It’s VAST Data’s turn to harness NVIDIA’s confidential computing capabilities.
Since the Hopper generation (2022), unified enclaves between CPU and GPU can be created. Communications between VMs pass through an encrypted channel.
Building on this foundation, VAST Data has grafted a DataEnclave module onto its DataEngine orchestrator. It is in a prerelease phase (commercial launch planned for the first quarter of 2027). The objective is to prevent deployments on infrastructures that are not under full control, benefiting both model providers and the enterprises that deploy and use them.
A attestation system guarantees that workloads run only within the appropriate confidential VMs. The server sits outside the cluster. VAST specifically ensures compatibility with the Trustee protocol of the CNCF Confidential Containers project. For disconnected deployments, it relies on Fortanix. Both parties bring their own Key Management System (KMS).
Within the VMs, the decryption of weights is performed with keys separate from those used to decrypt the workloads. The workloads are delivered as VAST applications—combining the inference service with the DataEngine configuration.
DataEnclave can, for state management, create encrypted volumes that are both ephemeral and persistent. It can potentially be deployed with partner models preloaded into the integrated registry. Cohere, Deepgram and NVIDIA, among others. On the server side, Cisco and Supermicro have joined the initiative. In the cloud arena, the current partners are four: Buzz, G42, Nscale and Sharon AI.
VAST Data acknowledges that there is a compute overhead… which “can exceed 10%.”
