Sovereign AI

National inference with encrypted prompts and outputs

Citizen-facing or civil-service AI must answer without sending plaintext prompts to a foreign or commercial host. Data-protection authorities ask where inference runs; the honest answer today is often 'abroad, decrypted.'

How it runs

APIs accept encrypted inputs at ministry gateways. Model execution on Umbra N8 nodes stays on ciphertext; decryption occurs only inside ministry key domains under statute. Prompts, context, and draft outputs never traverse foreign infrastructure in the clear.

What changes

Domestic AI services that can demonstrate data never resolved during processing. Sovereignty claims rest on cryptography, not on vendor contractual promises alone.

Citizen-data model training

A national model needs population-scale features without a shared plaintext training lake. Privacy impact assessments reject centralised copies; federated plaintext approaches still expose shards at the GPU.

How it runs

Feature engineering and gradient aggregation run on encrypted shards distributed across ministry data centres. Training scales on N8 nodes; identifiable records are never staged in the clear for the training cluster.

What changes

Model quality without collapsing privacy and sovereignty requirements. Programmes advance national AI capability while PIA reviewers see ciphertext-only processing at the compute tier.

Cross-ministry analytics

Health, treasury, and interior need joint policy models: fraud, economic planning, benefits integrity: without building a super-database of raw citizen files. Legal mandates restrict purpose and retention; plaintext merges violate both.

How it runs

Ministries contribute encrypted holdings under purpose-limited key governance. Policy models correlate in encrypted space on shared Umbra nodes; outputs are encrypted statistics released only to authorised analytical units.

What changes

Whole-of-government insight that survives privacy impact assessment. Ministries collaborate on models without surrendering raw holdings to a central plaintext warehouse.

Secure vendor and research collaboration

Governments engage academic and commercial partners on sensitive datasets under strict data-sharing agreements. Plaintext export to partner GPUs is prohibited; synthetic data alone insufficient for model quality.

How it runs

Partners run agreed models against encrypted citizen or operational datasets inside government-accredited Umbra nodes. Partners receive encrypted model outputs and metrics: never plaintext records: under contractually defined key release.

What changes

Research and vendor partnerships proceed without data export. Programmes retain custody while partners contribute model architecture and training expertise inside cryptographic boundaries.