Hardware-backed security at global scale: Confidential G4 with NVIDIA Blackwell is GA

Google Cloud has launched Confidential G4 instances powered by NVIDIA RTX PRO 6000 Blackwell Server GPUs as generally available, integrating hardware-backed security into the AI Hypercomputer. The instances are now accessible across 36 global zones using multiple consumption models on Compute Engine VMs and GKE nodes, enabling enterprises to process sensitive AI workloads with verifiable cryptographic isolation. Support for Confidential Space, which facilitates secure multi-party data collaboration, is also rolling out to general availability in the coming weeks, expanding secure AI deployment options.

The Confidential G4 instances use non-exportable hardware keys, verifiable remote attestation, and transparent memory encryption to protect data in use without compromising GPU throughput or operational agility. Enterprises can deploy models using standard Ubuntu OS images, NVIDIA drivers, and CUDA-X libraries without requiring code refactoring, ensuring compatibility with existing workflows while meeting strict regulatory and compliance requirements.

Benchmarking by Google Cloud and customers shows that Confidential G4 delivers performance comparable to standard G4 instances for AI inference workloads, with less than 3.5% degradation in throughput for models like Llama 3, Llama 3.3, Resnet-50, and SDXL. The instances provide hardware-level assurance that prevents unauthorized access to sensitive data during processing, validated through real-world deployments and testing conditions.

Confidential G4 enables organizations to scale trusted AI on Google Cloud while maintaining control, confidentiality, and accountability, as highlighted by partners like Polaris Cloud and Vertebrae. The solution allows enterprises to combine sensitive datasets securely for collaborative analytics and joint model training using Confidential Space, ensuring data privacy even during multi-party computations.