Defence

Coalition analytics without declassification

Two or more nations need joint correlation on holdings that cannot leave national key domains. Release authorities govern every output; source material must never resolve on a shared fabric in the clear. Prior programmes stalled on declassification timelines measured in months.

How it runs

Each partner contributes ciphertext under national keys. Encrypted search, graph correlation, and pattern joins run on Umbra N8 nodes inside accredited facility power and airflow. Queries and intermediate tensors stay encrypted; only policy-governed aggregates exit under bilateral release rules.

What changes

Mission analytics proceed without a prior declassification campaign or a shared plaintext lake. Coalition staff work from encrypted correlation results that survive releasability review because the underlying collections never opened.

Encrypted signals processing

RF, COMINT, or sensor streams must be inspected at line rate without staging plaintext on analyst workstations or vendor analytics tiers. Throughput requirements exceed what lattice FHE prototypes can sustain; accreditation rejects any pipeline that decrypts at the accelerator.

How it runs

Feature extraction, dwell detection, and cross-channel matching run on Umbra against encrypted streams ingested at the collection boundary. A single U100 characterises the workload; production scales to N4 or N8 nodes sized to the mission envelope: 160M to 320M encrypted ops/s estimated at the node class from published specs.

What changes

Line-of-mission inspection stays inside cryptographic and accreditation controls already familiar to the programme. Analysts receive actionable indicators without a plaintext staging tier between collection and compute.

Multi-level secure inference

Classification-aware models must answer prompts that span levels the host environment cannot hold in the clear. Commercial AI stacks assume a single trust domain at the GPU; MLS policy blocks them from production networks.

How it runs

Prompts, context windows, and model weights traverse the stack as ciphertext. Inference executes on Umbra; results are labelled and routed by policy without a trusted enclave holding every level as plaintext simultaneously.

What changes

Inference becomes available where MLS policy previously blocked commercial AI stacks. Programmes gain model-assisted analysis without collapsing compartment boundaries at the compute tier.

Cross-domain fusion without a plaintext merge

All-source fusion requires correlates from HUMINT, GEOINT, and SIGINT programmes that cannot be co-located in one database. Manual fusion does not scale; automated fusion on plaintext violates every compartment rule.

How it runs

Programmes export encrypted feature vectors and entity graphs to a fusion node. Correlation models run in ciphertext; only matched entity handles and confidence scores release under dual-key governance.

What changes

Fusion timelines compress from analyst-weeks to machine-hours without a super-database of raw holdings. Accreditation reviewers see a smaller plaintext scope because fusion never required decryption at the compute boundary.