Why does confidential AI need a common language for trust?

Confidential computing used to be an internal concern. Operators ran attestation inside their data centers to decide whether a server could join a cluster. The trust boundary stopped at the rack.

AI changed that boundary. A single workload now spans CPUs and GPUs, accelerators, firmware, and services from different suppliers. Model providers want proof their weights run only in protected environments. Users want proof their data stays encrypted during inference. Infrastructure operators still need to verify the platform. Each party needs evidence, but the evidence lives in different formats behind different interfaces.

Remote attestation lets a system produce evidence about its identity and state. A verifier evaluates that evidence against authoritative information — endorsements that describe hardware capabilities, reference values that describe expected firmware and software measurements.

The integration tax nobody budgeted for

In a homogeneous stack, one supplier provides both. In a heterogeneous AI system, the CPU vendor publishes artifacts. The GPU vendor publishes artifacts. The firmware author and the OS distributor each publish their own artifacts in their own format. A verifier must understand every format. Every new component adds another bespoke integration. Multiply that across platform configurations and the number of verifiers — some remote services, some local agents — and the integration tax compounds. The document calls this a scaling problem the industry needs to address. It gets worse fast.

CoSERV gives verifiers a single query language

The IETF RATS working group already standardized data formats. Entity Attestation Token (EAT) carries evidence. Concise Reference Integrity Manifest (CoRIM) carries reference values. Formats alone do not solve discovery. A verifier still needs to know where to fetch the right CoRIM for a specific GPU firmware version on a specific platform.

CoSERV, the Concise Selector for Endorsements and Reference Values, is an adopted work item in RATS that defines a common API for that discovery. Through one HTTP interface a verifier can query for relevant artifacts, collect them into bundles, and cache them using standard HTTP caching semantics. The specification does not dictate how suppliers create their trust information. It dictates how verifiers retrieve it.

Arm, Fujitsu, and Linaro are implementing CoSERV in Veraison. This open-source project serves as both a proving ground for emerging standards and a library of reusable components. Standards inform implementation. Implementation surfaces practical questions that specifications miss. Those lessons flow back into the standard before wide deployment. CoSERV remains in pre-RFC status — not finalized, still open to vendor input.

Trust must scale with the system

The work will be presented at OCP Global Summit 2026. The goal is not a single root of trust. The goal is interoperable building blocks so that every new combination of CPU, accelerator, firmware, and verifier does not require a custom integration. Confidential AI is the first workload that makes the problem visible. It will not be the last.