Cohere brings confidential computing to Model Vault, hiding inference data from cloud providers and itself

Cohere announced on September 17, 2026 that Model Vault, its single-tenant inference product, now supports confidential computing. The capability encrypts and isolates memory during AI inference, so that neither cloud providers nor Cohere itself can read enterprise customer data. The feature went live on September 17, 2026 at no additional cost for existing Model Vault customers.

According to Manoj Govindassamy, Cohere's director of serving inference, data is decrypted only at the point where it must actually be processed inside the protected CPU and GPU execution environment, with processor state and registers included in the confidential computing boundary. Model Vault's Encrypted tier combines a confidential VM on the CPU side with Nvidia GPUs running in confidential computing mode, using Intel TDX or AMD SEV-SNP on the CPU side. Protection now spans CPU, GPU and the interconnects between them, rather than stopping at the CPU.

Cohere said every inference returns an attestation report that lets customers verify the exact hardware, software and security policies protecting their workload. The company also plans to open source the full serving stack used in Model Vault, so independent auditors can validate that it does not log, export or leak data. Hardware and software golden values are registered with Intel's Trust Authority, and customers can use open-source tools to independently confirm their workloads are running inside the verified environment.

The launch comes amid heightened scrutiny over what AI vendors can see from customer data. In September 2026, reports said Nvidia, Palantir and Booz Allen Hamilton had restricted their use of Anthropic's models after a data retention policy change. Cohere signed a definitive merger agreement with Germany's Aleph Alpha on the same day, a roughly $20 billion deal, though the company said the two announcements are not connected beyond the calendar.