Cohere AI Updates: October 23, 2026

1. Cohere Launches Confidential Computing on Model Vault in Public Beta

Cohere. Cohere launched Confidential Computing for Model Vault, its dedicated managed inference offering, as a public beta that customers can deploy now through vault.cohere.com using the existing Cohere APIs. The beta runs on NVIDIA H100 GPUs on Microsoft Azure, with the model-serving stack in a confidential VM, NVIDIA Confidential Computing protecting GPU memory, and CPU-to-GPU traffic encrypted; Cohere says NVIDIA B200 support is planned for general availability. Requests are encrypted with Oblivious HTTP (OHTTP) to a key held only inside the attested environment, and the client verifies remote attestation evidence against an approved policy before sending any payload, refusing to send if verification fails. A companion technical post notes that metadata such as timing, message size, and model identifier stays outside the encryption, that side-channel attacks and prompt injection are out of scope, and that Cohere intends to open-source the stack, which builds on Kata Containers and Confidential Containers. Source Source