DataEnclave is VAST’s confidential container runtime. It is a feature of VAST DataEngine that uses confidential VMs, with GPU support, to run containerized applications. It also includes an attestation framework to ensure that confidential resources (secrets) are only released into confidential VMs after those VMs have proven their configuration to be secure.
It allows proprietary LLMs to be deployed into untrusted environments in a way that prevents the administrators of that environment from being able to access model weights or other proprietary model details. In plain terms, it would allow Anthropic or OpenAI models to run on-prem without exposing the model’s source to anyone regardless of admin privileges.
Its architecture follows the CNCF CoCo implementation. I authored two technical documents that describe how DataEnclave works for VAST: