Model Vault is the company’s managed, dedicated inference offering; the beta is intended for workloads involving sensitive enterprise data.
The service uses NVIDIA H100 GPUs with NVIDIA Confidential Computing on Microsoft Azure. Cohere says inference runs inside a hardware-backed confidential environment spanning CPU and GPU, with protected GPU memory and encrypted CPU-to-GPU traffic. Customer-side software performs remote attestation of signed hardware, firmware, and software configuration evidence before sending a request.
Cohere also says application payloads are encrypted end-to-end to the attested environment, while customers keep using Model Vault APIs and familiar request formats. Encrypted vaults can be created and managed in the public beta.
The announcement does not establish independent security results, pricing, regional availability, performance overhead, or a general-availability date. Cohere explicitly says the service does not replace application security, compliance, contractual controls, or data-residency requirements.