context, task-specific working surfaces sharing one engine, and an offline-verifiable attestation record for provenance. We discuss the operational model, in which capability scales with owned hardware and there is no usage metering, and a model strategy that is agnostic to any particular foundation model while supporting the later training of operator-specific models on proprietary data. The design trades the elasticity of the cloud for custody, provability and predictable cost, a trade many regulated buyers prefer.