AI startup Cohere announced on September 16, 2026 that it has added a "Confidential Computing" capability to Model Vault, its enterprise AI inference platform. The feature encrypts the entire processing pipeline, from the moment a prompt is submitted to the moment a response is generated, so that Cohere itself can no longer see the contents of customer data.

※Thumbnail image is an AI-generated illustration.

Encrypting Data "In Use," Not Just at Rest and in Transit

Cloud services have traditionally encrypted data while it sits in storage and while it travels across a network, but data has typically been handled as plain text in memory while an AI model is actually running inference on it. Cohere's new confidential computing feature closes that gap. The entire pipeline, from a customer's prompt to the model's generated response, now runs inside a hardware-protected environment that neither Cohere's own staff nor the administrators of the underlying cloud infrastructure running Model Vault can inspect.

Isolating Both the CPU and the GPU

On the technical side, Cohere pairs CPU-based confidential virtual machines, using Intel TDX or AMD SEV-SNP, with NVIDIA GPUs running in their own confidential-computing mode. The link between the CPU and GPU is itself authenticated and encrypted through an SPDM-based secure session (SPDM is an industry standard for verifying the identity of connected hardware). Decryption keys are generated only inside the processor handling the workload and are never exposed to outside software or infrastructure operators. Clients also encrypt their requests directly to this protected environment using a method called Oblivious HTTP (OHTTP), so the keys never pass through Cohere's own systems.

Customers can independently verify that their workloads are actually protected. For every inference request, Cohere issues a "remote attestation" report proving that the environment is genuine confidential-computing hardware and specifying exactly which firmware, code, and configuration were loaded, using Intel's Trust Authority as the verification service. Cohere also says it plans to open-source the full Model Vault serving stack, the software that actually processes inference requests, so that independent auditors can verify the implementation themselves.

Built for Banks, Hospitals, and Government Agencies

Cohere positions the feature for customers such as banks, hospitals, and government agencies that handle regulated or highly sensitive data and require single-tenant deployments backed by vendor-signed attestations. The move reflects a growing demand among enterprises adopting generative AI for core operations: not wanting even their cloud provider to be able to see their data. According to Manoj Govindassamy, Cohere's Director of Serving Inference, the feature is available to existing Model Vault customers at no additional cost.

Summary

Cohere has added confidential computing to Model Vault, extending encryption to cover data in use in addition to data at rest and in transit. The approach combines Intel TDX or AMD SEV-SNP with NVIDIA GPUs running in confidential-computing mode, backed by remote attestation for verification and a planned open-sourcing of the serving stack. Offered at no extra cost, the feature is another sign that enterprise AI adoption in regulated industries is increasingly being shaped by demands for verifiable data sovereignty.