Financial services

Encrypted inference on customer records

Credit, fraud, or underwriting models must score features the institution cannot send to a third-party GPU cloud. Model refresh cycles are weekly; lattice FHE cannot keep pace. Regulators and internal audit ask where plaintext exists: today, the answer is everywhere.

How it runs

Customer feature tensors remain encrypted in the bank's key domain. Models run against ciphertext on Umbra U100 or N8 nodes co-located with the core banking tier. Prompts, activations, and scores stay encrypted until deliberate decryption inside the institution's HSM boundary.

What changes

Production ML without exporting identifiable rows to an external plaintext training or inference host. PCI and privacy assessments shrink because the compute tier never held customer records in the clear.

Cross-institution analytics without data sharing

Several institutions need joint AML, fraud-ring, or stress-test models without exchanging customer files. Legal agreements take months; a shared database is politically and regulatorily unacceptable.

How it runs

Each participant contributes encrypted datasets under contractual key governance. Joint models train and score in encrypted federated space on N8 nodes hosted at a neutral facility or rotated across participants. No party receives another's plaintext holdings; gradients and aggregates remain ciphertext throughout.

What changes

Sector-level detection quality without a shared customer database. Institutions gain network-effect model quality while legal and compliance teams retain a defensible no-pooling architecture.

Privacy-preserving supervisory reporting

A supervisor needs concentration, exposure, or systemic-risk aggregates that today force over-exposure of customer detail in regulatory extracts. Institutions resist granular submission; supervisors lack timely visibility.

How it runs

Encrypted aggregation produces capital, concentration, and counterparty statistics on Umbra. Supervisors receive ciphertext aggregates released through agreed key ceremony: not raw customer rows into a general-purpose analytics environment.

What changes

Regulatory visibility with a smaller plaintext scope for both institution and supervisor. Reporting cycles accelerate because aggregation runs continuously on encrypted holdings rather than on batch plaintext extracts.

Trading and market-data analytics under confidentiality

Desks and infrastructure providers need joint analytics on order flow and position data bound by Chinese walls and client confidentiality. Plaintext consolidation violates both; slow legal review blocks time-sensitive risk models.

How it runs

Counterparties contribute encrypted order and position features. Risk and surveillance models run on Umbra nodes inside each firm's trust boundary or at a jointly governed neutral node. Outputs are encrypted risk metrics, not raw tape.

What changes

Joint risk visibility without a consolidated plaintext tape. Compliance retains wall integrity because the compute path never required decryption at a shared analytics tier.