Cohere opens confidential computing beta for Model Vault
Cohere announced a public beta of Confidential Computing on Model Vault on October 22, 2026. The capability is designed to protect data while models process it, using a hardware-backed environment spanning the CPU and GPU.

Key takeaways · 5
- 01
The public beta brings confidential computing to Cohere’s managed Model Vault inference service.
- 02
The beta runs on NVIDIA H100 GPUs on Microsoft Azure.
- 03
Customers seeking beta access should contact Cohere; Cohere says encrypted workloads can be deployed today.
- 04
Model Vault Encrypted is priced separately from Standard Vault, and its model lineup may differ.
- 05
Confidential computing does not eliminate risks such as hardware side-channel and physical attacks.
What Cohere announced
Cohere announced the public beta of Confidential Computing on Model Vault on October 22, 2026.[1] Model Vault is a dedicated, fully managed model-inference service for enterprises, according to Cohere.[2] The new capability is designed to protect data while a model is processing it, which Cohere describes as “data in use.”[2] Inference runs in a hardware-backed confidential environment spanning the CPU and GPU.[2] For teams assessing hosted inference, the focus is on protecting information during processing, not just while it is being transmitted or stored.
How the protections work
Cohere says hardware-enforced isolation shields workload memory from infrastructure software and operators outside the confidential boundary.[2] The beta runs on NVIDIA H100 GPUs on Microsoft Azure, with Azure providing the underlying confidential-computing infrastructure.[2] NVIDIA Confidential Computing protects GPU memory and encrypts traffic between the CPU and GPU during inference.[2] Model Vault also encrypts application payloads end-to-end to the attested confidential environment; encrypted requests can pass through intermediary infrastructure without exposing their contents, and only that environment can decrypt them.[2] Cohere says responses are protected on the return path as well.[2]
Verification and API workflow
Customers continue to use the Model Vault API and familiar request formats, Cohere says.[2] Before sensitive data is sent, customer-side software cryptographically verifies the processing environment.[2] The checks include the environment’s hardware, firmware and software configuration, and the client refuses to send a request if the environment fails the approved security policy.[2] This puts a verification step on the customer side of the workflow. Teams evaluating the beta should account for that check when planning how requests are prepared and sent, rather than treating encryption alone as the complete control.
Access, pricing and limits
Cohere says customers can start deploying encrypted Model Vault workloads today, while its documentation directs customers seeking beta access to contact Cohere.[2][3] Model Vault Encrypted is priced separately from a Standard Vault.[3] It uses a confidential virtual machine and an NVIDIA GPU in confidential-computing mode.[4] Cohere says the encrypted-vault model lineup is still expanding and may differ from Standard Vault availability.[4] Its documentation also notes that confidential computing does not eliminate risks such as hardware side-channel attacks and physical attacks.[3] Those qualifications matter when comparing deployment options and setting security expectations.
The beta gives enterprise teams another way to assess managed inference when protecting data during processing is a requirement. Before adopting it, teams should check access, model availability, separate pricing and how the client-side verification fits their workflow.
Why it matters
Bring this to your team — one session a day, built for your industry.
kju.ai is built for teams: daily sessions, team analytics, and rollout support for your organization.
For teamsHow this developed
23 October 2026
Cohere opens confidential computing beta for Model Vault