Protection for data in the one state that was never covered
Encryption covers data at rest and in transit. The gap regulators and security teams are now focused on is the third state: data while it's actually being used — loaded into memory, being processed, being read by a model. Luminous is built to close that gap.
The third state
For decades, data protection meant two things — encrypt it on disk, encrypt it in the network. Both are mature, well-understood, and widely deployed. The state that was left open is the moment data is actually being worked on: unencrypted in memory, visible to whatever process is running on the machine, including the cloud provider's own infrastructure and, in an AI system, the model itself. That gap is what confidential computing — hardware-backed memory encryption plus remote attestation — closes, and it's what most AI deployments today don't use.
Not gated on the hardware
Luminous doesn't require confidential-compute hardware to run — the governance and enforcement engine works on standard infrastructure. Confidential compute is a deployment option layered on top, selected per environment, not a prerequisite for the platform to function. That matters practically: a pilot can start on infrastructure the customer already has, and move up the assurance ladder later without a re-architecture.
Three tiers, one engine
| Tier | What it means | When it fits |
|---|---|---|
| Assertion | Standard compute; policy and controls enforced in software | Pilots, low-sensitivity workloads, fast time-to-value |
| Attestation | Confidential-compute hardware; memory encrypted, attestation verifiable by the customer | Regulated production workloads — PHI, financial data, controlled information |
| Sovereign | Attestation plus the infrastructure provider is removed from the trust base entirely | Environments where even the cloud operator must be an untrusted party |
The tier is a deployment choice, not a different product. The same enforcement logic runs at every tier; what changes is how strongly the underlying infrastructure can prove it did what it says it did.
Cross Fabric — beyond one node
Confidential computing today is mostly a single-node guarantee — the attestation and encrypted-memory protections that make a workload verifiable stop at the edge of one machine. Most production AI workloads don't fit on one machine. Luminous's Cross Fabric is built to extend confidential-compute guarantees across a cluster of nodes, so a deployment isn't capped at what confidential compute can do on a single box. It's the piece that makes the Attested and Sovereign tiers usable at real production scale, not just in a single-node demo.
Built accelerator-agnostic
The platform's approach to confidential computing is described and implemented at a general level — it isn't wired to one chip vendor's specific confidential-computing implementation. That's a deliberate choice: the confidential-compute hardware landscape is still moving, and a customer's infrastructure choice today shouldn't lock Luminous out tomorrow.
Why this is the control regulators are asking about
Financial-services and healthcare regulators increasingly ask not just "is the data encrypted" but "who can see it while it's in use — including your own vendor." A platform that can answer "no one, including us, given attestation" is answering a question most AI deployments today can't. The framework detail sits on Standards & Validation.