NVIDIA Confidential Computing Expands, Enabling Frontier Model Builders to Securely Scale Across Global Infrastructure

The new deployment option lets model builders run sensitive training and inference on dedicated third-party infrastructure without giving up control of models or data.
Key points
- Adds hardware-level access restrictions, platform integrity measurement, and in-transit data protection across third-party clouds and bare-metal environments.
- Expands NVIDIA Confidential Computing support to Grace Blackwell NVL72 rack-scale systems and x86-based HGX Rubin NVL8 and DGX Rubin NVL8 systems.
- Google Cloud will support Confidential Bare Metal on A4X Max instances powered by GB300 NVL72 systems.
- Cohere will integrate Confidential Bare Metal into Cohere Model Vault to secure GPU memory and interconnects during execution.
Why it matters
AI developers can now use bare-metal third-party cloud capacity for sensitive workloads while maintaining hardware-enforced protection for proprietary model weights and training data.
Sources
- NVIDIA Confidential Computing Expands, Enabling Frontier Model Builders to Securely Scale Across Global Infrastructure (NVIDIA AI, Oct 21, 2026, 11:18 AM)